Skip to content

GitLab

"llama/vscode:/vscode.git/clone" did not exist on "f86d00cd9572868d837be5b81342c7a1d5f2e8fc"
Commit cf1ed55a authored by suily's avatar suily
Browse files

Initial commit

parents
Pipeline #3338 canceled with stages
Hide whitespace changes
Inline Side-by-side
Showing with 4166 additions and 0 deletions
utils/loggers/clearml/clearml_utils.py 0 → 100644
View file @ cf1ed55a
"""Main Logger class for ClearML experiment tracking."""
import glob
import re
from pathlib import Path
import numpy as np
import yaml
from utils.plots import Annotator, colors
try:
import clearml
from clearml import Dataset, Task
assert hasattr(clearml, '__version__') # verify package import not local dir
except (ImportError, AssertionError):
clearml = None
def construct_dataset(clearml_info_string):
"""Load in a clearml dataset and fill the internal data_dict with its contents.
"""
dataset_id = clearml_info_string.replace('clearml://', '')
dataset = Dataset.get(dataset_id=dataset_id)
dataset_root_path = Path(dataset.get_local_copy())
# We'll search for the yaml file definition in the dataset
yaml_filenames = list(glob.glob(str(dataset_root_path / "*.yaml")) + glob.glob(str(dataset_root_path / "*.yml")))
if len(yaml_filenames) > 1:
raise ValueError('More than one yaml file was found in the dataset root, cannot determine which one contains '
'the dataset definition this way.')
elif len(yaml_filenames) == 0:
raise ValueError('No yaml definition found in dataset root path, check that there is a correct yaml file '
'inside the dataset root path.')
with open(yaml_filenames[0]) as f:
dataset_definition = yaml.safe_load(f)
assert set(dataset_definition.keys()).issuperset(
{'train', 'test', 'val', 'nc', 'names'}
), "The right keys were not found in the yaml file, make sure it at least has the following keys: ('train', 'test', 'val', 'nc', 'names')"
data_dict = dict()
data_dict['train'] = str(
(dataset_root_path / dataset_definition['train']).resolve()) if dataset_definition['train'] else None
data_dict['test'] = str(
(dataset_root_path / dataset_definition['test']).resolve()) if dataset_definition['test'] else None
data_dict['val'] = str(
(dataset_root_path / dataset_definition['val']).resolve()) if dataset_definition['val'] else None
data_dict['nc'] = dataset_definition['nc']
data_dict['names'] = dataset_definition['names']
return data_dict
class ClearmlLogger:
"""Log training runs, datasets, models, and predictions to ClearML.
This logger sends information to ClearML at app.clear.ml or to your own hosted server. By default,
this information includes hyperparameters, system configuration and metrics, model metrics, code information and
basic data metrics and analyses.
By providing additional command line arguments to train.py, datasets,
models and predictions can also be logged.
"""
def __init__(self, opt, hyp):
"""
- Initialize ClearML Task, this object will capture the experiment
- Upload dataset version to ClearML Data if opt.upload_dataset is True
arguments:
opt (namespace) -- Commandline arguments for this run
hyp (dict) -- Hyperparameters for this run
"""
self.current_epoch = 0
# Keep tracked of amount of logged images to enforce a limit
self.current_epoch_logged_images = set()
# Maximum number of images to log to clearML per epoch
self.max_imgs_to_log_per_epoch = 16
# Get the interval of epochs when bounding box images should be logged
self.bbox_interval = opt.bbox_interval
self.clearml = clearml
self.task = None
self.data_dict = None
if self.clearml:
self.task = Task.init(
project_name=opt.project if opt.project != 'runs/train' else 'YOLOv5',
task_name=opt.name if opt.name != 'exp' else 'Training',
tags=['YOLOv5'],
output_uri=True,
auto_connect_frameworks={'pytorch': False}
# We disconnect pytorch auto-detection, because we added manual model save points in the code
)
# ClearML's hooks will already grab all general parameters
# Only the hyperparameters coming from the yaml config file
# will have to be added manually!
self.task.connect(hyp, name='Hyperparameters')
# Get ClearML Dataset Version if requested
if opt.data.startswith('clearml://'):
# data_dict should have the following keys:
# names, nc (number of classes), test, train, val (all three relative paths to ../datasets)
self.data_dict = construct_dataset(opt.data)
# Set data to data_dict because wandb will crash without this information and opt is the best way
# to give it to them
opt.data = self.data_dict
def log_debug_samples(self, files, title='Debug Samples'):
"""
Log files (images) as debug samples in the ClearML task.
arguments:
files (List(PosixPath)) a list of file paths in PosixPath format
title (str) A title that groups together images with the same values
"""
for f in files:
if f.exists():
it = re.search(r'_batch(\d+)', f.name)
iteration = int(it.groups()[0]) if it else 0
self.task.get_logger().report_image(title=title,
series=f.name.replace(it.group(), ''),
local_path=str(f),
iteration=iteration)
def log_image_with_boxes(self, image_path, boxes, class_names, image, conf_threshold=0.25):
"""
Draw the bounding boxes on a single image and report the result as a ClearML debug sample.
arguments:
image_path (PosixPath) the path the original image file
boxes (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class]
class_names (dict): dict containing mapping of class int to class name
image (Tensor): A torch tensor containing the actual image data
"""
if len(self.current_epoch_logged_images) < self.max_imgs_to_log_per_epoch and self.current_epoch >= 0:
# Log every bbox_interval times and deduplicate for any intermittend extra eval runs
if self.current_epoch % self.bbox_interval == 0 and image_path not in self.current_epoch_logged_images:
im = np.ascontiguousarray(np.moveaxis(image.mul(255).clamp(0, 255).byte().cpu().numpy(), 0, 2))
annotator = Annotator(im=im, pil=True)
for i, (conf, class_nr, box) in enumerate(zip(boxes[:, 4], boxes[:, 5], boxes[:, :4])):
color = colors(i)
class_name = class_names[int(class_nr)]
confidence_percentage = round(float(conf) * 100, 2)
label = f"{class_name}: {confidence_percentage}%"
if conf > conf_threshold:
annotator.rectangle(box.cpu().numpy(), outline=color)
annotator.box_label(box.cpu().numpy(), label=label, color=color)
annotated_image = annotator.result()
self.task.get_logger().report_image(title='Bounding Boxes',
series=image_path.name,
iteration=self.current_epoch,
image=annotated_image)
self.current_epoch_logged_images.add(image_path)
utils/loggers/clearml/hpo.py 0 → 100644
View file @ cf1ed55a
from clearml import Task
# Connecting ClearML with the current process,
# from here on everything is logged automatically
from clearml.automation import HyperParameterOptimizer, UniformParameterRange
from clearml.automation.optuna import OptimizerOptuna
task = Task.init(project_name='Hyper-Parameter Optimization',
task_name='YOLOv5',
task_type=Task.TaskTypes.optimizer,
reuse_last_task_id=False)
# Example use case:
optimizer = HyperParameterOptimizer(
# This is the experiment we want to optimize
base_task_id='<your_template_task_id>',
# here we define the hyper-parameters to optimize
# Notice: The parameter name should exactly match what you see in the UI: <section_name>/<parameter>
# For Example, here we see in the base experiment a section Named: "General"
# under it a parameter named "batch_size", this becomes "General/batch_size"
# If you have `argparse` for example, then arguments will appear under the "Args" section,
# and you should instead pass "Args/batch_size"
hyper_parameters=[
UniformParameterRange('Hyperparameters/lr0', min_value=1e-5, max_value=1e-1),
UniformParameterRange('Hyperparameters/lrf', min_value=0.01, max_value=1.0),
UniformParameterRange('Hyperparameters/momentum', min_value=0.6, max_value=0.98),
UniformParameterRange('Hyperparameters/weight_decay', min_value=0.0, max_value=0.001),
UniformParameterRange('Hyperparameters/warmup_epochs', min_value=0.0, max_value=5.0),
UniformParameterRange('Hyperparameters/warmup_momentum', min_value=0.0, max_value=0.95),
UniformParameterRange('Hyperparameters/warmup_bias_lr', min_value=0.0, max_value=0.2),
UniformParameterRange('Hyperparameters/box', min_value=0.02, max_value=0.2),
UniformParameterRange('Hyperparameters/cls', min_value=0.2, max_value=4.0),
UniformParameterRange('Hyperparameters/cls_pw', min_value=0.5, max_value=2.0),
UniformParameterRange('Hyperparameters/obj', min_value=0.2, max_value=4.0),
UniformParameterRange('Hyperparameters/obj_pw', min_value=0.5, max_value=2.0),
UniformParameterRange('Hyperparameters/iou_t', min_value=0.1, max_value=0.7),
UniformParameterRange('Hyperparameters/anchor_t', min_value=2.0, max_value=8.0),
UniformParameterRange('Hyperparameters/fl_gamma', min_value=0.0, max_value=4.0),
UniformParameterRange('Hyperparameters/hsv_h', min_value=0.0, max_value=0.1),
UniformParameterRange('Hyperparameters/hsv_s', min_value=0.0, max_value=0.9),
UniformParameterRange('Hyperparameters/hsv_v', min_value=0.0, max_value=0.9),
UniformParameterRange('Hyperparameters/degrees', min_value=0.0, max_value=45.0),
UniformParameterRange('Hyperparameters/translate', min_value=0.0, max_value=0.9),
UniformParameterRange('Hyperparameters/scale', min_value=0.0, max_value=0.9),
UniformParameterRange('Hyperparameters/shear', min_value=0.0, max_value=10.0),
UniformParameterRange('Hyperparameters/perspective', min_value=0.0, max_value=0.001),
UniformParameterRange('Hyperparameters/flipud', min_value=0.0, max_value=1.0),
UniformParameterRange('Hyperparameters/fliplr', min_value=0.0, max_value=1.0),
UniformParameterRange('Hyperparameters/mosaic', min_value=0.0, max_value=1.0),
UniformParameterRange('Hyperparameters/mixup', min_value=0.0, max_value=1.0),
UniformParameterRange('Hyperparameters/copy_paste', min_value=0.0, max_value=1.0)],
# this is the objective metric we want to maximize/minimize
objective_metric_title='metrics',
objective_metric_series='mAP_0.5',
# now we decide if we want to maximize it or minimize it (accuracy we maximize)
objective_metric_sign='max',
# let us limit the number of concurrent experiments,
# this in turn will make sure we do dont bombard the scheduler with experiments.
# if we have an auto-scaler connected, this, by proxy, will limit the number of machine
max_number_of_concurrent_tasks=1,
# this is the optimizer class (actually doing the optimization)
# Currently, we can choose from GridSearch, RandomSearch or OptimizerBOHB (Bayesian optimization Hyper-Band)
optimizer_class=OptimizerOptuna,
# If specified only the top K performing Tasks will be kept, the others will be automatically archived
save_top_k_tasks_only=5, # 5,
compute_time_limit=None,
total_max_jobs=20,
min_iteration_per_job=None,
max_iteration_per_job=None,
)
# report every 10 seconds, this is way too often, but we are testing here
optimizer.set_report_period(10 / 60)
# You can also use the line below instead to run all the optimizer tasks locally, without using queues or agent
# an_optimizer.start_locally(job_complete_callback=job_complete_callback)
# set the time limit for the optimization process (2 hours)
optimizer.set_time_limit(in_minutes=120.0)
# Start the optimization process in the local environment
optimizer.start_locally()
# wait until process is done (notice we are controlling the optimization process in the background)
optimizer.wait()
# make sure background optimization stopped
optimizer.stop()
print('We are done, good bye')
utils/loggers/comet/README.md 0 → 100644
View file @ cf1ed55a
<img src="https://cdn.comet.ml/img/notebook_logo.png">
# YOLOv5 with Comet
This guide will cover how to use YOLOv5 with [Comet](https://bit.ly/yolov5-readme-comet)
# About Comet
Comet builds tools that help data scientists, engineers, and team leaders accelerate and optimize machine learning and deep learning models.
Track and visualize model metrics in real time, save your hyperparameters, datasets, and model checkpoints, and visualize your model predictions with [Comet Custom Panels](https://bit.ly/yolov5-colab-comet-panels)!
Comet makes sure you never lose track of your work and makes it easy to share results and collaborate across teams of all sizes!
# Getting Started
## Install Comet
```shell
pip install comet_ml
```
## Configure Comet Credentials
There are two ways to configure Comet with YOLOv5.
You can either set your credentials through enviroment variables
**Environment Variables**
```shell
export COMET_API_KEY=<Your Comet API Key>
export COMET_PROJECT_NAME=<Your Comet Project Name> # This will default to 'yolov5'
```
Or create a `.comet.config` file in your working directory and set your credentials there.
**Comet Configuration File**
```
[comet]
api_key=<Your Comet API Key>
project_name=<Your Comet Project Name> # This will default to 'yolov5'
```
## Run the Training Script
```shell
# Train YOLOv5s on COCO128 for 5 epochs
python train.py --img 640 --batch 16 --epochs 5 --data coco128.yaml --weights yolov5s.pt
```
That's it! Comet will automatically log your hyperparameters, command line arguments, training and valiation metrics. You can visualize and analyze your runs in the Comet UI
<img width="1920" alt="yolo-ui" src="https://user-images.githubusercontent.com/26833433/202851203-164e94e1-2238-46dd-91f8-de020e9d6b41.png">
# Try out an Example!
Check out an example of a [completed run here](https://www.comet.com/examples/comet-example-yolov5/a0e29e0e9b984e4a822db2a62d0cb357?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step&ref=yolov5&utm_source=yolov5&utm_medium=affilliate&utm_campaign=yolov5_comet_integration)
Or better yet, try it out yourself in this Colab Notebook
[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1RG0WOQyxlDlo5Km8GogJpIEJlg_5lyYO?usp=sharing)
# Log automatically
By default, Comet will log the following items
## Metrics
- Box Loss, Object Loss, Classification Loss for the training and validation data
- mAP_0.5, mAP_0.5:0.95 metrics for the validation data.
- Precision and Recall for the validation data
## Parameters
- Model Hyperparameters
- All parameters passed through the command line options
## Visualizations
- Confusion Matrix of the model predictions on the validation data
- Plots for the PR and F1 curves across all classes
- Correlogram of the Class Labels
# Configure Comet Logging
Comet can be configured to log additional data either through command line flags passed to the training script
or through environment variables.
```shell
export COMET_MODE=online # Set whether to run Comet in 'online' or 'offline' mode. Defaults to online
export COMET_MODEL_NAME=<your model name> #Set the name for the saved model. Defaults to yolov5
export COMET_LOG_CONFUSION_MATRIX=false # Set to disable logging a Comet Confusion Matrix. Defaults to true
export COMET_MAX_IMAGE_UPLOADS=<number of allowed images to upload to Comet> # Controls how many total image predictions to log to Comet. Defaults to 100.
export COMET_LOG_PER_CLASS_METRICS=true # Set to log evaluation metrics for each detected class at the end of training. Defaults to false
export COMET_DEFAULT_CHECKPOINT_FILENAME=<your checkpoint filename> # Set this if you would like to resume training from a different checkpoint. Defaults to 'last.pt'
export COMET_LOG_BATCH_LEVEL_METRICS=true # Set this if you would like to log training metrics at the batch level. Defaults to false.
export COMET_LOG_PREDICTIONS=true # Set this to false to disable logging model predictions
```
## Logging Checkpoints with Comet
Logging Models to Comet is disabled by default. To enable it, pass the `save-period` argument to the training script. This will save the
logged checkpoints to Comet based on the interval value provided by `save-period`
```shell
python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data coco128.yaml \
--weights yolov5s.pt \
--save-period 1
```
## Logging Model Predictions
By default, model predictions (images, ground truth labels and bounding boxes) will be logged to Comet.
You can control the frequency of logged predictions and the associated images by passing the `bbox_interval` command line argument. Predictions can be visualized using Comet's Object Detection Custom Panel. This frequency corresponds to every Nth batch of data per epoch. In the example below, we are logging every 2nd batch of data for each epoch.
**Note:** The YOLOv5 validation dataloader will default to a batch size of 32, so you will have to set the logging frequency accordingly.
Here is an [example project using the Panel](https://www.comet.com/examples/comet-example-yolov5?shareable=YcwMiJaZSXfcEXpGOHDD12vA1&ref=yolov5&utm_source=yolov5&utm_medium=affilliate&utm_campaign=yolov5_comet_integration)
```shell
python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data coco128.yaml \
--weights yolov5s.pt \
--bbox_interval 2
```
### Controlling the number of Prediction Images logged to Comet
When logging predictions from YOLOv5, Comet will log the images associated with each set of predictions. By default a maximum of 100 validation images are logged. You can increase or decrease this number using the `COMET_MAX_IMAGE_UPLOADS` environment variable.
```shell
env COMET_MAX_IMAGE_UPLOADS=200 python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data coco128.yaml \
--weights yolov5s.pt \
--bbox_interval 1
```
### Logging Class Level Metrics
Use the `COMET_LOG_PER_CLASS_METRICS` environment variable to log mAP, precision, recall, f1 for each class.
```shell
env COMET_LOG_PER_CLASS_METRICS=true python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data coco128.yaml \
--weights yolov5s.pt
```
## Uploading a Dataset to Comet Artifacts
If you would like to store your data using [Comet Artifacts](https://www.comet.com/docs/v2/guides/data-management/using-artifacts/#learn-more?ref=yolov5&utm_source=yolov5&utm_medium=affilliate&utm_campaign=yolov5_comet_integration), you can do so using the `upload_dataset` flag.
The dataset be organized in the way described in the [YOLOv5 documentation](https://docs.ultralytics.com/tutorials/train-custom-datasets/#3-organize-directories). The dataset config `yaml` file must follow the same format as that of the `coco128.yaml` file.
```shell
python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data coco128.yaml \
--weights yolov5s.pt \
--upload_dataset
```
You can find the uploaded dataset in the Artifacts tab in your Comet Workspace
<img width="1073" alt="artifact-1" src="https://user-images.githubusercontent.com/7529846/186929193-162718bf-ec7b-4eb9-8c3b-86b3763ef8ea.png">
You can preview the data directly in the Comet UI.
<img width="1082" alt="artifact-2" src="https://user-images.githubusercontent.com/7529846/186929215-432c36a9-c109-4eb0-944b-84c2786590d6.png">
Artifacts are versioned and also support adding metadata about the dataset. Comet will automatically log the metadata from your dataset `yaml` file
<img width="963" alt="artifact-3" src="https://user-images.githubusercontent.com/7529846/186929256-9d44d6eb-1a19-42de-889a-bcbca3018f2e.png">
### Using a saved Artifact
If you would like to use a dataset from Comet Artifacts, set the `path` variable in your dataset `yaml` file to point to the following Artifact resource URL.
```
# contents of artifact.yaml file
path: "comet://<workspace name>/<artifact name>:<artifact version or alias>"
```
Then pass this file to your training script in the following way
```shell
python train.py \
--img 640 \
--batch 16 \
--epochs 5 \
--data artifact.yaml \
--weights yolov5s.pt
```
Artifacts also allow you to track the lineage of data as it flows through your Experimentation workflow. Here you can see a graph that shows you all the experiments that have used your uploaded dataset.
<img width="1391" alt="artifact-4" src="https://user-images.githubusercontent.com/7529846/186929264-4c4014fa-fe51-4f3c-a5c5-f6d24649b1b4.png">
## Resuming a Training Run
If your training run is interrupted for any reason, e.g. disrupted internet connection, you can resume the run using the `resume` flag and the Comet Run Path.
The Run Path has the following format `comet://<your workspace name>/<your project name>/<experiment id>`.
This will restore the run to its state before the interruption, which includes restoring the model from a checkpoint, restoring all hyperparameters and training arguments and downloading Comet dataset Artifacts if they were used in the original run. The resumed run will continue logging to the existing Experiment in the Comet UI
```shell
python train.py \
--resume "comet://<your run path>"
```
## Hyperparameter Search with the Comet Optimizer
YOLOv5 is also integrated with Comet's Optimizer, making is simple to visualie hyperparameter sweeps in the Comet UI.
### Configuring an Optimizer Sweep
To configure the Comet Optimizer, you will have to create a JSON file with the information about the sweep. An example file has been provided in `utils/loggers/comet/optimizer_config.json`
```shell
python utils/loggers/comet/hpo.py \
--comet_optimizer_config "utils/loggers/comet/optimizer_config.json"
```
The `hpo.py` script accepts the same arguments as `train.py`. If you wish to pass additional arguments to your sweep simply add them after
the script.
```shell
python utils/loggers/comet/hpo.py \
--comet_optimizer_config "utils/loggers/comet/optimizer_config.json" \
--save-period 1 \
--bbox_interval 1
```
### Running a Sweep in Parallel
```shell
comet optimizer -j <set number of workers> utils/loggers/comet/hpo.py \
utils/loggers/comet/optimizer_config.json"
```
### Visualizing Results
Comet provides a number of ways to visualize the results of your sweep. Take a look at a [project with a completed sweep here](https://www.comet.com/examples/comet-example-yolov5/view/PrlArHGuuhDTKC1UuBmTtOSXD/panels?ref=yolov5&utm_source=yolov5&utm_medium=affilliate&utm_campaign=yolov5_comet_integration)
<img width="1626" alt="hyperparameter-yolo" src="https://user-images.githubusercontent.com/7529846/186914869-7dc1de14-583f-4323-967b-c9a66a29e495.png">
utils/loggers/comet/__init__.py 0 → 100644
View file @ cf1ed55a
import glob
import json
import logging
import os
import sys
from pathlib import Path
logger = logging.getLogger(__name__)
FILE = Path(__file__).resolve()
ROOT = FILE.parents[3] # YOLOv5 root directory
if str(ROOT) not in sys.path:
sys.path.append(str(ROOT)) # add ROOT to PATH
try:
import comet_ml
# Project Configuration
config = comet_ml.config.get_config()
COMET_PROJECT_NAME = config.get_string(os.getenv("COMET_PROJECT_NAME"), "comet.project_name", default="yolov5")
except (ModuleNotFoundError, ImportError):
comet_ml = None
COMET_PROJECT_NAME = None
import PIL
import torch
import torchvision.transforms as T
import yaml
from utils.dataloaders import img2label_paths
from utils.general import check_dataset, scale_boxes, xywh2xyxy
from utils.metrics import box_iou
COMET_PREFIX = "comet://"
COMET_MODE = os.getenv("COMET_MODE", "online")
# Model Saving Settings
COMET_MODEL_NAME = os.getenv("COMET_MODEL_NAME", "yolov5")
# Dataset Artifact Settings
COMET_UPLOAD_DATASET = os.getenv("COMET_UPLOAD_DATASET", "false").lower() == "true"
# Evaluation Settings
COMET_LOG_CONFUSION_MATRIX = os.getenv("COMET_LOG_CONFUSION_MATRIX", "true").lower() == "true"
COMET_LOG_PREDICTIONS = os.getenv("COMET_LOG_PREDICTIONS", "true").lower() == "true"
COMET_MAX_IMAGE_UPLOADS = int(os.getenv("COMET_MAX_IMAGE_UPLOADS", 100))
# Confusion Matrix Settings
CONF_THRES = float(os.getenv("CONF_THRES", 0.001))
IOU_THRES = float(os.getenv("IOU_THRES", 0.6))
# Batch Logging Settings
COMET_LOG_BATCH_METRICS = os.getenv("COMET_LOG_BATCH_METRICS", "false").lower() == "true"
COMET_BATCH_LOGGING_INTERVAL = os.getenv("COMET_BATCH_LOGGING_INTERVAL", 1)
COMET_PREDICTION_LOGGING_INTERVAL = os.getenv("COMET_PREDICTION_LOGGING_INTERVAL", 1)
COMET_LOG_PER_CLASS_METRICS = os.getenv("COMET_LOG_PER_CLASS_METRICS", "false").lower() == "true"
RANK = int(os.getenv("RANK", -1))
to_pil = T.ToPILImage()
class CometLogger:
"""Log metrics, parameters, source code, models and much more
with Comet
"""
def __init__(self, opt, hyp, run_id=None, job_type="Training", **experiment_kwargs) -> None:
self.job_type = job_type
self.opt = opt
self.hyp = hyp
# Comet Flags
self.comet_mode = COMET_MODE
self.save_model = opt.save_period > -1
self.model_name = COMET_MODEL_NAME
# Batch Logging Settings
self.log_batch_metrics = COMET_LOG_BATCH_METRICS
self.comet_log_batch_interval = COMET_BATCH_LOGGING_INTERVAL
# Dataset Artifact Settings
self.upload_dataset = self.opt.upload_dataset if self.opt.upload_dataset else COMET_UPLOAD_DATASET
self.resume = self.opt.resume
# Default parameters to pass to Experiment objects
self.default_experiment_kwargs = {
"log_code": False,
"log_env_gpu": True,
"log_env_cpu": True,
"project_name": COMET_PROJECT_NAME,}
self.default_experiment_kwargs.update(experiment_kwargs)
self.experiment = self._get_experiment(self.comet_mode, run_id)
self.data_dict = self.check_dataset(self.opt.data)
self.class_names = self.data_dict["names"]
self.num_classes = self.data_dict["nc"]
self.logged_images_count = 0
self.max_images = COMET_MAX_IMAGE_UPLOADS
if run_id is None:
self.experiment.log_other("Created from", "YOLOv5")
if not isinstance(self.experiment, comet_ml.OfflineExperiment):
workspace, project_name, experiment_id = self.experiment.url.split("/")[-3:]
self.experiment.log_other(
"Run Path",
f"{workspace}/{project_name}/{experiment_id}",
)
self.log_parameters(vars(opt))
self.log_parameters(self.opt.hyp)
self.log_asset_data(
self.opt.hyp,
name="hyperparameters.json",
metadata={"type": "hyp-config-file"},
)
self.log_asset(
f"{self.opt.save_dir}/opt.yaml",
metadata={"type": "opt-config-file"},
)
self.comet_log_confusion_matrix = COMET_LOG_CONFUSION_MATRIX
if hasattr(self.opt, "conf_thres"):
self.conf_thres = self.opt.conf_thres
else:
self.conf_thres = CONF_THRES
if hasattr(self.opt, "iou_thres"):
self.iou_thres = self.opt.iou_thres
else:
self.iou_thres = IOU_THRES
self.log_parameters({"val_iou_threshold": self.iou_thres, "val_conf_threshold": self.conf_thres})
self.comet_log_predictions = COMET_LOG_PREDICTIONS
if self.opt.bbox_interval == -1:
self.comet_log_prediction_interval = 1 if self.opt.epochs < 10 else self.opt.epochs // 10
else:
self.comet_log_prediction_interval = self.opt.bbox_interval
if self.comet_log_predictions:
self.metadata_dict = {}
self.logged_image_names = []
self.comet_log_per_class_metrics = COMET_LOG_PER_CLASS_METRICS
self.experiment.log_others({
"comet_mode": COMET_MODE,
"comet_max_image_uploads": COMET_MAX_IMAGE_UPLOADS,
"comet_log_per_class_metrics": COMET_LOG_PER_CLASS_METRICS,
"comet_log_batch_metrics": COMET_LOG_BATCH_METRICS,
"comet_log_confusion_matrix": COMET_LOG_CONFUSION_MATRIX,
"comet_model_name": COMET_MODEL_NAME,})
# Check if running the Experiment with the Comet Optimizer
if hasattr(self.opt, "comet_optimizer_id"):
self.experiment.log_other("optimizer_id", self.opt.comet_optimizer_id)
self.experiment.log_other("optimizer_objective", self.opt.comet_optimizer_objective)
self.experiment.log_other("optimizer_metric", self.opt.comet_optimizer_metric)
self.experiment.log_other("optimizer_parameters", json.dumps(self.hyp))
def _get_experiment(self, mode, experiment_id=None):
if mode == "offline":
if experiment_id is not None:
return comet_ml.ExistingOfflineExperiment(
previous_experiment=experiment_id,
**self.default_experiment_kwargs,
)
return comet_ml.OfflineExperiment(**self.default_experiment_kwargs,)
else:
try:
if experiment_id is not None:
return comet_ml.ExistingExperiment(
previous_experiment=experiment_id,
**self.default_experiment_kwargs,
)
return comet_ml.Experiment(**self.default_experiment_kwargs)
except ValueError:
logger.warning("COMET WARNING: "
"Comet credentials have not been set. "
"Comet will default to offline logging. "
"Please set your credentials to enable online logging.")
return self._get_experiment("offline", experiment_id)
return
def log_metrics(self, log_dict, **kwargs):
self.experiment.log_metrics(log_dict, **kwargs)
def log_parameters(self, log_dict, **kwargs):
self.experiment.log_parameters(log_dict, **kwargs)
def log_asset(self, asset_path, **kwargs):
self.experiment.log_asset(asset_path, **kwargs)
def log_asset_data(self, asset, **kwargs):
self.experiment.log_asset_data(asset, **kwargs)
def log_image(self, img, **kwargs):
self.experiment.log_image(img, **kwargs)
def log_model(self, path, opt, epoch, fitness_score, best_model=False):
if not self.save_model:
return
model_metadata = {
"fitness_score": fitness_score[-1],
"epochs_trained": epoch + 1,
"save_period": opt.save_period,
"total_epochs": opt.epochs,}
model_files = glob.glob(f"{path}/*.pt")
for model_path in model_files:
name = Path(model_path).name
self.experiment.log_model(
self.model_name,
file_or_folder=model_path,
file_name=name,
metadata=model_metadata,
overwrite=True,
)
def check_dataset(self, data_file):
with open(data_file) as f:
data_config = yaml.safe_load(f)
if data_config['path'].startswith(COMET_PREFIX):
path = data_config['path'].replace(COMET_PREFIX, "")
data_dict = self.download_dataset_artifact(path)
return data_dict
self.log_asset(self.opt.data, metadata={"type": "data-config-file"})
return check_dataset(data_file)
def log_predictions(self, image, labelsn, path, shape, predn):
if self.logged_images_count >= self.max_images:
return
detections = predn[predn[:, 4] > self.conf_thres]
iou = box_iou(labelsn[:, 1:], detections[:, :4])
mask, _ = torch.where(iou > self.iou_thres)
if len(mask) == 0:
return
filtered_detections = detections[mask]
filtered_labels = labelsn[mask]
image_id = path.split("/")[-1].split(".")[0]
image_name = f"{image_id}_curr_epoch_{self.experiment.curr_epoch}"
if image_name not in self.logged_image_names:
native_scale_image = PIL.Image.open(path)
self.log_image(native_scale_image, name=image_name)
self.logged_image_names.append(image_name)
metadata = []
for cls, *xyxy in filtered_labels.tolist():
metadata.append({
"label": f"{self.class_names[int(cls)]}-gt",
"score": 100,
"box": {
"x": xyxy[0],
"y": xyxy[1],
"x2": xyxy[2],
"y2": xyxy[3]},})
for *xyxy, conf, cls in filtered_detections.tolist():
metadata.append({
"label": f"{self.class_names[int(cls)]}",
"score": conf * 100,
"box": {
"x": xyxy[0],
"y": xyxy[1],
"x2": xyxy[2],
"y2": xyxy[3]},})
self.metadata_dict[image_name] = metadata
self.logged_images_count += 1
return
def preprocess_prediction(self, image, labels, shape, pred):
nl, _ = labels.shape[0], pred.shape[0]
# Predictions
if self.opt.single_cls:
pred[:, 5] = 0
predn = pred.clone()
scale_boxes(image.shape[1:], predn[:, :4], shape[0], shape[1])
labelsn = None
if nl:
tbox = xywh2xyxy(labels[:, 1:5]) # target boxes
scale_boxes(image.shape[1:], tbox, shape[0], shape[1]) # native-space labels
labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels
scale_boxes(image.shape[1:], predn[:, :4], shape[0], shape[1]) # native-space pred
return predn, labelsn
def add_assets_to_artifact(self, artifact, path, asset_path, split):
img_paths = sorted(glob.glob(f"{asset_path}/*"))
label_paths = img2label_paths(img_paths)
for image_file, label_file in zip(img_paths, label_paths):
image_logical_path, label_logical_path = map(lambda x: os.path.relpath(x, path), [image_file, label_file])
try:
artifact.add(image_file, logical_path=image_logical_path, metadata={"split": split})
artifact.add(label_file, logical_path=label_logical_path, metadata={"split": split})
except ValueError as e:
logger.error('COMET ERROR: Error adding file to Artifact. Skipping file.')
logger.error(f"COMET ERROR: {e}")
continue
return artifact
def upload_dataset_artifact(self):
dataset_name = self.data_dict.get("dataset_name", "yolov5-dataset")
path = str((ROOT / Path(self.data_dict["path"])).resolve())
metadata = self.data_dict.copy()
for key in ["train", "val", "test"]:
split_path = metadata.get(key)
if split_path is not None:
metadata[key] = split_path.replace(path, "")
artifact = comet_ml.Artifact(name=dataset_name, artifact_type="dataset", metadata=metadata)
for key in metadata.keys():
if key in ["train", "val", "test"]:
if isinstance(self.upload_dataset, str) and (key != self.upload_dataset):
continue
asset_path = self.data_dict.get(key)
if asset_path is not None:
artifact = self.add_assets_to_artifact(artifact, path, asset_path, key)
self.experiment.log_artifact(artifact)
return
def download_dataset_artifact(self, artifact_path):
logged_artifact = self.experiment.get_artifact(artifact_path)
artifact_save_dir = str(Path(self.opt.save_dir) / logged_artifact.name)
logged_artifact.download(artifact_save_dir)
metadata = logged_artifact.metadata
data_dict = metadata.copy()
data_dict["path"] = artifact_save_dir
metadata_names = metadata.get("names")
if type(metadata_names) == dict:
data_dict["names"] = {int(k): v for k, v in metadata.get("names").items()}
elif type(metadata_names) == list:
data_dict["names"] = {int(k): v for k, v in zip(range(len(metadata_names)), metadata_names)}
else:
raise "Invalid 'names' field in dataset yaml file. Please use a list or dictionary"
data_dict = self.update_data_paths(data_dict)
return data_dict
def update_data_paths(self, data_dict):
path = data_dict.get("path", "")
for split in ["train", "val", "test"]:
if data_dict.get(split):
split_path = data_dict.get(split)
data_dict[split] = (f"{path}/{split_path}" if isinstance(split, str) else [
f"{path}/{x}" for x in split_path])
return data_dict
def on_pretrain_routine_end(self, paths):
if self.opt.resume:
return
for path in paths:
self.log_asset(str(path))
if self.upload_dataset:
if not self.resume:
self.upload_dataset_artifact()
return
def on_train_start(self):
self.log_parameters(self.hyp)
def on_train_epoch_start(self):
return
def on_train_epoch_end(self, epoch):
self.experiment.curr_epoch = epoch
return
def on_train_batch_start(self):
return
def on_train_batch_end(self, log_dict, step):
self.experiment.curr_step = step
if self.log_batch_metrics and (step % self.comet_log_batch_interval == 0):
self.log_metrics(log_dict, step=step)
return
def on_train_end(self, files, save_dir, last, best, epoch, results):
if self.comet_log_predictions:
curr_epoch = self.experiment.curr_epoch
self.experiment.log_asset_data(self.metadata_dict, "image-metadata.json", epoch=curr_epoch)
for f in files:
self.log_asset(f, metadata={"epoch": epoch})
self.log_asset(f"{save_dir}/results.csv", metadata={"epoch": epoch})
if not self.opt.evolve:
model_path = str(best if best.exists() else last)
name = Path(model_path).name
if self.save_model:
self.experiment.log_model(
self.model_name,
file_or_folder=model_path,
file_name=name,
overwrite=True,
)
# Check if running Experiment with Comet Optimizer
if hasattr(self.opt, 'comet_optimizer_id'):
metric = results.get(self.opt.comet_optimizer_metric)
self.experiment.log_other('optimizer_metric_value', metric)
self.finish_run()
def on_val_start(self):
return
def on_val_batch_start(self):
return
def on_val_batch_end(self, batch_i, images, targets, paths, shapes, outputs):
if not (self.comet_log_predictions and ((batch_i + 1) % self.comet_log_prediction_interval == 0)):
return
for si, pred in enumerate(outputs):
if len(pred) == 0:
continue
image = images[si]
labels = targets[targets[:, 0] == si, 1:]
shape = shapes[si]
path = paths[si]
predn, labelsn = self.preprocess_prediction(image, labels, shape, pred)
if labelsn is not None:
self.log_predictions(image, labelsn, path, shape, predn)
return
def on_val_end(self, nt, tp, fp, p, r, f1, ap, ap50, ap_class, confusion_matrix):
if self.comet_log_per_class_metrics:
if self.num_classes > 1:
for i, c in enumerate(ap_class):
class_name = self.class_names[c]
self.experiment.log_metrics(
{
'mAP@.5': ap50[i],
'mAP@.5:.95': ap[i],
'precision': p[i],
'recall': r[i],
'f1': f1[i],
'true_positives': tp[i],
'false_positives': fp[i],
'support': nt[c]},
prefix=class_name)
if self.comet_log_confusion_matrix:
epoch = self.experiment.curr_epoch
class_names = list(self.class_names.values())
class_names.append("background")
num_classes = len(class_names)
self.experiment.log_confusion_matrix(
matrix=confusion_matrix.matrix,
max_categories=num_classes,
labels=class_names,
epoch=epoch,
column_label='Actual Category',
row_label='Predicted Category',
file_name=f"confusion-matrix-epoch-{epoch}.json",
)
def on_fit_epoch_end(self, result, epoch):
self.log_metrics(result, epoch=epoch)
def on_model_save(self, last, epoch, final_epoch, best_fitness, fi):
if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1:
self.log_model(last.parent, self.opt, epoch, fi, best_model=best_fitness == fi)
def on_params_update(self, params):
self.log_parameters(params)
def finish_run(self):
self.experiment.end()
utils/loggers/comet/comet_utils.py 0 → 100644
View file @ cf1ed55a
import logging
import os
from urllib.parse import urlparse
try:
import comet_ml
except (ModuleNotFoundError, ImportError):
comet_ml = None
import yaml
logger = logging.getLogger(__name__)
COMET_PREFIX = "comet://"
COMET_MODEL_NAME = os.getenv("COMET_MODEL_NAME", "yolov5")
COMET_DEFAULT_CHECKPOINT_FILENAME = os.getenv("COMET_DEFAULT_CHECKPOINT_FILENAME", "last.pt")
def download_model_checkpoint(opt, experiment):
model_dir = f"{opt.project}/{experiment.name}"
os.makedirs(model_dir, exist_ok=True)
model_name = COMET_MODEL_NAME
model_asset_list = experiment.get_model_asset_list(model_name)
if len(model_asset_list) == 0:
logger.error(f"COMET ERROR: No checkpoints found for model name : {model_name}")
return
model_asset_list = sorted(
model_asset_list,
key=lambda x: x["step"],
reverse=True,
)
logged_checkpoint_map = {asset["fileName"]: asset["assetId"] for asset in model_asset_list}
resource_url = urlparse(opt.weights)
checkpoint_filename = resource_url.query
if checkpoint_filename:
asset_id = logged_checkpoint_map.get(checkpoint_filename)
else:
asset_id = logged_checkpoint_map.get(COMET_DEFAULT_CHECKPOINT_FILENAME)
checkpoint_filename = COMET_DEFAULT_CHECKPOINT_FILENAME
if asset_id is None:
logger.error(f"COMET ERROR: Checkpoint {checkpoint_filename} not found in the given Experiment")
return
try:
logger.info(f"COMET INFO: Downloading checkpoint {checkpoint_filename}")
asset_filename = checkpoint_filename
model_binary = experiment.get_asset(asset_id, return_type="binary", stream=False)
model_download_path = f"{model_dir}/{asset_filename}"
with open(model_download_path, "wb") as f:
f.write(model_binary)
opt.weights = model_download_path
except Exception as e:
logger.warning("COMET WARNING: Unable to download checkpoint from Comet")
logger.exception(e)
def set_opt_parameters(opt, experiment):
"""Update the opts Namespace with parameters
from Comet's ExistingExperiment when resuming a run
Args:
opt (argparse.Namespace): Namespace of command line options
experiment (comet_ml.APIExperiment): Comet API Experiment object
"""
asset_list = experiment.get_asset_list()
resume_string = opt.resume
for asset in asset_list:
if asset["fileName"] == "opt.yaml":
asset_id = asset["assetId"]
asset_binary = experiment.get_asset(asset_id, return_type="binary", stream=False)
opt_dict = yaml.safe_load(asset_binary)
for key, value in opt_dict.items():
setattr(opt, key, value)
opt.resume = resume_string
# Save hyperparameters to YAML file
# Necessary to pass checks in training script
save_dir = f"{opt.project}/{experiment.name}"
os.makedirs(save_dir, exist_ok=True)
hyp_yaml_path = f"{save_dir}/hyp.yaml"
with open(hyp_yaml_path, "w") as f:
yaml.dump(opt.hyp, f)
opt.hyp = hyp_yaml_path
def check_comet_weights(opt):
"""Downloads model weights from Comet and updates the
weights path to point to saved weights location
Args:
opt (argparse.Namespace): Command Line arguments passed
to YOLOv5 training script
Returns:
None/bool: Return True if weights are successfully downloaded
else return None
"""
if comet_ml is None:
return
if isinstance(opt.weights, str):
if opt.weights.startswith(COMET_PREFIX):
api = comet_ml.API()
resource = urlparse(opt.weights)
experiment_path = f"{resource.netloc}{resource.path}"
experiment = api.get(experiment_path)
download_model_checkpoint(opt, experiment)
return True
return None
def check_comet_resume(opt):
"""Restores run parameters to its original state based on the model checkpoint
and logged Experiment parameters.
Args:
opt (argparse.Namespace): Command Line arguments passed
to YOLOv5 training script
Returns:
None/bool: Return True if the run is restored successfully
else return None
"""
if comet_ml is None:
return
if isinstance(opt.resume, str):
if opt.resume.startswith(COMET_PREFIX):
api = comet_ml.API()
resource = urlparse(opt.resume)
experiment_path = f"{resource.netloc}{resource.path}"
experiment = api.get(experiment_path)
set_opt_parameters(opt, experiment)
download_model_checkpoint(opt, experiment)
return True
return None
utils/loggers/comet/hpo.py 0 → 100644
View file @ cf1ed55a
import argparse
import json
import logging
import os
import sys
from pathlib import Path
import comet_ml
logger = logging.getLogger(__name__)
FILE = Path(__file__).resolve()
ROOT = FILE.parents[3] # YOLOv5 root directory
if str(ROOT) not in sys.path:
sys.path.append(str(ROOT)) # add ROOT to PATH
from train import train
from utils.callbacks import Callbacks
from utils.general import increment_path
from utils.torch_utils import select_device
# Project Configuration
config = comet_ml.config.get_config()
COMET_PROJECT_NAME = config.get_string(os.getenv("COMET_PROJECT_NAME"), "comet.project_name", default="yolov5")
def get_args(known=False):
parser = argparse.ArgumentParser()
parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')
parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
parser.add_argument('--epochs', type=int, default=300, help='total training epochs')
parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
parser.add_argument('--rect', action='store_true', help='rectangular training')
parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
parser.add_argument('--noval', action='store_true', help='only validate final epoch')
parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
parser.add_argument('--noplots', action='store_true', help='save no plot files')
parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"')
parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer')
parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
parser.add_argument('--name', default='exp', help='save to project/name')
parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
parser.add_argument('--quad', action='store_true', help='quad dataloader')
parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
parser.add_argument('--seed', type=int, default=0, help='Global training seed')
parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
# Weights & Biases arguments
parser.add_argument('--entity', default=None, help='W&B: Entity')
parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='W&B: Upload data, "val" option')
parser.add_argument('--bbox_interval', type=int, default=-1, help='W&B: Set bounding-box image logging interval')
parser.add_argument('--artifact_alias', type=str, default='latest', help='W&B: Version of dataset artifact to use')
# Comet Arguments
parser.add_argument("--comet_optimizer_config", type=str, help="Comet: Path to a Comet Optimizer Config File.")
parser.add_argument("--comet_optimizer_id", type=str, help="Comet: ID of the Comet Optimizer sweep.")
parser.add_argument("--comet_optimizer_objective", type=str, help="Comet: Set to 'minimize' or 'maximize'.")
parser.add_argument("--comet_optimizer_metric", type=str, help="Comet: Metric to Optimize.")
parser.add_argument("--comet_optimizer_workers",
type=int,
default=1,
help="Comet: Number of Parallel Workers to use with the Comet Optimizer.")
return parser.parse_known_args()[0] if known else parser.parse_args()
def run(parameters, opt):
hyp_dict = {k: v for k, v in parameters.items() if k not in ["epochs", "batch_size"]}
opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok or opt.evolve))
opt.batch_size = parameters.get("batch_size")
opt.epochs = parameters.get("epochs")
device = select_device(opt.device, batch_size=opt.batch_size)
train(hyp_dict, opt, device, callbacks=Callbacks())
if __name__ == "__main__":
opt = get_args(known=True)
opt.weights = str(opt.weights)
opt.cfg = str(opt.cfg)
opt.data = str(opt.data)
opt.project = str(opt.project)
optimizer_id = os.getenv("COMET_OPTIMIZER_ID")
if optimizer_id is None:
with open(opt.comet_optimizer_config) as f:
optimizer_config = json.load(f)
optimizer = comet_ml.Optimizer(optimizer_config)
else:
optimizer = comet_ml.Optimizer(optimizer_id)
opt.comet_optimizer_id = optimizer.id
status = optimizer.status()
opt.comet_optimizer_objective = status["spec"]["objective"]
opt.comet_optimizer_metric = status["spec"]["metric"]
logger.info("COMET INFO: Starting Hyperparameter Sweep")
for parameter in optimizer.get_parameters():
run(parameter["parameters"], opt)
utils/loggers/wandb/README.md 0 → 100644
View file @ cf1ed55a
📚 This guide explains how to use **Weights & Biases** (W&B) with YOLOv5 🚀. UPDATED 29 September 2021.
- [About Weights & Biases](#about-weights-&-biases)
- [First-Time Setup](#first-time-setup)
- [Viewing runs](#viewing-runs)
- [Disabling wandb](#disabling-wandb)
- [Advanced Usage: Dataset Versioning and Evaluation](#advanced-usage)
- [Reports: Share your work with the world!](#reports)
## About Weights & Biases
Think of [W&B](https://wandb.ai/site?utm_campaign=repo_yolo_wandbtutorial) like GitHub for machine learning models. With a few lines of code, save everything you need to debug, compare and reproduce your models — architecture, hyperparameters, git commits, model weights, GPU usage, and even datasets and predictions.
Used by top researchers including teams at OpenAI, Lyft, Github, and MILA, W&B is part of the new standard of best practices for machine learning. How W&B can help you optimize your machine learning workflows:
- [Debug](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Free-2) model performance in real time
- [GPU usage](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#System-4) visualized automatically
- [Custom charts](https://wandb.ai/wandb/customizable-charts/reports/Powerful-Custom-Charts-To-Debug-Model-Peformance--VmlldzoyNzY4ODI) for powerful, extensible visualization
- [Share insights](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Share-8) interactively with collaborators
- [Optimize hyperparameters](https://docs.wandb.com/sweeps) efficiently
- [Track](https://docs.wandb.com/artifacts) datasets, pipelines, and production models
## First-Time Setup
<details open>
<summary> Toggle Details </summary>
When you first train, W&B will prompt you to create a new account and will generate an **API key** for you. If you are an existing user you can retrieve your key from https://wandb.ai/authorize. This key is used to tell W&B where to log your data. You only need to supply your key once, and then it is remembered on the same device.
W&B will create a cloud **project** (default is 'YOLOv5') for your training runs, and each new training run will be provided a unique run **name** within that project as project/name. You can also manually set your project and run name as:
```shell
$ python train.py --project ... --name ...
```
YOLOv5 notebook example: <a href="https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a> <a href="https://www.kaggle.com/ultralytics/yolov5"><img src="https://kaggle.com/static/images/open-in-kaggle.svg" alt="Open In Kaggle"></a>
<img width="960" alt="Screen Shot 2021-09-29 at 10 23 13 PM" src="https://user-images.githubusercontent.com/26833433/135392431-1ab7920a-c49d-450a-b0b0-0c86ec86100e.png">
</details>
## Viewing Runs
<details open>
<summary> Toggle Details </summary>
Run information streams from your environment to the W&B cloud console as you train. This allows you to monitor and even cancel runs in <b>realtime</b> . All important information is logged:
- Training & Validation losses
- Metrics: Precision, Recall, mAP@0.5, mAP@0.5:0.95
- Learning Rate over time
- A bounding box debugging panel, showing the training progress over time
- GPU: Type, **GPU Utilization**, power, temperature, **CUDA memory usage**
- System: Disk I/0, CPU utilization, RAM memory usage
- Your trained model as W&B Artifact
- Environment: OS and Python types, Git repository and state, **training command**
<p align="center"><img width="900" alt="Weights & Biases dashboard" src="https://user-images.githubusercontent.com/26833433/135390767-c28b050f-8455-4004-adb0-3b730386e2b2.png"></p>
</details>
## Disabling wandb
- training after running `wandb disabled` inside that directory creates no wandb run
![Screenshot (84)](https://user-images.githubusercontent.com/15766192/143441777-c780bdd7-7cb4-4404-9559-b4316030a985.png)
- To enable wandb again, run `wandb online`
![Screenshot (85)](https://user-images.githubusercontent.com/15766192/143441866-7191b2cb-22f0-4e0f-ae64-2dc47dc13078.png)
## Advanced Usage
You can leverage W&B artifacts and Tables integration to easily visualize and manage your datasets, models and training evaluations. Here are some quick examples to get you started.
<details open>
<h3> 1: Train and Log Evaluation simultaneousy </h3>
This is an extension of the previous section, but it'll also training after uploading the dataset. <b> This also evaluation Table</b>
Evaluation table compares your predictions and ground truths across the validation set for each epoch. It uses the references to the already uploaded datasets,
so no images will be uploaded from your system more than once.
<details open>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python train.py --upload_data val</code>
![Screenshot from 2021-11-21 17-40-06](https://user-images.githubusercontent.com/15766192/142761183-c1696d8c-3f38-45ab-991a-bb0dfd98ae7d.png)
</details>
<h3>2. Visualize and Version Datasets</h3>
Log, visualize, dynamically query, and understand your data with <a href='https://docs.wandb.ai/guides/data-vis/tables'>W&B Tables</a>. You can use the following command to log your dataset as a W&B Table. This will generate a <code>{dataset}_wandb.yaml</code> file which can be used to train from dataset artifact.
<details>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python utils/logger/wandb/log_dataset.py --project ... --name ... --data .. </code>
![Screenshot (64)](https://user-images.githubusercontent.com/15766192/128486078-d8433890-98a3-4d12-8986-b6c0e3fc64b9.png)
</details>
<h3> 3: Train using dataset artifact </h3>
When you upload a dataset as described in the first section, you get a new config file with an added `_wandb` to its name. This file contains the information that
can be used to train a model directly from the dataset artifact. <b> This also logs evaluation </b>
<details>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python train.py --data {data}_wandb.yaml </code>
![Screenshot (72)](https://user-images.githubusercontent.com/15766192/128979739-4cf63aeb-a76f-483f-8861-1c0100b938a5.png)
</details>
<h3> 4: Save model checkpoints as artifacts </h3>
To enable saving and versioning checkpoints of your experiment, pass `--save_period n` with the base cammand, where `n` represents checkpoint interval.
You can also log both the dataset and model checkpoints simultaneously. If not passed, only the final model will be logged
<details>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python train.py --save_period 1 </code>
![Screenshot (68)](https://user-images.githubusercontent.com/15766192/128726138-ec6c1f60-639d-437d-b4ee-3acd9de47ef3.png)
</details>
</details>
<h3> 5: Resume runs from checkpoint artifacts. </h3>
Any run can be resumed using artifacts if the <code>--resume</code> argument starts with <code>wandb-artifact://</code> prefix followed by the run path, i.e, <code>wandb-artifact://username/project/runid </code>. This doesn't require the model checkpoint to be present on the local system.
<details>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python train.py --resume wandb-artifact://{run_path} </code>
![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png)
</details>
<h3> 6: Resume runs from dataset artifact & checkpoint artifacts. </h3>
<b> Local dataset or model checkpoints are not required. This can be used to resume runs directly on a different device </b>
The syntax is same as the previous section, but you'll need to lof both the dataset and model checkpoints as artifacts, i.e, set bot <code>--upload_dataset</code> or
train from <code>_wandb.yaml</code> file and set <code>--save_period</code>
<details>
<summary> <b>Usage</b> </summary>
<b>Code</b> <code> $ python train.py --resume wandb-artifact://{run_path} </code>
![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png)
</details>
</details>
<h3> Reports </h3>
W&B Reports can be created from your saved runs for sharing online. Once a report is created you will receive a link you can use to publically share your results. Here is an example report created from the COCO128 tutorial trainings of all four YOLOv5 models ([link](https://wandb.ai/glenn-jocher/yolov5_tutorial/reports/YOLOv5-COCO128-Tutorial-Results--VmlldzozMDI5OTY)).
<img width="900" alt="Weights & Biases Reports" src="https://user-images.githubusercontent.com/26833433/135394029-a17eaf86-c6c1-4b1d-bb80-b90e83aaffa7.png">
## Environments
YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled):
- **Google Colab and Kaggle** notebooks with free GPU: <a href="https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a> <a href="https://www.kaggle.com/ultralytics/yolov5"><img src="https://kaggle.com/static/images/open-in-kaggle.svg" alt="Open In Kaggle"></a>
- **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)
- **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart)
- **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) <a href="https://hub.docker.com/r/ultralytics/yolov5"><img src="https://img.shields.io/docker/pulls/ultralytics/yolov5?logo=docker" alt="Docker Pulls"></a>
## Status
![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg)
If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), validation ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit.
utils/loggers/wandb/log_dataset.py 0 → 100644
View file @ cf1ed55a
import argparse
from wandb_utils import WandbLogger
from utils.general import LOGGER
WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
def create_dataset_artifact(opt):
logger = WandbLogger(opt, None, job_type='Dataset Creation') # TODO: return value unused
if not logger.wandb:
LOGGER.info("install wandb using `pip install wandb` to log the dataset")
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path')
parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project')
parser.add_argument('--entity', default=None, help='W&B entity')
parser.add_argument('--name', type=str, default='log dataset', help='name of W&B run')
opt = parser.parse_args()
opt.resume = False # Explicitly disallow resume check for dataset upload job
create_dataset_artifact(opt)
utils/loggers/wandb/sweep.py 0 → 100644
View file @ cf1ed55a
import sys
from pathlib import Path
import wandb
FILE = Path(__file__).resolve()
ROOT = FILE.parents[3] # YOLOv5 root directory
if str(ROOT) not in sys.path:
sys.path.append(str(ROOT)) # add ROOT to PATH
from train import parse_opt, train
from utils.callbacks import Callbacks
from utils.general import increment_path
from utils.torch_utils import select_device
def sweep():
wandb.init()
# Get hyp dict from sweep agent. Copy because train() modifies parameters which confused wandb.
hyp_dict = vars(wandb.config).get("_items").copy()
# Workaround: get necessary opt args
opt = parse_opt(known=True)
opt.batch_size = hyp_dict.get("batch_size")
opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok or opt.evolve))
opt.epochs = hyp_dict.get("epochs")
opt.nosave = True
opt.data = hyp_dict.get("data")
opt.weights = str(opt.weights)
opt.cfg = str(opt.cfg)
opt.data = str(opt.data)
opt.hyp = str(opt.hyp)
opt.project = str(opt.project)
device = select_device(opt.device, batch_size=opt.batch_size)
# train
train(hyp_dict, opt, device, callbacks=Callbacks())
if __name__ == "__main__":
sweep()
utils/loggers/wandb/sweep.yaml 0 → 100644
View file @ cf1ed55a
# Hyperparameters for training
# To set range-
# Provide min and max values as:
# parameter:
#
# min: scalar
# max: scalar
# OR
#
# Set a specific list of search space-
# parameter:
# values: [scalar1, scalar2, scalar3...]
#
# You can use grid, bayesian and hyperopt search strategy
# For more info on configuring sweeps visit - https://docs.wandb.ai/guides/sweeps/configuration
program: utils/loggers/wandb/sweep.py
method: random
metric:
name: metrics/mAP_0.5
goal: maximize
parameters:
# hyperparameters: set either min, max range or values list
data:
value: "data/coco128.yaml"
batch_size:
values: [64]
epochs:
values: [10]
lr0:
distribution: uniform
min: 1e-5
max: 1e-1
lrf:
distribution: uniform
min: 0.01
max: 1.0
momentum:
distribution: uniform
min: 0.6
max: 0.98
weight_decay:
distribution: uniform
min: 0.0
max: 0.001
warmup_epochs:
distribution: uniform
min: 0.0
max: 5.0
warmup_momentum:
distribution: uniform
min: 0.0
max: 0.95
warmup_bias_lr:
distribution: uniform
min: 0.0
max: 0.2
box:
distribution: uniform
min: 0.02
max: 0.2
cls:
distribution: uniform
min: 0.2
max: 4.0
cls_pw:
distribution: uniform
min: 0.5
max: 2.0
obj:
distribution: uniform
min: 0.2
max: 4.0
obj_pw:
distribution: uniform
min: 0.5
max: 2.0
iou_t:
distribution: uniform
min: 0.1
max: 0.7
anchor_t:
distribution: uniform
min: 2.0
max: 8.0
fl_gamma:
distribution: uniform
min: 0.0
max: 4.0
hsv_h:
distribution: uniform
min: 0.0
max: 0.1
hsv_s:
distribution: uniform
min: 0.0
max: 0.9
hsv_v:
distribution: uniform
min: 0.0
max: 0.9
degrees:
distribution: uniform
min: 0.0
max: 45.0
translate:
distribution: uniform
min: 0.0
max: 0.9
scale:
distribution: uniform
min: 0.0
max: 0.9
shear:
distribution: uniform
min: 0.0
max: 10.0
perspective:
distribution: uniform
min: 0.0
max: 0.001
flipud:
distribution: uniform
min: 0.0
max: 1.0
fliplr:
distribution: uniform
min: 0.0
max: 1.0
mosaic:
distribution: uniform
min: 0.0
max: 1.0
mixup:
distribution: uniform
min: 0.0
max: 1.0
copy_paste:
distribution: uniform
min: 0.0
max: 1.0
utils/loggers/wandb/wandb_utils.py 0 → 100644
View file @ cf1ed55a
"""Utilities and tools for tracking runs with Weights & Biases."""
import logging
import os
import sys
from contextlib import contextmanager
from pathlib import Path
from typing import Dict
import yaml
from tqdm import tqdm
FILE = Path(__file__).resolve()
ROOT = FILE.parents[3] # YOLOv5 root directory
if str(ROOT) not in sys.path:
sys.path.append(str(ROOT)) # add ROOT to PATH
from utils.dataloaders import LoadImagesAndLabels, img2label_paths
from utils.general import LOGGER, check_dataset, check_file
try:
import wandb
assert hasattr(wandb, '__version__') # verify package import not local dir
except (ImportError, AssertionError):
wandb = None
RANK = int(os.getenv('RANK', -1))
WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX):
return from_string[len(prefix):]
def check_wandb_config_file(data_config_file):
wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1)) # updated data.yaml path
if Path(wandb_config).is_file():
return wandb_config
return data_config_file
def check_wandb_dataset(data_file):
is_trainset_wandb_artifact = False
is_valset_wandb_artifact = False
if isinstance(data_file, dict):
# In that case another dataset manager has already processed it and we don't have to
return data_file
if check_file(data_file) and data_file.endswith('.yaml'):
with open(data_file, errors='ignore') as f:
data_dict = yaml.safe_load(f)
is_trainset_wandb_artifact = isinstance(data_dict['train'],
str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX)
is_valset_wandb_artifact = isinstance(data_dict['val'],
str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX)
if is_trainset_wandb_artifact or is_valset_wandb_artifact:
return data_dict
else:
return check_dataset(data_file)
def get_run_info(run_path):
run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX))
run_id = run_path.stem
project = run_path.parent.stem
entity = run_path.parent.parent.stem
model_artifact_name = 'run_' + run_id + '_model'
return entity, project, run_id, model_artifact_name
def check_wandb_resume(opt):
process_wandb_config_ddp_mode(opt) if RANK not in [-1, 0] else None
if isinstance(opt.resume, str):
if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
if RANK not in [-1, 0]: # For resuming DDP runs
entity, project, run_id, model_artifact_name = get_run_info(opt.resume)
api = wandb.Api()
artifact = api.artifact(entity + '/' + project + '/' + model_artifact_name + ':latest')
modeldir = artifact.download()
opt.weights = str(Path(modeldir) / "last.pt")
return True
return None
def process_wandb_config_ddp_mode(opt):
with open(check_file(opt.data), errors='ignore') as f:
data_dict = yaml.safe_load(f) # data dict
train_dir, val_dir = None, None
if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX):
api = wandb.Api()
train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias)
train_dir = train_artifact.download()
train_path = Path(train_dir) / 'data/images/'
data_dict['train'] = str(train_path)
if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX):
api = wandb.Api()
val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias)
val_dir = val_artifact.download()
val_path = Path(val_dir) / 'data/images/'
data_dict['val'] = str(val_path)
if train_dir or val_dir:
ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml')
with open(ddp_data_path, 'w') as f:
yaml.safe_dump(data_dict, f)
opt.data = ddp_data_path
class WandbLogger():
"""Log training runs, datasets, models, and predictions to Weights & Biases.
This logger sends information to W&B at wandb.ai. By default, this information
includes hyperparameters, system configuration and metrics, model metrics,
and basic data metrics and analyses.
By providing additional command line arguments to train.py, datasets,
models and predictions can also be logged.
For more on how this logger is used, see the Weights & Biases documentation:
https://docs.wandb.com/guides/integrations/yolov5
"""
def __init__(self, opt, run_id=None, job_type='Training'):
"""
- Initialize WandbLogger instance
- Upload dataset if opt.upload_dataset is True
- Setup training processes if job_type is 'Training'
arguments:
opt (namespace) -- Commandline arguments for this run
run_id (str) -- Run ID of W&B run to be resumed
job_type (str) -- To set the job_type for this run
"""
# Temporary-fix
if opt.upload_dataset:
opt.upload_dataset = False
# LOGGER.info("Uploading Dataset functionality is not being supported temporarily due to a bug.")
# Pre-training routine --
self.job_type = job_type
self.wandb, self.wandb_run = wandb, None if not wandb else wandb.run
self.val_artifact, self.train_artifact = None, None
self.train_artifact_path, self.val_artifact_path = None, None
self.result_artifact = None
self.val_table, self.result_table = None, None
self.bbox_media_panel_images = []
self.val_table_path_map = None
self.max_imgs_to_log = 16
self.wandb_artifact_data_dict = None
self.data_dict = None
# It's more elegant to stick to 1 wandb.init call,
# but useful config data is overwritten in the WandbLogger's wandb.init call
if isinstance(opt.resume, str): # checks resume from artifact
if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
entity, project, run_id, model_artifact_name = get_run_info(opt.resume)
model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name
assert wandb, 'install wandb to resume wandb runs'
# Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config
self.wandb_run = wandb.init(id=run_id,
project=project,
entity=entity,
resume='allow',
allow_val_change=True)
opt.resume = model_artifact_name
elif self.wandb:
self.wandb_run = wandb.init(config=opt,
resume="allow",
project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem,
entity=opt.entity,
name=opt.name if opt.name != 'exp' else None,
job_type=job_type,
id=run_id,
allow_val_change=True) if not wandb.run else wandb.run
if self.wandb_run:
if self.job_type == 'Training':
if opt.upload_dataset:
if not opt.resume:
self.wandb_artifact_data_dict = self.check_and_upload_dataset(opt)
if isinstance(opt.data, dict):
# This means another dataset manager has already processed the dataset info (e.g. ClearML)
# and they will have stored the already processed dict in opt.data
self.data_dict = opt.data
elif opt.resume:
# resume from artifact
if isinstance(opt.resume, str) and opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
self.data_dict = dict(self.wandb_run.config.data_dict)
else: # local resume
self.data_dict = check_wandb_dataset(opt.data)
else:
self.data_dict = check_wandb_dataset(opt.data)
self.wandb_artifact_data_dict = self.wandb_artifact_data_dict or self.data_dict
# write data_dict to config. useful for resuming from artifacts. Do this only when not resuming.
self.wandb_run.config.update({'data_dict': self.wandb_artifact_data_dict}, allow_val_change=True)
self.setup_training(opt)
if self.job_type == 'Dataset Creation':
self.wandb_run.config.update({"upload_dataset": True})
self.data_dict = self.check_and_upload_dataset(opt)
def check_and_upload_dataset(self, opt):
"""
Check if the dataset format is compatible and upload it as W&B artifact
arguments:
opt (namespace)-- Commandline arguments for current run
returns:
Updated dataset info dictionary where local dataset paths are replaced by WAND_ARFACT_PREFIX links.
"""
assert wandb, 'Install wandb to upload dataset'
config_path = self.log_dataset_artifact(opt.data, opt.single_cls,
'YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem)
with open(config_path, errors='ignore') as f:
wandb_data_dict = yaml.safe_load(f)
return wandb_data_dict
def setup_training(self, opt):
"""
Setup the necessary processes for training YOLO models:
- Attempt to download model checkpoint and dataset artifacts if opt.resume stats with WANDB_ARTIFACT_PREFIX
- Update data_dict, to contain info of previous run if resumed and the paths of dataset artifact if downloaded
- Setup log_dict, initialize bbox_interval
arguments:
opt (namespace) -- commandline arguments for this run
"""
self.log_dict, self.current_epoch = {}, 0
self.bbox_interval = opt.bbox_interval
if isinstance(opt.resume, str):
modeldir, _ = self.download_model_artifact(opt)
if modeldir:
self.weights = Path(modeldir) / "last.pt"
config = self.wandb_run.config
opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp, opt.imgsz = str(
self.weights), config.save_period, config.batch_size, config.bbox_interval, config.epochs,\
config.hyp, config.imgsz
data_dict = self.data_dict
if self.val_artifact is None: # If --upload_dataset is set, use the existing artifact, don't download
self.train_artifact_path, self.train_artifact = self.download_dataset_artifact(
data_dict.get('train'), opt.artifact_alias)
self.val_artifact_path, self.val_artifact = self.download_dataset_artifact(
data_dict.get('val'), opt.artifact_alias)
if self.train_artifact_path is not None:
train_path = Path(self.train_artifact_path) / 'data/images/'
data_dict['train'] = str(train_path)
if self.val_artifact_path is not None:
val_path = Path(self.val_artifact_path) / 'data/images/'
data_dict['val'] = str(val_path)
if self.val_artifact is not None:
self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
columns = ["epoch", "id", "ground truth", "prediction"]
columns.extend(self.data_dict['names'])
self.result_table = wandb.Table(columns)
self.val_table = self.val_artifact.get("val")
if self.val_table_path_map is None:
self.map_val_table_path()
if opt.bbox_interval == -1:
self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1
if opt.evolve or opt.noplots:
self.bbox_interval = opt.bbox_interval = opt.epochs + 1 # disable bbox_interval
train_from_artifact = self.train_artifact_path is not None and self.val_artifact_path is not None
# Update the the data_dict to point to local artifacts dir
if train_from_artifact:
self.data_dict = data_dict
def download_dataset_artifact(self, path, alias):
"""
download the model checkpoint artifact if the path starts with WANDB_ARTIFACT_PREFIX
arguments:
path -- path of the dataset to be used for training
alias (str)-- alias of the artifact to be download/used for training
returns:
(str, wandb.Artifact) -- path of the downladed dataset and it's corresponding artifact object if dataset
is found otherwise returns (None, None)
"""
if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX):
artifact_path = Path(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias)
dataset_artifact = wandb.use_artifact(artifact_path.as_posix().replace("\\", "/"))
assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'"
datadir = dataset_artifact.download()
return datadir, dataset_artifact
return None, None
def download_model_artifact(self, opt):
"""
download the model checkpoint artifact if the resume path starts with WANDB_ARTIFACT_PREFIX
arguments:
opt (namespace) -- Commandline arguments for this run
"""
if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest")
assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist'
modeldir = model_artifact.download()
# epochs_trained = model_artifact.metadata.get('epochs_trained')
total_epochs = model_artifact.metadata.get('total_epochs')
is_finished = total_epochs is None
assert not is_finished, 'training is finished, can only resume incomplete runs.'
return modeldir, model_artifact
return None, None
def log_model(self, path, opt, epoch, fitness_score, best_model=False):
"""
Log the model checkpoint as W&B artifact
arguments:
path (Path) -- Path of directory containing the checkpoints
opt (namespace) -- Command line arguments for this run
epoch (int) -- Current epoch number
fitness_score (float) -- fitness score for current epoch
best_model (boolean) -- Boolean representing if the current checkpoint is the best yet.
"""
model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model',
type='model',
metadata={
'original_url': str(path),
'epochs_trained': epoch + 1,
'save period': opt.save_period,
'project': opt.project,
'total_epochs': opt.epochs,
'fitness_score': fitness_score})
model_artifact.add_file(str(path / 'last.pt'), name='last.pt')
wandb.log_artifact(model_artifact,
aliases=['latest', 'last', 'epoch ' + str(self.current_epoch), 'best' if best_model else ''])
LOGGER.info(f"Saving model artifact on epoch {epoch + 1}")
def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False):
"""
Log the dataset as W&B artifact and return the new data file with W&B links
arguments:
data_file (str) -- the .yaml file with information about the dataset like - path, classes etc.
single_class (boolean) -- train multi-class data as single-class
project (str) -- project name. Used to construct the artifact path
overwrite_config (boolean) -- overwrites the data.yaml file if set to true otherwise creates a new
file with _wandb postfix. Eg -> data_wandb.yaml
returns:
the new .yaml file with artifact links. it can be used to start training directly from artifacts
"""
upload_dataset = self.wandb_run.config.upload_dataset
log_val_only = isinstance(upload_dataset, str) and upload_dataset == 'val'
self.data_dict = check_dataset(data_file) # parse and check
data = dict(self.data_dict)
nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names'])
names = {k: v for k, v in enumerate(names)} # to index dictionary
# log train set
if not log_val_only:
self.train_artifact = self.create_dataset_table(LoadImagesAndLabels(data['train'], rect=True, batch_size=1),
names,
name='train') if data.get('train') else None
if data.get('train'):
data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train')
self.val_artifact = self.create_dataset_table(
LoadImagesAndLabels(data['val'], rect=True, batch_size=1), names, name='val') if data.get('val') else None
if data.get('val'):
data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val')
path = Path(data_file)
# create a _wandb.yaml file with artifacts links if both train and test set are logged
if not log_val_only:
path = (path.stem if overwrite_config else path.stem + '_wandb') + '.yaml' # updated data.yaml path
path = ROOT / 'data' / path
data.pop('download', None)
data.pop('path', None)
with open(path, 'w') as f:
yaml.safe_dump(data, f)
LOGGER.info(f"Created dataset config file {path}")
if self.job_type == 'Training': # builds correct artifact pipeline graph
if not log_val_only:
self.wandb_run.log_artifact(
self.train_artifact) # calling use_artifact downloads the dataset. NOT NEEDED!
self.wandb_run.use_artifact(self.val_artifact)
self.val_artifact.wait()
self.val_table = self.val_artifact.get('val')
self.map_val_table_path()
else:
self.wandb_run.log_artifact(self.train_artifact)
self.wandb_run.log_artifact(self.val_artifact)
return path
def map_val_table_path(self):
"""
Map the validation dataset Table like name of file -> it's id in the W&B Table.
Useful for - referencing artifacts for evaluation.
"""
self.val_table_path_map = {}
LOGGER.info("Mapping dataset")
for i, data in enumerate(tqdm(self.val_table.data)):
self.val_table_path_map[data[3]] = data[0]
def create_dataset_table(self, dataset: LoadImagesAndLabels, class_to_id: Dict[int, str], name: str = 'dataset'):
"""
Create and return W&B artifact containing W&B Table of the dataset.
arguments:
dataset -- instance of LoadImagesAndLabels class used to iterate over the data to build Table
class_to_id -- hash map that maps class ids to labels
name -- name of the artifact
returns:
dataset artifact to be logged or used
"""
# TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging
artifact = wandb.Artifact(name=name, type="dataset")
img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None
img_files = tqdm(dataset.im_files) if not img_files else img_files
for img_file in img_files:
if Path(img_file).is_dir():
artifact.add_dir(img_file, name='data/images')
labels_path = 'labels'.join(dataset.path.rsplit('images', 1))
artifact.add_dir(labels_path, name='data/labels')
else:
artifact.add_file(img_file, name='data/images/' + Path(img_file).name)
label_file = Path(img2label_paths([img_file])[0])
artifact.add_file(str(label_file), name='data/labels/' +
label_file.name) if label_file.exists() else None
table = wandb.Table(columns=["id", "train_image", "Classes", "name"])
class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()])
for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)):
box_data, img_classes = [], {}
for cls, *xywh in labels[:, 1:].tolist():
cls = int(cls)
box_data.append({
"position": {
"middle": [xywh[0], xywh[1]],
"width": xywh[2],
"height": xywh[3]},
"class_id": cls,
"box_caption": "%s" % (class_to_id[cls])})
img_classes[cls] = class_to_id[cls]
boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space
table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), list(img_classes.values()),
Path(paths).name)
artifact.add(table, name)
return artifact
def log_training_progress(self, predn, path, names):
"""
Build evaluation Table. Uses reference from validation dataset table.
arguments:
predn (list): list of predictions in the native space in the format - [xmin, ymin, xmax, ymax, confidence, class]
path (str): local path of the current evaluation image
names (dict(int, str)): hash map that maps class ids to labels
"""
class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()])
box_data = []
avg_conf_per_class = [0] * len(self.data_dict['names'])
pred_class_count = {}
for *xyxy, conf, cls in predn.tolist():
if conf >= 0.25:
cls = int(cls)
box_data.append({
"position": {
"minX": xyxy[0],
"minY": xyxy[1],
"maxX": xyxy[2],
"maxY": xyxy[3]},
"class_id": cls,
"box_caption": f"{names[cls]} {conf:.3f}",
"scores": {
"class_score": conf},
"domain": "pixel"})
avg_conf_per_class[cls] += conf
if cls in pred_class_count:
pred_class_count[cls] += 1
else:
pred_class_count[cls] = 1
for pred_class in pred_class_count.keys():
avg_conf_per_class[pred_class] = avg_conf_per_class[pred_class] / pred_class_count[pred_class]
boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space
id = self.val_table_path_map[Path(path).name]
self.result_table.add_data(self.current_epoch, id, self.val_table.data[id][1],
wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set),
*avg_conf_per_class)
def val_one_image(self, pred, predn, path, names, im):
"""
Log validation data for one image. updates the result Table if validation dataset is uploaded and log bbox media panel
arguments:
pred (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class]
predn (list): list of predictions in the native space - [xmin, ymin, xmax, ymax, confidence, class]
path (str): local path of the current evaluation image
"""
if self.val_table and self.result_table: # Log Table if Val dataset is uploaded as artifact
self.log_training_progress(predn, path, names)
if len(self.bbox_media_panel_images) < self.max_imgs_to_log and self.current_epoch > 0:
if self.current_epoch % self.bbox_interval == 0:
box_data = [{
"position": {
"minX": xyxy[0],
"minY": xyxy[1],
"maxX": xyxy[2],
"maxY": xyxy[3]},
"class_id": int(cls),
"box_caption": f"{names[int(cls)]} {conf:.3f}",
"scores": {
"class_score": conf},
"domain": "pixel"} for *xyxy, conf, cls in pred.tolist()]
boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space
self.bbox_media_panel_images.append(wandb.Image(im, boxes=boxes, caption=path.name))
def log(self, log_dict):
"""
save the metrics to the logging dictionary
arguments:
log_dict (Dict) -- metrics/media to be logged in current step
"""
if self.wandb_run:
for key, value in log_dict.items():
self.log_dict[key] = value
def end_epoch(self, best_result=False):
"""
commit the log_dict, model artifacts and Tables to W&B and flush the log_dict.
arguments:
best_result (boolean): Boolean representing if the result of this evaluation is best or not
"""
if self.wandb_run:
with all_logging_disabled():
if self.bbox_media_panel_images:
self.log_dict["BoundingBoxDebugger"] = self.bbox_media_panel_images
try:
wandb.log(self.log_dict)
except BaseException as e:
LOGGER.info(
f"An error occurred in wandb logger. The training will proceed without interruption. More info\n{e}"
)
self.wandb_run.finish()
self.wandb_run = None
self.log_dict = {}
self.bbox_media_panel_images = []
if self.result_artifact:
self.result_artifact.add(self.result_table, 'result')
wandb.log_artifact(self.result_artifact,
aliases=[
'latest', 'last', 'epoch ' + str(self.current_epoch),
('best' if best_result else '')])
wandb.log({"evaluation": self.result_table})
columns = ["epoch", "id", "ground truth", "prediction"]
columns.extend(self.data_dict['names'])
self.result_table = wandb.Table(columns)
self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
def finish_run(self):
"""
Log metrics if any and finish the current W&B run
"""
if self.wandb_run:
if self.log_dict:
with all_logging_disabled():
wandb.log(self.log_dict)
wandb.run.finish()
@contextmanager
def all_logging_disabled(highest_level=logging.CRITICAL):
""" source - https://gist.github.com/simon-weber/7853144
A context manager that will prevent any logging messages triggered during the body from being processed.
:param highest_level: the maximum logging level in use.
This would only need to be changed if a custom level greater than CRITICAL is defined.
"""
previous_level = logging.root.manager.disable
logging.disable(highest_level)
try:
yield
finally:
logging.disable(previous_level)
utils/loss.py 0 → 100644
View file @ cf1ed55a
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Loss functions
"""
import torch
import torch.nn as nn
from utils.metrics import bbox_iou
from utils.torch_utils import de_parallel
def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
# return positive, negative label smoothing BCE targets
return 1.0 - 0.5 * eps, 0.5 * eps
class BCEBlurWithLogitsLoss(nn.Module):
# BCEwithLogitLoss() with reduced missing label effects.
def __init__(self, alpha=0.05):
super().__init__()
self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss()
self.alpha = alpha
def forward(self, pred, true):
loss = self.loss_fcn(pred, true)
pred = torch.sigmoid(pred) # prob from logits
dx = pred - true # reduce only missing label effects
# dx = (pred - true).abs() # reduce missing label and false label effects
alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
loss *= alpha_factor
return loss.mean()
class FocalLoss(nn.Module):
# Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
super().__init__()
self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
self.gamma = gamma
self.alpha = alpha
self.reduction = loss_fcn.reduction
self.loss_fcn.reduction = 'none' # required to apply FL to each element
def forward(self, pred, true):
loss = self.loss_fcn(pred, true)
# p_t = torch.exp(-loss)
# loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability
# TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
pred_prob = torch.sigmoid(pred) # prob from logits
p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
modulating_factor = (1.0 - p_t) ** self.gamma
loss *= alpha_factor * modulating_factor
if self.reduction == 'mean':
return loss.mean()
elif self.reduction == 'sum':
return loss.sum()
else: # 'none'
return loss
class QFocalLoss(nn.Module):
# Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
super().__init__()
self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
self.gamma = gamma
self.alpha = alpha
self.reduction = loss_fcn.reduction
self.loss_fcn.reduction = 'none' # required to apply FL to each element
def forward(self, pred, true):
loss = self.loss_fcn(pred, true)
pred_prob = torch.sigmoid(pred) # prob from logits
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
modulating_factor = torch.abs(true - pred_prob) ** self.gamma
loss *= alpha_factor * modulating_factor
if self.reduction == 'mean':
return loss.mean()
elif self.reduction == 'sum':
return loss.sum()
else: # 'none'
return loss
class ComputeLoss:
sort_obj_iou = False
# Compute losses
def __init__(self, model, autobalance=False):
device = next(model.parameters()).device # get model device
h = model.hyp # hyperparameters
# Define criteria
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
# Focal loss
g = h['fl_gamma'] # focal loss gamma
if g > 0:
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
m = de_parallel(model).model[-1] # Detect() module
self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02]) # P3-P7
self.ssi = list(m.stride).index(16) if autobalance else 0 # stride 16 index
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
self.na = m.na # number of anchors
self.nc = m.nc # number of classes
self.nl = m.nl # number of layers
self.anchors = m.anchors
self.device = device
def __call__(self, p, targets): # predictions, targets
lcls = torch.zeros(1, device=self.device) # class loss
lbox = torch.zeros(1, device=self.device) # box loss
lobj = torch.zeros(1, device=self.device) # object loss
tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
# Losses
for i, pi in enumerate(p): # layer index, layer predictions
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device) # target obj
n = b.shape[0] # number of targets
if n:
# pxy, pwh, _, pcls = pi[b, a, gj, gi].tensor_split((2, 4, 5), dim=1) # faster, requires torch 1.8.0
pxy, pwh, _, pcls = pi[b, a, gj, gi].split((2, 2, 1, self.nc), 1) # target-subset of predictions
# Regression
pxy = pxy.sigmoid() * 2 - 0.5
pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i]
pbox = torch.cat((pxy, pwh), 1) # predicted box
iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze() # iou(prediction, target)
lbox += (1.0 - iou).mean() # iou loss
# Objectness
iou = iou.detach().clamp(0).type(tobj.dtype)
if self.sort_obj_iou:
j = iou.argsort()
b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j]
if self.gr < 1:
iou = (1.0 - self.gr) + self.gr * iou
tobj[b, a, gj, gi] = iou # iou ratio
# Classification
if self.nc > 1: # cls loss (only if multiple classes)
t = torch.full_like(pcls, self.cn, device=self.device) # targets
t[range(n), tcls[i]] = self.cp
lcls += self.BCEcls(pcls, t) # BCE
# Append targets to text file
# with open('targets.txt', 'a') as file:
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
obji = self.BCEobj(pi[..., 4], tobj)
lobj += obji * self.balance[i] # obj loss
if self.autobalance:
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
if self.autobalance:
self.balance = [x / self.balance[self.ssi] for x in self.balance]
lbox *= self.hyp['box']
lobj *= self.hyp['obj']
lcls *= self.hyp['cls']
bs = tobj.shape[0] # batch size
return (lbox + lobj + lcls) * bs, torch.cat((lbox, lobj, lcls)).detach()
def build_targets(self, p, targets):
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
na, nt = self.na, targets.shape[0] # number of anchors, targets
tcls, tbox, indices, anch = [], [], [], []
gain = torch.ones(7, device=self.device) # normalized to gridspace gain
ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None]), 2) # append anchor indices
g = 0.5 # bias
off = torch.tensor(
[
[0, 0],
[1, 0],
[0, 1],
[-1, 0],
[0, -1], # j,k,l,m
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
],
device=self.device).float() * g # offsets
for i in range(self.nl):
anchors, shape = self.anchors[i], p[i].shape
gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]] # xyxy gain
# Match targets to anchors
t = targets * gain # shape(3,n,7)
if nt:
# Matches
r = t[..., 4:6] / anchors[:, None] # wh ratio
j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t'] # compare
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
t = t[j] # filter
# Offsets
gxy = t[:, 2:4] # grid xy
gxi = gain[[2, 3]] - gxy # inverse
j, k = ((gxy % 1 < g) & (gxy > 1)).T
l, m = ((gxi % 1 < g) & (gxi > 1)).T
j = torch.stack((torch.ones_like(j), j, k, l, m))
t = t.repeat((5, 1, 1))[j]
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
else:
t = targets[0]
offsets = 0
# Define
bc, gxy, gwh, a = t.chunk(4, 1) # (image, class), grid xy, grid wh, anchors
a, (b, c) = a.long().view(-1), bc.long().T # anchors, image, class
gij = (gxy - offsets).long()
gi, gj = gij.T # grid indices
# Append
indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1))) # image, anchor, grid
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
anch.append(anchors[a]) # anchors
tcls.append(c) # class
return tcls, tbox, indices, anch
utils/metrics.py 0 → 100644
View file @ cf1ed55a
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Model validation metrics
"""
import math
import warnings
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import torch
from utils import TryExcept, threaded
def fitness(x):
# Model fitness as a weighted combination of metrics
w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
return (x[:, :4] * w).sum(1)
def smooth(y, f=0.05):
# Box filter of fraction f
nf = round(len(y) * f * 2) // 2 + 1 # number of filter elements (must be odd)
p = np.ones(nf // 2) # ones padding
yp = np.concatenate((p * y[0], y, p * y[-1]), 0) # y padded
return np.convolve(yp, np.ones(nf) / nf, mode='valid') # y-smoothed
def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16, prefix=""):
""" Compute the average precision, given the recall and precision curves.
Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
# Arguments
tp: True positives (nparray, nx1 or nx10).
conf: Objectness value from 0-1 (nparray).
pred_cls: Predicted object classes (nparray).
target_cls: True object classes (nparray).
plot: Plot precision-recall curve at mAP@0.5
save_dir: Plot save directory
# Returns
The average precision as computed in py-faster-rcnn.
"""
# Sort by objectness
i = np.argsort(-conf)
tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
# Find unique classes
unique_classes, nt = np.unique(target_cls, return_counts=True)
nc = unique_classes.shape[0] # number of classes, number of detections
# Create Precision-Recall curve and compute AP for each class
px, py = np.linspace(0, 1, 1000), [] # for plotting
ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
for ci, c in enumerate(unique_classes):
i = pred_cls == c
n_l = nt[ci] # number of labels
n_p = i.sum() # number of predictions
if n_p == 0 or n_l == 0:
continue
# Accumulate FPs and TPs
fpc = (1 - tp[i]).cumsum(0)
tpc = tp[i].cumsum(0)
# Recall
recall = tpc / (n_l + eps) # recall curve
r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases
# Precision
precision = tpc / (tpc + fpc) # precision curve
p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score
# AP from recall-precision curve
for j in range(tp.shape[1]):
ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
if plot and j == 0:
py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5
# Compute F1 (harmonic mean of precision and recall)
f1 = 2 * p * r / (p + r + eps)
names = [v for k, v in names.items() if k in unique_classes] # list: only classes that have data
names = dict(enumerate(names)) # to dict
if plot:
plot_pr_curve(px, py, ap, Path(save_dir) / f'{prefix}PR_curve.png', names)
plot_mc_curve(px, f1, Path(save_dir) / f'{prefix}F1_curve.png', names, ylabel='F1')
plot_mc_curve(px, p, Path(save_dir) / f'{prefix}P_curve.png', names, ylabel='Precision')
plot_mc_curve(px, r, Path(save_dir) / f'{prefix}R_curve.png', names, ylabel='Recall')
i = smooth(f1.mean(0), 0.1).argmax() # max F1 index
p, r, f1 = p[:, i], r[:, i], f1[:, i]
tp = (r * nt).round() # true positives
fp = (tp / (p + eps) - tp).round() # false positives
return tp, fp, p, r, f1, ap, unique_classes.astype(int)
def compute_ap(recall, precision):
""" Compute the average precision, given the recall and precision curves
# Arguments
recall: The recall curve (list)
precision: The precision curve (list)
# Returns
Average precision, precision curve, recall curve
"""
# Append sentinel values to beginning and end
mrec = np.concatenate(([0.0], recall, [1.0]))
mpre = np.concatenate(([1.0], precision, [0.0]))
# Compute the precision envelope
mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
# Integrate area under curve
method = 'interp' # methods: 'continuous', 'interp'
if method == 'interp':
x = np.linspace(0, 1, 101) # 101-point interp (COCO)
ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate
else: # 'continuous'
i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes
ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve
return ap, mpre, mrec
class ConfusionMatrix:
# Updated version of https://github.com/kaanakan/object_detection_confusion_matrix
def __init__(self, nc, conf=0.25, iou_thres=0.45):
self.matrix = np.zeros((nc + 1, nc + 1))
self.nc = nc # number of classes
self.conf = conf
self.iou_thres = iou_thres
def process_batch(self, detections, labels):
"""
Return intersection-over-union (Jaccard index) of boxes.
Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
Arguments:
detections (Array[N, 6]), x1, y1, x2, y2, conf, class
labels (Array[M, 5]), class, x1, y1, x2, y2
Returns:
None, updates confusion matrix accordingly
"""
if detections is None:
gt_classes = labels.int()
for gc in gt_classes:
self.matrix[self.nc, gc] += 1 # background FN
return
detections = detections[detections[:, 4] > self.conf]
gt_classes = labels[:, 0].int()
detection_classes = detections[:, 5].int()
iou = box_iou(labels[:, 1:], detections[:, :4])
x = torch.where(iou > self.iou_thres)
if x[0].shape[0]:
matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
if x[0].shape[0] > 1:
matches = matches[matches[:, 2].argsort()[::-1]]
matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
matches = matches[matches[:, 2].argsort()[::-1]]
matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
else:
matches = np.zeros((0, 3))
n = matches.shape[0] > 0
m0, m1, _ = matches.transpose().astype(int)
for i, gc in enumerate(gt_classes):
j = m0 == i
if n and sum(j) == 1:
self.matrix[detection_classes[m1[j]], gc] += 1 # correct
else:
self.matrix[self.nc, gc] += 1 # true background
if n:
for i, dc in enumerate(detection_classes):
if not any(m1 == i):
self.matrix[dc, self.nc] += 1 # predicted background
def matrix(self):
return self.matrix
def tp_fp(self):
tp = self.matrix.diagonal() # true positives
fp = self.matrix.sum(1) - tp # false positives
# fn = self.matrix.sum(0) - tp # false negatives (missed detections)
return tp[:-1], fp[:-1] # remove background class
@TryExcept('WARNING ⚠️ ConfusionMatrix plot failure')
def plot(self, normalize=True, save_dir='', names=()):
import seaborn as sn
array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1E-9) if normalize else 1) # normalize columns
array[array < 0.005] = np.nan # don't annotate (would appear as 0.00)
fig, ax = plt.subplots(1, 1, figsize=(12, 9), tight_layout=True)
nc, nn = self.nc, len(names) # number of classes, names
sn.set(font_scale=1.0 if nc < 50 else 0.8) # for label size
labels = (0 < nn < 99) and (nn == nc) # apply names to ticklabels
ticklabels = (names + ['background']) if labels else "auto"
with warnings.catch_warnings():
warnings.simplefilter('ignore') # suppress empty matrix RuntimeWarning: All-NaN slice encountered
sn.heatmap(array,
ax=ax,
annot=nc < 30,
annot_kws={
"size": 8},
cmap='Blues',
fmt='.2f',
square=True,
vmin=0.0,
xticklabels=ticklabels,
yticklabels=ticklabels).set_facecolor((1, 1, 1))
ax.set_ylabel('True')
ax.set_ylabel('Predicted')
ax.set_title('Confusion Matrix')
fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250)
plt.close(fig)
def print(self):
for i in range(self.nc + 1):
print(' '.join(map(str, self.matrix[i])))
def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
# Returns Intersection over Union (IoU) of box1(1,4) to box2(n,4)
# Get the coordinates of bounding boxes
if xywh: # transform from xywh to xyxy
(x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, -1), box2.chunk(4, -1)
w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2
b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_
b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_
else: # x1, y1, x2, y2 = box1
b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, -1)
b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, -1)
w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
# Intersection area
inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
# Union Area
union = w1 * h1 + w2 * h2 - inter + eps
# IoU
iou = inter / union
if CIoU or DIoU or GIoU:
cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width
ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height
if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared
rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center dist ** 2
if CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
with torch.no_grad():
alpha = v / (v - iou + (1 + eps))
return iou - (rho2 / c2 + v * alpha) # CIoU
return iou - rho2 / c2 # DIoU
c_area = cw * ch + eps # convex area
return iou - (c_area - union) / c_area # GIoU https://arxiv.org/pdf/1902.09630.pdf
return iou # IoU
def box_iou(box1, box2, eps=1e-7):
# https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
"""
Return intersection-over-union (Jaccard index) of boxes.
Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
Arguments:
box1 (Tensor[N, 4])
box2 (Tensor[M, 4])
Returns:
iou (Tensor[N, M]): the NxM matrix containing the pairwise
IoU values for every element in boxes1 and boxes2
"""
# inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
(a1, a2), (b1, b2) = box1.unsqueeze(1).chunk(2, 2), box2.unsqueeze(0).chunk(2, 2)
inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp(0).prod(2)
# IoU = inter / (area1 + area2 - inter)
return inter / ((a2 - a1).prod(2) + (b2 - b1).prod(2) - inter + eps)
def bbox_ioa(box1, box2, eps=1e-7):
""" Returns the intersection over box2 area given box1, box2. Boxes are x1y1x2y2
box1: np.array of shape(4)
box2: np.array of shape(nx4)
returns: np.array of shape(n)
"""
# Get the coordinates of bounding boxes
b1_x1, b1_y1, b1_x2, b1_y2 = box1
b2_x1, b2_y1, b2_x2, b2_y2 = box2.T
# Intersection area
inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
(np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)
# box2 area
box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + eps
# Intersection over box2 area
return inter_area / box2_area
def wh_iou(wh1, wh2, eps=1e-7):
# Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2
wh1 = wh1[:, None] # [N,1,2]
wh2 = wh2[None] # [1,M,2]
inter = torch.min(wh1, wh2).prod(2) # [N,M]
return inter / (wh1.prod(2) + wh2.prod(2) - inter + eps) # iou = inter / (area1 + area2 - inter)
# Plots ----------------------------------------------------------------------------------------------------------------
@threaded
def plot_pr_curve(px, py, ap, save_dir=Path('pr_curve.png'), names=()):
# Precision-recall curve
fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
py = np.stack(py, axis=1)
if 0 < len(names) < 21: # display per-class legend if < 21 classes
for i, y in enumerate(py.T):
ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision)
else:
ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision)
ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean())
ax.set_xlabel('Recall')
ax.set_ylabel('Precision')
ax.set_xlim(0, 1)
ax.set_ylim(0, 1)
ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
ax.set_title('Precision-Recall Curve')
fig.savefig(save_dir, dpi=250)
plt.close(fig)
@threaded
def plot_mc_curve(px, py, save_dir=Path('mc_curve.png'), names=(), xlabel='Confidence', ylabel='Metric'):
# Metric-confidence curve
fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
if 0 < len(names) < 21: # display per-class legend if < 21 classes
for i, y in enumerate(py):
ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric)
else:
ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric)
y = smooth(py.mean(0), 0.05)
ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}')
ax.set_xlabel(xlabel)
ax.set_ylabel(ylabel)
ax.set_xlim(0, 1)
ax.set_ylim(0, 1)
ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
ax.set_title(f'{ylabel}-Confidence Curve')
fig.savefig(save_dir, dpi=250)
plt.close(fig)
utils/plots.py 0 → 100644
View file @ cf1ed55a
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Plotting utils
"""
import contextlib
import math
import os
from copy import copy
from pathlib import Path
from urllib.error import URLError
import cv2
import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sn
import torch
from PIL import Image, ImageDraw, ImageFont
from utils import TryExcept, threaded
from utils.general import (CONFIG_DIR, FONT, LOGGER, check_font, check_requirements, clip_boxes, increment_path,
is_ascii, xywh2xyxy, xyxy2xywh)
from utils.metrics import fitness
from utils.segment.general import scale_image
# Settings
RANK = int(os.getenv('RANK', -1))
matplotlib.rc('font', **{'size': 11})
matplotlib.use('Agg') # for writing to files only
class Colors:
# Ultralytics color palette https://ultralytics.com/
def __init__(self):
# hex = matplotlib.colors.TABLEAU_COLORS.values()
hexs = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB',
'2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7')
self.palette = [self.hex2rgb(f'#{c}') for c in hexs]
self.n = len(self.palette)
def __call__(self, i, bgr=False):
c = self.palette[int(i) % self.n]
return (c[2], c[1], c[0]) if bgr else c
@staticmethod
def hex2rgb(h): # rgb order (PIL)
return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
colors = Colors() # create instance for 'from utils.plots import colors'
def check_pil_font(font=FONT, size=10):
# Return a PIL TrueType Font, downloading to CONFIG_DIR if necessary
font = Path(font)
font = font if font.exists() else (CONFIG_DIR / font.name)
try:
return ImageFont.truetype(str(font) if font.exists() else font.name, size)
except Exception: # download if missing
try:
check_font(font)
return ImageFont.truetype(str(font), size)
except TypeError:
check_requirements('Pillow>=8.4.0') # known issue https://github.com/ultralytics/yolov5/issues/5374
except URLError: # not online
return ImageFont.load_default()
class Annotator:
# YOLOv5 Annotator for train/val mosaics and jpgs and detect/hub inference annotations
def __init__(self, im, line_width=None, font_size=None, font='Arial.ttf', pil=False, example='abc'):
assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to Annotator() input images.'
non_ascii = not is_ascii(example) # non-latin labels, i.e. asian, arabic, cyrillic
self.pil = pil or non_ascii
if self.pil: # use PIL
self.im = im if isinstance(im, Image.Image) else Image.fromarray(im)
self.draw = ImageDraw.Draw(self.im)
self.font = check_pil_font(font='Arial.Unicode.ttf' if non_ascii else font,
size=font_size or max(round(sum(self.im.size) / 2 * 0.035), 12))
else: # use cv2
self.im = im
self.lw = line_width or max(round(sum(im.shape) / 2 * 0.003), 2) # line width
def box_label(self, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)):
# Add one xyxy box to image with label
if self.pil or not is_ascii(label):
self.draw.rectangle(box, width=self.lw, outline=color) # box
if label:
left, top, right, bottom = self.font.getbbox(label)
w, h = right - left, bottom - top
outside = box[1] - h >= 0 # label fits outside box
self.draw.rectangle(
(box[0], box[1] - h if outside else box[1], box[0] + w + 1,
box[1] + 1 if outside else box[1] + h + 1),
fill=color,
)
# self.draw.text((box[0], box[1]), label, fill=txt_color, font=self.font, anchor='ls') # for PIL>8.0
self.draw.text((box[0], box[1] - h if outside else box[1]), label, fill=txt_color, font=self.font)
else: # cv2
p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3]))
cv2.rectangle(self.im, p1, p2, color, thickness=self.lw, lineType=cv2.LINE_AA)
if label:
tf = max(self.lw - 1, 1) # font thickness
w, h = cv2.getTextSize(label, 0, fontScale=self.lw / 3, thickness=tf)[0] # text width, height
outside = p1[1] - h >= 3
p2 = p1[0] + w, p1[1] - h - 3 if outside else p1[1] + h + 3
cv2.rectangle(self.im, p1, p2, color, -1, cv2.LINE_AA) # filled
cv2.putText(self.im,
label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2),
0,
self.lw / 3,
txt_color,
thickness=tf,
lineType=cv2.LINE_AA)
def masks(self, masks, colors, im_gpu=None, alpha=0.5):
"""Plot masks at once.
Args:
masks (tensor): predicted masks on cuda, shape: [n, h, w]
colors (List[List[Int]]): colors for predicted masks, [[r, g, b] * n]
im_gpu (tensor): img is in cuda, shape: [3, h, w], range: [0, 1]
alpha (float): mask transparency: 0.0 fully transparent, 1.0 opaque
"""
if self.pil:
# convert to numpy first
self.im = np.asarray(self.im).copy()
if im_gpu is None:
# Add multiple masks of shape(h,w,n) with colors list([r,g,b], [r,g,b], ...)
if len(masks) == 0:
return
if isinstance(masks, torch.Tensor):
masks = torch.as_tensor(masks, dtype=torch.uint8)
masks = masks.permute(1, 2, 0).contiguous()
masks = masks.cpu().numpy()
# masks = np.ascontiguousarray(masks.transpose(1, 2, 0))
masks = scale_image(masks.shape[:2], masks, self.im.shape)
masks = np.asarray(masks, dtype=np.float32)
colors = np.asarray(colors, dtype=np.float32) # shape(n,3)
s = masks.sum(2, keepdims=True).clip(0, 1) # add all masks together
masks = (masks @ colors).clip(0, 255) # (h,w,n) @ (n,3) = (h,w,3)
self.im[:] = masks * alpha + self.im * (1 - s * alpha)
else:
if len(masks) == 0:
self.im[:] = im_gpu.permute(1, 2, 0).contiguous().cpu().numpy() * 255
colors = torch.tensor(colors, device=im_gpu.device, dtype=torch.float32) / 255.0
colors = colors[:, None, None] # shape(n,1,1,3)
masks = masks.unsqueeze(3) # shape(n,h,w,1)
masks_color = masks * (colors * alpha) # shape(n,h,w,3)
inv_alph_masks = (1 - masks * alpha).cumprod(0) # shape(n,h,w,1)
mcs = (masks_color * inv_alph_masks).sum(0) * 2 # mask color summand shape(n,h,w,3)
im_gpu = im_gpu.flip(dims=[0]) # flip channel
im_gpu = im_gpu.permute(1, 2, 0).contiguous() # shape(h,w,3)
im_gpu = im_gpu * inv_alph_masks[-1] + mcs
im_mask = (im_gpu * 255).byte().cpu().numpy()
self.im[:] = scale_image(im_gpu.shape, im_mask, self.im.shape)
if self.pil:
# convert im back to PIL and update draw
self.fromarray(self.im)
def rectangle(self, xy, fill=None, outline=None, width=1):
# Add rectangle to image (PIL-only)
self.draw.rectangle(xy, fill, outline, width)
def text(self, xy, text, txt_color=(255, 255, 255), anchor='top'):
# Add text to image (PIL-only)
if anchor == 'bottom': # start y from font bottom
w, h = self.font.getsize(text) # text width, height
xy[1] += 1 - h
self.draw.text(xy, text, fill=txt_color, font=self.font)
def fromarray(self, im):
# Update self.im from a numpy array
self.im = im if isinstance(im, Image.Image) else Image.fromarray(im)
self.draw = ImageDraw.Draw(self.im)
def result(self):
# Return annotated image as array
return np.asarray(self.im)
def feature_visualization(x, module_type, stage, n=32, save_dir=Path('runs/detect/exp')):
"""
x: Features to be visualized
module_type: Module type
stage: Module stage within model
n: Maximum number of feature maps to plot
save_dir: Directory to save results
"""
if 'Detect' not in module_type:
batch, channels, height, width = x.shape # batch, channels, height, width
if height > 1 and width > 1:
f = save_dir / f"stage{stage}_{module_type.split('.')[-1]}_features.png" # filename
blocks = torch.chunk(x[0].cpu(), channels, dim=0) # select batch index 0, block by channels
n = min(n, channels) # number of plots
fig, ax = plt.subplots(math.ceil(n / 8), 8, tight_layout=True) # 8 rows x n/8 cols
ax = ax.ravel()
plt.subplots_adjust(wspace=0.05, hspace=0.05)
for i in range(n):
ax[i].imshow(blocks[i].squeeze()) # cmap='gray'
ax[i].axis('off')
LOGGER.info(f'Saving {f}... ({n}/{channels})')
plt.savefig(f, dpi=300, bbox_inches='tight')
plt.close()
np.save(str(f.with_suffix('.npy')), x[0].cpu().numpy()) # npy save
def hist2d(x, y, n=100):
# 2d histogram used in labels.png and evolve.png
xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n)
hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges))
xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1)
yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1)
return np.log(hist[xidx, yidx])
def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5):
from scipy.signal import butter, filtfilt
# https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy
def butter_lowpass(cutoff, fs, order):
nyq = 0.5 * fs
normal_cutoff = cutoff / nyq
return butter(order, normal_cutoff, btype='low', analog=False)
b, a = butter_lowpass(cutoff, fs, order=order)
return filtfilt(b, a, data) # forward-backward filter
def output_to_target(output, max_det=300):
# Convert model output to target format [batch_id, class_id, x, y, w, h, conf] for plotting
targets = []
for i, o in enumerate(output):
box, conf, cls = o[:max_det, :6].cpu().split((4, 1, 1), 1)
j = torch.full((conf.shape[0], 1), i)
targets.append(torch.cat((j, cls, xyxy2xywh(box), conf), 1))
return torch.cat(targets, 0).numpy()
@threaded
def plot_images(images, targets, paths=None, fname='images.jpg', names=None):
# Plot image grid with labels
if isinstance(images, torch.Tensor):
images = images.cpu().float().numpy()
if isinstance(targets, torch.Tensor):
targets = targets.cpu().numpy()
max_size = 1920 # max image size
max_subplots = 16 # max image subplots, i.e. 4x4
bs, _, h, w = images.shape # batch size, _, height, width
bs = min(bs, max_subplots) # limit plot images
ns = np.ceil(bs ** 0.5) # number of subplots (square)
if np.max(images[0]) <= 1:
images *= 255 # de-normalise (optional)
# Build Image
mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init
for i, im in enumerate(images):
if i == max_subplots: # if last batch has fewer images than we expect
break
x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin
im = im.transpose(1, 2, 0)
mosaic[y:y + h, x:x + w, :] = im
# Resize (optional)
scale = max_size / ns / max(h, w)
if scale < 1:
h = math.ceil(scale * h)
w = math.ceil(scale * w)
mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h)))
# Annotate
fs = int((h + w) * ns * 0.01) # font size
annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs, pil=True, example=names)
for i in range(i + 1):
x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin
annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2) # borders
if paths:
annotator.text((x + 5, y + 5), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220)) # filenames
if len(targets) > 0:
ti = targets[targets[:, 0] == i] # image targets
boxes = xywh2xyxy(ti[:, 2:6]).T
classes = ti[:, 1].astype('int')
labels = ti.shape[1] == 6 # labels if no conf column
conf = None if labels else ti[:, 6] # check for confidence presence (label vs pred)
if boxes.shape[1]:
if boxes.max() <= 1.01: # if normalized with tolerance 0.01
boxes[[0, 2]] *= w # scale to pixels
boxes[[1, 3]] *= h
elif scale < 1: # absolute coords need scale if image scales
boxes *= scale
boxes[[0, 2]] += x
boxes[[1, 3]] += y
for j, box in enumerate(boxes.T.tolist()):
cls = classes[j]
color = colors(cls)
cls = names[cls] if names else cls
if labels or conf[j] > 0.25: # 0.25 conf thresh
label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}'
annotator.box_label(box, label, color=color)
annotator.im.save(fname) # save
def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''):
# Plot LR simulating training for full epochs
optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals
y = []
for _ in range(epochs):
scheduler.step()
y.append(optimizer.param_groups[0]['lr'])
plt.plot(y, '.-', label='LR')
plt.xlabel('epoch')
plt.ylabel('LR')
plt.grid()
plt.xlim(0, epochs)
plt.ylim(0)
plt.savefig(Path(save_dir) / 'LR.png', dpi=200)
plt.close()
def plot_val_txt(): # from utils.plots import *; plot_val()
# Plot val.txt histograms
x = np.loadtxt('val.txt', dtype=np.float32)
box = xyxy2xywh(x[:, :4])
cx, cy = box[:, 0], box[:, 1]
fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True)
ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0)
ax.set_aspect('equal')
plt.savefig('hist2d.png', dpi=300)
fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True)
ax[0].hist(cx, bins=600)
ax[1].hist(cy, bins=600)
plt.savefig('hist1d.png', dpi=200)
def plot_targets_txt(): # from utils.plots import *; plot_targets_txt()
# Plot targets.txt histograms
x = np.loadtxt('targets.txt', dtype=np.float32).T
s = ['x targets', 'y targets', 'width targets', 'height targets']
fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)
ax = ax.ravel()
for i in range(4):
ax[i].hist(x[i], bins=100, label=f'{x[i].mean():.3g} +/- {x[i].std():.3g}')
ax[i].legend()
ax[i].set_title(s[i])
plt.savefig('targets.jpg', dpi=200)
def plot_val_study(file='', dir='', x=None): # from utils.plots import *; plot_val_study()
# Plot file=study.txt generated by val.py (or plot all study*.txt in dir)
save_dir = Path(file).parent if file else Path(dir)
plot2 = False # plot additional results
if plot2:
ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)[1].ravel()
fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
# for f in [save_dir / f'study_coco_{x}.txt' for x in ['yolov5n6', 'yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]:
for f in sorted(save_dir.glob('study*.txt')):
y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T
x = np.arange(y.shape[1]) if x is None else np.array(x)
if plot2:
s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_preprocess (ms/img)', 't_inference (ms/img)', 't_NMS (ms/img)']
for i in range(7):
ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8)
ax[i].set_title(s[i])
j = y[3].argmax() + 1
ax2.plot(y[5, 1:j],
y[3, 1:j] * 1E2,
'.-',
linewidth=2,
markersize=8,
label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5],
'k.-',
linewidth=2,
markersize=8,
alpha=.25,
label='EfficientDet')
ax2.grid(alpha=0.2)
ax2.set_yticks(np.arange(20, 60, 5))
ax2.set_xlim(0, 57)
ax2.set_ylim(25, 55)
ax2.set_xlabel('GPU Speed (ms/img)')
ax2.set_ylabel('COCO AP val')
ax2.legend(loc='lower right')
f = save_dir / 'study.png'
print(f'Saving {f}...')
plt.savefig(f, dpi=300)
@TryExcept() # known issue https://github.com/ultralytics/yolov5/issues/5395
def plot_labels(labels, names=(), save_dir=Path('')):
# plot dataset labels
LOGGER.info(f"Plotting labels to {save_dir / 'labels.jpg'}... ")
c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes
nc = int(c.max() + 1) # number of classes
x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height'])
# seaborn correlogram
sn.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9))
plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200)
plt.close()
# matplotlib labels
matplotlib.use('svg') # faster
ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel()
y = ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8)
with contextlib.suppress(Exception): # color histogram bars by class
[y[2].patches[i].set_color([x / 255 for x in colors(i)]) for i in range(nc)] # known issue #3195
ax[0].set_ylabel('instances')
if 0 < len(names) < 30:
ax[0].set_xticks(range(len(names)))
ax[0].set_xticklabels(list(names.values()), rotation=90, fontsize=10)
else:
ax[0].set_xlabel('classes')
sn.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9)
sn.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9)
# rectangles
labels[:, 1:3] = 0.5 # center
labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000
img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255)
for cls, *box in labels[:1000]:
ImageDraw.Draw(img).rectangle(box, width=1, outline=colors(cls)) # plot
ax[1].imshow(img)
ax[1].axis('off')
for a in [0, 1, 2, 3]:
for s in ['top', 'right', 'left', 'bottom']:
ax[a].spines[s].set_visible(False)
plt.savefig(save_dir / 'labels.jpg', dpi=200)
matplotlib.use('Agg')
plt.close()
def imshow_cls(im, labels=None, pred=None, names=None, nmax=25, verbose=False, f=Path('images.jpg')):
# Show classification image grid with labels (optional) and predictions (optional)
from utils.augmentations import denormalize
names = names or [f'class{i}' for i in range(1000)]
blocks = torch.chunk(denormalize(im.clone()).cpu().float(), len(im),
dim=0) # select batch index 0, block by channels
n = min(len(blocks), nmax) # number of plots
m = min(8, round(n ** 0.5)) # 8 x 8 default
fig, ax = plt.subplots(math.ceil(n / m), m) # 8 rows x n/8 cols
ax = ax.ravel() if m > 1 else [ax]
# plt.subplots_adjust(wspace=0.05, hspace=0.05)
for i in range(n):
ax[i].imshow(blocks[i].squeeze().permute((1, 2, 0)).numpy().clip(0.0, 1.0))
ax[i].axis('off')
if labels is not None:
s = names[labels[i]] + (f'—{names[pred[i]]}' if pred is not None else '')
ax[i].set_title(s, fontsize=8, verticalalignment='top')
plt.savefig(f, dpi=300, bbox_inches='tight')
plt.close()
if verbose:
LOGGER.info(f"Saving {f}")
if labels is not None:
LOGGER.info('True: ' + ' '.join(f'{names[i]:3s}' for i in labels[:nmax]))
if pred is not None:
LOGGER.info('Predicted:' + ' '.join(f'{names[i]:3s}' for i in pred[:nmax]))
return f
def plot_evolve(evolve_csv='path/to/evolve.csv'): # from utils.plots import *; plot_evolve()
# Plot evolve.csv hyp evolution results
evolve_csv = Path(evolve_csv)
data = pd.read_csv(evolve_csv)
keys = [x.strip() for x in data.columns]
x = data.values
f = fitness(x)
j = np.argmax(f) # max fitness index
plt.figure(figsize=(10, 12), tight_layout=True)
matplotlib.rc('font', **{'size': 8})
print(f'Best results from row {j} of {evolve_csv}:')
for i, k in enumerate(keys[7:]):
v = x[:, 7 + i]
mu = v[j] # best single result
plt.subplot(6, 5, i + 1)
plt.scatter(v, f, c=hist2d(v, f, 20), cmap='viridis', alpha=.8, edgecolors='none')
plt.plot(mu, f.max(), 'k+', markersize=15)
plt.title(f'{k} = {mu:.3g}', fontdict={'size': 9}) # limit to 40 characters
if i % 5 != 0:
plt.yticks([])
print(f'{k:>15}: {mu:.3g}')
f = evolve_csv.with_suffix('.png') # filename
plt.savefig(f, dpi=200)
plt.close()
print(f'Saved {f}')
def plot_results(file='path/to/results.csv', dir=''):
# Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv')
save_dir = Path(file).parent if file else Path(dir)
fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True)
ax = ax.ravel()
files = list(save_dir.glob('results*.csv'))
assert len(files), f'No results.csv files found in {save_dir.resolve()}, nothing to plot.'
for f in files:
try:
data = pd.read_csv(f)
s = [x.strip() for x in data.columns]
x = data.values[:, 0]
for i, j in enumerate([1, 2, 3, 4, 5, 8, 9, 10, 6, 7]):
y = data.values[:, j].astype('float')
# y[y == 0] = np.nan # don't show zero values
ax[i].plot(x, y, marker='.', label=f.stem, linewidth=2, markersize=8)
ax[i].set_title(s[j], fontsize=12)
# if j in [8, 9, 10]: # share train and val loss y axes
# ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
except Exception as e:
LOGGER.info(f'Warning: Plotting error for {f}: {e}')
ax[1].legend()
fig.savefig(save_dir / 'results.png', dpi=200)
plt.close()
def profile_idetection(start=0, stop=0, labels=(), save_dir=''):
# Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection()
ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel()
s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS']
files = list(Path(save_dir).glob('frames*.txt'))
for fi, f in enumerate(files):
try:
results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows
n = results.shape[1] # number of rows
x = np.arange(start, min(stop, n) if stop else n)
results = results[:, x]
t = (results[0] - results[0].min()) # set t0=0s
results[0] = x
for i, a in enumerate(ax):
if i < len(results):
label = labels[fi] if len(labels) else f.stem.replace('frames_', '')
a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5)
a.set_title(s[i])
a.set_xlabel('time (s)')
# if fi == len(files) - 1:
# a.set_ylim(bottom=0)
for side in ['top', 'right']:
a.spines[side].set_visible(False)
else:
a.remove()
except Exception as e:
print(f'Warning: Plotting error for {f}; {e}')
ax[1].legend()
plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200)
def save_one_box(xyxy, im, file=Path('im.jpg'), gain=1.02, pad=10, square=False, BGR=False, save=True):
# Save image crop as {file} with crop size multiple {gain} and {pad} pixels. Save and/or return crop
xyxy = torch.tensor(xyxy).view(-1, 4)
b = xyxy2xywh(xyxy) # boxes
if square:
b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # attempt rectangle to square
b[:, 2:] = b[:, 2:] * gain + pad # box wh * gain + pad
xyxy = xywh2xyxy(b).long()
clip_boxes(xyxy, im.shape)
crop = im[int(xyxy[0, 1]):int(xyxy[0, 3]), int(xyxy[0, 0]):int(xyxy[0, 2]), ::(1 if BGR else -1)]
if save:
file.parent.mkdir(parents=True, exist_ok=True) # make directory
f = str(increment_path(file).with_suffix('.jpg'))
# cv2.imwrite(f, crop) # save BGR, https://github.com/ultralytics/yolov5/issues/7007 chroma subsampling issue
Image.fromarray(crop[..., ::-1]).save(f, quality=95, subsampling=0) # save RGB
return crop
utils/segment/augmentations.py 0 → 100644
View file @ cf1ed55a
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Image augmentation functions
"""
import math
import random
import cv2
import numpy as np
from ..augmentations import box_candidates
from ..general import resample_segments, segment2box
def mixup(im, labels, segments, im2, labels2, segments2):
# Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0
im = (im * r + im2 * (1 - r)).astype(np.uint8)
labels = np.concatenate((labels, labels2), 0)
segments = np.concatenate((segments, segments2), 0)
return im, labels, segments
def random_perspective(im,
targets=(),
segments=(),
degrees=10,
translate=.1,
scale=.1,
shear=10,
perspective=0.0,
border=(0, 0)):
# torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
# targets = [cls, xyxy]
height = im.shape[0] + border[0] * 2 # shape(h,w,c)
width = im.shape[1] + border[1] * 2
# Center
C = np.eye(3)
C[0, 2] = -im.shape[1] / 2 # x translation (pixels)
C[1, 2] = -im.shape[0] / 2 # y translation (pixels)
# Perspective
P = np.eye(3)
P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y)
P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x)
# Rotation and Scale
R = np.eye(3)
a = random.uniform(-degrees, degrees)
# a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations
s = random.uniform(1 - scale, 1 + scale)
# s = 2 ** random.uniform(-scale, scale)
R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
# Shear
S = np.eye(3)
S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)
S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)
# Translation
T = np.eye(3)
T[0, 2] = (random.uniform(0.5 - translate, 0.5 + translate) * width) # x translation (pixels)
T[1, 2] = (random.uniform(0.5 - translate, 0.5 + translate) * height) # y translation (pixels)
# Combined rotation matrix
M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT
if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed
if perspective:
im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
else: # affine
im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
# Visualize
# import matplotlib.pyplot as plt
# ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
# ax[0].imshow(im[:, :, ::-1]) # base
# ax[1].imshow(im2[:, :, ::-1]) # warped
# Transform label coordinates
n = len(targets)
new_segments = []
if n:
new = np.zeros((n, 4))
segments = resample_segments(segments) # upsample
for i, segment in enumerate(segments):
xy = np.ones((len(segment), 3))
xy[:, :2] = segment
xy = xy @ M.T # transform
xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]) # perspective rescale or affine
# clip
new[i] = segment2box(xy, width, height)
new_segments.append(xy)
# filter candidates
i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01)
targets = targets[i]
targets[:, 1:5] = new[i]
new_segments = np.array(new_segments)[i]
return im, targets, new_segments
utils/segment/dataloaders.py 0 → 100644
View file @ cf1ed55a
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
Dataloaders
"""
import os
import random
import cv2
import numpy as np
import torch
from torch.utils.data import DataLoader, distributed
from ..augmentations import augment_hsv, copy_paste, letterbox
from ..dataloaders import InfiniteDataLoader, LoadImagesAndLabels, seed_worker
from ..general import LOGGER, xyn2xy, xywhn2xyxy, xyxy2xywhn
from ..torch_utils import torch_distributed_zero_first
from .augmentations import mixup, random_perspective
RANK = int(os.getenv('RANK', -1))
def create_dataloader(path,
imgsz,
batch_size,
stride,
single_cls=False,
hyp=None,
augment=False,
cache=False,
pad=0.0,
rect=False,
rank=-1,
workers=8,
image_weights=False,
quad=False,
prefix='',
shuffle=False,
mask_downsample_ratio=1,
overlap_mask=False):
if rect and shuffle:
LOGGER.warning('WARNING ⚠️ --rect is incompatible with DataLoader shuffle, setting shuffle=False')
shuffle = False
with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP
dataset = LoadImagesAndLabelsAndMasks(
path,
imgsz,
batch_size,
augment=augment, # augmentation
hyp=hyp, # hyperparameters
rect=rect, # rectangular batches
cache_images=cache,
single_cls=single_cls,
stride=int(stride),
pad=pad,
image_weights=image_weights,
prefix=prefix,
downsample_ratio=mask_downsample_ratio,
overlap=overlap_mask)
batch_size = min(batch_size, len(dataset))
nd = torch.cuda.device_count() # number of CUDA devices
nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers]) # number of workers
sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle)
loader = DataLoader if image_weights else InfiniteDataLoader # only DataLoader allows for attribute updates
generator = torch.Generator()
generator.manual_seed(6148914691236517205 + RANK)
return loader(
dataset,
batch_size=batch_size,
shuffle=shuffle and sampler is None,
num_workers=nw,
sampler=sampler,
pin_memory=True,
collate_fn=LoadImagesAndLabelsAndMasks.collate_fn4 if quad else LoadImagesAndLabelsAndMasks.collate_fn,
worker_init_fn=seed_worker,
generator=generator,
), dataset
class LoadImagesAndLabelsAndMasks(LoadImagesAndLabels): # for training/testing
def __init__(
self,
path,
img_size=640,
batch_size=16,
augment=False,
hyp=None,
rect=False,
image_weights=False,
cache_images=False,
single_cls=False,
stride=32,
pad=0,
min_items=0,
prefix="",
downsample_ratio=1,
overlap=False,
):
super().__init__(path, img_size, batch_size, augment, hyp, rect, image_weights, cache_images, single_cls,
stride, pad, min_items, prefix)
self.downsample_ratio = downsample_ratio
self.overlap = overlap
def __getitem__(self, index):
index = self.indices[index] # linear, shuffled, or image_weights
hyp = self.hyp
mosaic = self.mosaic and random.random() < hyp['mosaic']
masks = []
if mosaic:
# Load mosaic
img, labels, segments = self.load_mosaic(index)
shapes = None
# MixUp augmentation
if random.random() < hyp["mixup"]:
img, labels, segments = mixup(img, labels, segments, *self.load_mosaic(random.randint(0, self.n - 1)))
else:
# Load image
img, (h0, w0), (h, w) = self.load_image(index)
# Letterbox
shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape
img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling
labels = self.labels[index].copy()
# [array, array, ....], array.shape=(num_points, 2), xyxyxyxy
segments = self.segments[index].copy()
if len(segments):
for i_s in range(len(segments)):
segments[i_s] = xyn2xy(
segments[i_s],
ratio[0] * w,
ratio[1] * h,
padw=pad[0],
padh=pad[1],
)
if labels.size: # normalized xywh to pixel xyxy format
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
if self.augment:
img, labels, segments = random_perspective(img,
labels,
segments=segments,
degrees=hyp["degrees"],
translate=hyp["translate"],
scale=hyp["scale"],
shear=hyp["shear"],
perspective=hyp["perspective"])
nl = len(labels) # number of labels
if nl:
labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1e-3)
if self.overlap:
masks, sorted_idx = polygons2masks_overlap(img.shape[:2],
segments,
downsample_ratio=self.downsample_ratio)
masks = masks[None] # (640, 640) -> (1, 640, 640)
labels = labels[sorted_idx]
else:
masks = polygons2masks(img.shape[:2], segments, color=1, downsample_ratio=self.downsample_ratio)
masks = (torch.from_numpy(masks) if len(masks) else torch.zeros(1 if self.overlap else nl, img.shape[0] //
self.downsample_ratio, img.shape[1] //
self.downsample_ratio))
# TODO: albumentations support
if self.augment:
# Albumentations
# there are some augmentation that won't change boxes and masks,
# so just be it for now.
img, labels = self.albumentations(img, labels)
nl = len(labels) # update after albumentations
# HSV color-space
augment_hsv(img, hgain=hyp["hsv_h"], sgain=hyp["hsv_s"], vgain=hyp["hsv_v"])
# Flip up-down
if random.random() < hyp["flipud"]:
img = np.flipud(img)
if nl:
labels[:, 2] = 1 - labels[:, 2]
masks = torch.flip(masks, dims=[1])
# Flip left-right
if random.random() < hyp["fliplr"]:
img = np.fliplr(img)
if nl:
labels[:, 1] = 1 - labels[:, 1]
masks = torch.flip(masks, dims=[2])
# Cutouts # labels = cutout(img, labels, p=0.5)
labels_out = torch.zeros((nl, 6))
if nl:
labels_out[:, 1:] = torch.from_numpy(labels)
# Convert
img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB
img = np.ascontiguousarray(img)
return (torch.from_numpy(img), labels_out, self.im_files[index], shapes, masks)
def load_mosaic(self, index):
# YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic
labels4, segments4 = [], []
s = self.img_size
yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border) # mosaic center x, y
# 3 additional image indices
indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices
for i, index in enumerate(indices):
# Load image
img, _, (h, w) = self.load_image(index)
# place img in img4
if i == 0: # top left
img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image)
x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image)
elif i == 1: # top right
x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
elif i == 2: # bottom left
x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
elif i == 3: # bottom right
x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
padw = x1a - x1b
padh = y1a - y1b
labels, segments = self.labels[index].copy(), self.segments[index].copy()
if labels.size:
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format
segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
labels4.append(labels)
segments4.extend(segments)
# Concat/clip labels
labels4 = np.concatenate(labels4, 0)
for x in (labels4[:, 1:], *segments4):
np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
# img4, labels4 = replicate(img4, labels4) # replicate
# Augment
img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp["copy_paste"])
img4, labels4, segments4 = random_perspective(img4,
labels4,
segments4,
degrees=self.hyp["degrees"],
translate=self.hyp["translate"],
scale=self.hyp["scale"],
shear=self.hyp["shear"],
perspective=self.hyp["perspective"],
border=self.mosaic_border) # border to remove
return img4, labels4, segments4
@staticmethod
def collate_fn(batch):
img, label, path, shapes, masks = zip(*batch) # transposed
batched_masks = torch.cat(masks, 0)
for i, l in enumerate(label):
l[:, 0] = i # add target image index for build_targets()
return torch.stack(img, 0), torch.cat(label, 0), path, shapes, batched_masks
def polygon2mask(img_size, polygons, color=1, downsample_ratio=1):
"""
Args:
img_size (tuple): The image size.
polygons (np.ndarray): [N, M], N is the number of polygons,
M is the number of points(Be divided by 2).
"""
mask = np.zeros(img_size, dtype=np.uint8)
polygons = np.asarray(polygons)
polygons = polygons.astype(np.int32)
shape = polygons.shape
polygons = polygons.reshape(shape[0], -1, 2)
cv2.fillPoly(mask, polygons, color=color)
nh, nw = (img_size[0] // downsample_ratio, img_size[1] // downsample_ratio)
# NOTE: fillPoly firstly then resize is trying the keep the same way
# of loss calculation when mask-ratio=1.
mask = cv2.resize(mask, (nw, nh))
return mask
def polygons2masks(img_size, polygons, color, downsample_ratio=1):
"""
Args:
img_size (tuple): The image size.
polygons (list[np.ndarray]): each polygon is [N, M],
N is the number of polygons,
M is the number of points(Be divided by 2).
"""
masks = []
for si in range(len(polygons)):
mask = polygon2mask(img_size, [polygons[si].reshape(-1)], color, downsample_ratio)
masks.append(mask)
return np.array(masks)
def polygons2masks_overlap(img_size, segments, downsample_ratio=1):
"""Return a (640, 640) overlap mask."""
masks = np.zeros((img_size[0] // downsample_ratio, img_size[1] // downsample_ratio),
dtype=np.int32 if len(segments) > 255 else np.uint8)
areas = []
ms = []
for si in range(len(segments)):
mask = polygon2mask(
img_size,
[segments[si].reshape(-1)],
downsample_ratio=downsample_ratio,
color=1,
)
ms.append(mask)
areas.append(mask.sum())
areas = np.asarray(areas)
index = np.argsort(-areas)
ms = np.array(ms)[index]
for i in range(len(segments)):
mask = ms[i] * (i + 1)
masks = masks + mask
masks = np.clip(masks, a_min=0, a_max=i + 1)
return masks, index
utils/segment/general.py 0 → 100644
View file @ cf1ed55a
import cv2
import numpy as np
import torch
import torch.nn.functional as F
def crop_mask(masks, boxes):
"""
"Crop" predicted masks by zeroing out everything not in the predicted bbox.
Vectorized by Chong (thanks Chong).
Args:
- masks should be a size [h, w, n] tensor of masks
- boxes should be a size [n, 4] tensor of bbox coords in relative point form
"""
n, h, w = masks.shape
x1, y1, x2, y2 = torch.chunk(boxes[:, :, None], 4, 1) # x1 shape(1,1,n)
r = torch.arange(w, device=masks.device, dtype=x1.dtype)[None, None, :] # rows shape(1,w,1)
c = torch.arange(h, device=masks.device, dtype=x1.dtype)[None, :, None] # cols shape(h,1,1)
return masks * ((r >= x1) * (r < x2) * (c >= y1) * (c < y2))
def process_mask_upsample(protos, masks_in, bboxes, shape):
"""
Crop after upsample.
proto_out: [mask_dim, mask_h, mask_w]
out_masks: [n, mask_dim], n is number of masks after nms
bboxes: [n, 4], n is number of masks after nms
shape:input_image_size, (h, w)
return: h, w, n
"""
c, mh, mw = protos.shape # CHW
masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)
masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0] # CHW
masks = crop_mask(masks, bboxes) # CHW
return masks.gt_(0.5)
def process_mask(protos, masks_in, bboxes, shape, upsample=False):
"""
Crop before upsample.
proto_out: [mask_dim, mask_h, mask_w]
out_masks: [n, mask_dim], n is number of masks after nms
bboxes: [n, 4], n is number of masks after nms
shape:input_image_size, (h, w)
return: h, w, n
"""
c, mh, mw = protos.shape # CHW
ih, iw = shape
masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw) # CHW
downsampled_bboxes = bboxes.clone()
downsampled_bboxes[:, 0] *= mw / iw
downsampled_bboxes[:, 2] *= mw / iw
downsampled_bboxes[:, 3] *= mh / ih
downsampled_bboxes[:, 1] *= mh / ih
masks = crop_mask(masks, downsampled_bboxes) # CHW
if upsample:
masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0] # CHW
return masks.gt_(0.5)
def scale_image(im1_shape, masks, im0_shape, ratio_pad=None):
"""
img1_shape: model input shape, [h, w]
img0_shape: origin pic shape, [h, w, 3]
masks: [h, w, num]
"""
# Rescale coordinates (xyxy) from im1_shape to im0_shape
if ratio_pad is None: # calculate from im0_shape
gain = min(im1_shape[0] / im0_shape[0], im1_shape[1] / im0_shape[1]) # gain = old / new
pad = (im1_shape[1] - im0_shape[1] * gain) / 2, (im1_shape[0] - im0_shape[0] * gain) / 2 # wh padding
else:
pad = ratio_pad[1]
top, left = int(pad[1]), int(pad[0]) # y, x
bottom, right = int(im1_shape[0] - pad[1]), int(im1_shape[1] - pad[0])
if len(masks.shape) < 2:
raise ValueError(f'"len of masks shape" should be 2 or 3, but got {len(masks.shape)}')
masks = masks[top:bottom, left:right]
# masks = masks.permute(2, 0, 1).contiguous()
# masks = F.interpolate(masks[None], im0_shape[:2], mode='bilinear', align_corners=False)[0]
# masks = masks.permute(1, 2, 0).contiguous()
masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]))
if len(masks.shape) == 2:
masks = masks[:, :, None]
return masks
def mask_iou(mask1, mask2, eps=1e-7):
"""
mask1: [N, n] m1 means number of predicted objects
mask2: [M, n] m2 means number of gt objects
Note: n means image_w x image_h
return: masks iou, [N, M]
"""
intersection = torch.matmul(mask1, mask2.t()).clamp(0)
union = (mask1.sum(1)[:, None] + mask2.sum(1)[None]) - intersection # (area1 + area2) - intersection
return intersection / (union + eps)
def masks_iou(mask1, mask2, eps=1e-7):
"""
mask1: [N, n] m1 means number of predicted objects
mask2: [N, n] m2 means number of gt objects
Note: n means image_w x image_h
return: masks iou, (N, )
"""
intersection = (mask1 * mask2).sum(1).clamp(0) # (N, )
union = (mask1.sum(1) + mask2.sum(1))[None] - intersection # (area1 + area2) - intersection
return intersection / (union + eps)
def masks2segments(masks, strategy='largest'):
# Convert masks(n,160,160) into segments(n,xy)
segments = []
for x in masks.int().cpu().numpy().astype('uint8'):
c = cv2.findContours(x, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
if c:
if strategy == 'concat': # concatenate all segments
c = np.concatenate([x.reshape(-1, 2) for x in c])
elif strategy == 'largest': # select largest segment
c = np.array(c[np.array([len(x) for x in c]).argmax()]).reshape(-1, 2)
else:
c = np.zeros((0, 2)) # no segments found
segments.append(c.astype('float32'))
return segments
utils/segment/loss.py 0 → 100644
View file @ cf1ed55a
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..general import xywh2xyxy
from ..loss import FocalLoss, smooth_BCE
from ..metrics import bbox_iou
from ..torch_utils import de_parallel
from .general import crop_mask
class ComputeLoss:
# Compute losses
def __init__(self, model, autobalance=False, overlap=False):
self.sort_obj_iou = False
self.overlap = overlap
device = next(model.parameters()).device # get model device
h = model.hyp # hyperparameters
self.device = device
# Define criteria
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
# Focal loss
g = h['fl_gamma'] # focal loss gamma
if g > 0:
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
m = de_parallel(model).model[-1] # Detect() module
self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02]) # P3-P7
self.ssi = list(m.stride).index(16) if autobalance else 0 # stride 16 index
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
self.na = m.na # number of anchors
self.nc = m.nc # number of classes
self.nl = m.nl # number of layers
self.nm = m.nm # number of masks
self.anchors = m.anchors
self.device = device
def __call__(self, preds, targets, masks): # predictions, targets, model
p, proto = preds
bs, nm, mask_h, mask_w = proto.shape # batch size, number of masks, mask height, mask width
lcls = torch.zeros(1, device=self.device)
lbox = torch.zeros(1, device=self.device)
lobj = torch.zeros(1, device=self.device)
lseg = torch.zeros(1, device=self.device)
tcls, tbox, indices, anchors, tidxs, xywhn = self.build_targets(p, targets) # targets
# Losses
for i, pi in enumerate(p): # layer index, layer predictions
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device) # target obj
n = b.shape[0] # number of targets
if n:
pxy, pwh, _, pcls, pmask = pi[b, a, gj, gi].split((2, 2, 1, self.nc, nm), 1) # subset of predictions
# Box regression
pxy = pxy.sigmoid() * 2 - 0.5
pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i]
pbox = torch.cat((pxy, pwh), 1) # predicted box
iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze() # iou(prediction, target)
lbox += (1.0 - iou).mean() # iou loss
# Objectness
iou = iou.detach().clamp(0).type(tobj.dtype)
if self.sort_obj_iou:
j = iou.argsort()
b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j]
if self.gr < 1:
iou = (1.0 - self.gr) + self.gr * iou
tobj[b, a, gj, gi] = iou # iou ratio
# Classification
if self.nc > 1: # cls loss (only if multiple classes)
t = torch.full_like(pcls, self.cn, device=self.device) # targets
t[range(n), tcls[i]] = self.cp
lcls += self.BCEcls(pcls, t) # BCE
# Mask regression
if tuple(masks.shape[-2:]) != (mask_h, mask_w): # downsample
masks = F.interpolate(masks[None], (mask_h, mask_w), mode="nearest")[0]
marea = xywhn[i][:, 2:].prod(1) # mask width, height normalized
mxyxy = xywh2xyxy(xywhn[i] * torch.tensor([mask_w, mask_h, mask_w, mask_h], device=self.device))
for bi in b.unique():
j = b == bi # matching index
if self.overlap:
mask_gti = torch.where(masks[bi][None] == tidxs[i][j].view(-1, 1, 1), 1.0, 0.0)
else:
mask_gti = masks[tidxs[i]][j]
lseg += self.single_mask_loss(mask_gti, pmask[j], proto[bi], mxyxy[j], marea[j])
obji = self.BCEobj(pi[..., 4], tobj)
lobj += obji * self.balance[i] # obj loss
if self.autobalance:
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
if self.autobalance:
self.balance = [x / self.balance[self.ssi] for x in self.balance]
lbox *= self.hyp["box"]
lobj *= self.hyp["obj"]
lcls *= self.hyp["cls"]
lseg *= self.hyp["box"] / bs
loss = lbox + lobj + lcls + lseg
return loss * bs, torch.cat((lbox, lseg, lobj, lcls)).detach()
def single_mask_loss(self, gt_mask, pred, proto, xyxy, area):
# Mask loss for one image
pred_mask = (pred @ proto.view(self.nm, -1)).view(-1, *proto.shape[1:]) # (n,32) @ (32,80,80) -> (n,80,80)
loss = F.binary_cross_entropy_with_logits(pred_mask, gt_mask, reduction="none")
return (crop_mask(loss, xyxy).mean(dim=(1, 2)) / area).mean()
def build_targets(self, p, targets):
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
na, nt = self.na, targets.shape[0] # number of anchors, targets
tcls, tbox, indices, anch, tidxs, xywhn = [], [], [], [], [], []
gain = torch.ones(8, device=self.device) # normalized to gridspace gain
ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
if self.overlap:
batch = p[0].shape[0]
ti = []
for i in range(batch):
num = (targets[:, 0] == i).sum() # find number of targets of each image
ti.append(torch.arange(num, device=self.device).float().view(1, num).repeat(na, 1) + 1) # (na, num)
ti = torch.cat(ti, 1) # (na, nt)
else:
ti = torch.arange(nt, device=self.device).float().view(1, nt).repeat(na, 1)
targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None], ti[..., None]), 2) # append anchor indices
g = 0.5 # bias
off = torch.tensor(
[
[0, 0],
[1, 0],
[0, 1],
[-1, 0],
[0, -1], # j,k,l,m
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
],
device=self.device).float() * g # offsets
for i in range(self.nl):
anchors, shape = self.anchors[i], p[i].shape
gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]] # xyxy gain
# Match targets to anchors
t = targets * gain # shape(3,n,7)
if nt:
# Matches
r = t[..., 4:6] / anchors[:, None] # wh ratio
j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t'] # compare
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
t = t[j] # filter
# Offsets
gxy = t[:, 2:4] # grid xy
gxi = gain[[2, 3]] - gxy # inverse
j, k = ((gxy % 1 < g) & (gxy > 1)).T
l, m = ((gxi % 1 < g) & (gxi > 1)).T
j = torch.stack((torch.ones_like(j), j, k, l, m))
t = t.repeat((5, 1, 1))[j]
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
else:
t = targets[0]
offsets = 0
# Define
bc, gxy, gwh, at = t.chunk(4, 1) # (image, class), grid xy, grid wh, anchors
(a, tidx), (b, c) = at.long().T, bc.long().T # anchors, image, class
gij = (gxy - offsets).long()
gi, gj = gij.T # grid indices
# Append
indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1))) # image, anchor, grid
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
anch.append(anchors[a]) # anchors
tcls.append(c) # class
tidxs.append(tidx)
xywhn.append(torch.cat((gxy, gwh), 1) / gain[2:6]) # xywh normalized
return tcls, tbox, indices, anch, tidxs, xywhn
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment