Skip to content

GitLab

Commit 63e10e00 authored by limm's avatar limm
Browse files

Merge branch 'add_v0.6' into 'v0.6-release' 

support v0.6

See merge request !3
parents 2ca8e9fd b634945d
Hide whitespace changes
Inline Side-by-side
Showing with 4078 additions and 0 deletions
datasets/prepare_panoptic_fpn.py 0 → 100755
View file @ 63e10e00
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
import functools
import json
import multiprocessing as mp
import numpy as np
import os
import time
from fvcore.common.download import download
from panopticapi.utils import rgb2id
from PIL import Image
from detectron2.data.datasets.builtin_meta import COCO_CATEGORIES
def _process_panoptic_to_semantic(input_panoptic, output_semantic, segments, id_map):
panoptic = np.asarray(Image.open(input_panoptic), dtype=np.uint32)
panoptic = rgb2id(panoptic)
output = np.zeros_like(panoptic, dtype=np.uint8) + 255
for seg in segments:
cat_id = seg["category_id"]
new_cat_id = id_map[cat_id]
output[panoptic == seg["id"]] = new_cat_id
Image.fromarray(output).save(output_semantic)
def separate_coco_semantic_from_panoptic(panoptic_json, panoptic_root, sem_seg_root, categories):
"""
Create semantic segmentation annotations from panoptic segmentation
annotations, to be used by PanopticFPN.
It maps all thing categories to class 0, and maps all unlabeled pixels to class 255.
It maps all stuff categories to contiguous ids starting from 1.
Args:
panoptic_json (str): path to the panoptic json file, in COCO's format.
panoptic_root (str): a directory with panoptic annotation files, in COCO's format.
sem_seg_root (str): a directory to output semantic annotation files
categories (list[dict]): category metadata. Each dict needs to have:
"id": corresponds to the "category_id" in the json annotations
"isthing": 0 or 1
"""
os.makedirs(sem_seg_root, exist_ok=True)
stuff_ids = [k["id"] for k in categories if k["isthing"] == 0]
thing_ids = [k["id"] for k in categories if k["isthing"] == 1]
id_map = {} # map from category id to id in the output semantic annotation
assert len(stuff_ids) <= 254
for i, stuff_id in enumerate(stuff_ids):
id_map[stuff_id] = i + 1
for thing_id in thing_ids:
id_map[thing_id] = 0
id_map[0] = 255
with open(panoptic_json) as f:
obj = json.load(f)
pool = mp.Pool(processes=max(mp.cpu_count() // 2, 4))
def iter_annotations():
for anno in obj["annotations"]:
file_name = anno["file_name"]
segments = anno["segments_info"]
input = os.path.join(panoptic_root, file_name)
output = os.path.join(sem_seg_root, file_name)
yield input, output, segments
print("Start writing to {} ...".format(sem_seg_root))
start = time.time()
pool.starmap(
functools.partial(_process_panoptic_to_semantic, id_map=id_map),
iter_annotations(),
chunksize=100,
)
print("Finished. time: {:.2f}s".format(time.time() - start))
if __name__ == "__main__":
dataset_dir = os.path.join(os.getenv("DETECTRON2_DATASETS", "datasets"), "coco")
for s in ["val2017", "train2017"]:
separate_coco_semantic_from_panoptic(
os.path.join(dataset_dir, "annotations/panoptic_{}.json".format(s)),
os.path.join(dataset_dir, "panoptic_{}".format(s)),
os.path.join(dataset_dir, "panoptic_stuff_{}".format(s)),
COCO_CATEGORIES,
)
# Prepare val2017_100 for quick testing:
dest_dir = os.path.join(dataset_dir, "annotations/")
URL_PREFIX = "https://dl.fbaipublicfiles.com/detectron2/"
download(URL_PREFIX + "annotations/coco/panoptic_val2017_100.json", dest_dir)
with open(os.path.join(dest_dir, "panoptic_val2017_100.json")) as f:
obj = json.load(f)
def link_val100(dir_full, dir_100):
print("Creating " + dir_100 + " ...")
os.makedirs(dir_100, exist_ok=True)
for img in obj["images"]:
basename = os.path.splitext(img["file_name"])[0]
src = os.path.join(dir_full, basename + ".png")
dst = os.path.join(dir_100, basename + ".png")
src = os.path.relpath(src, start=dir_100)
os.symlink(src, dst)
link_val100(
os.path.join(dataset_dir, "panoptic_val2017"),
os.path.join(dataset_dir, "panoptic_val2017_100"),
)
link_val100(
os.path.join(dataset_dir, "panoptic_stuff_val2017"),
os.path.join(dataset_dir, "panoptic_stuff_val2017_100"),
)
demo/README.md 0 → 100644
View file @ 63e10e00
## Detectron2 Demo
We provide a command line tool to run a simple demo of builtin configs.
The usage is explained in [GETTING_STARTED.md](../GETTING_STARTED.md).
See our [blog post](https://ai.facebook.com/blog/-detectron2-a-pytorch-based-modular-object-detection-library-)
for a high-quality demo generated with this tool.
demo/demo.py 0 → 100755
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import argparse
import glob
import multiprocessing as mp
import numpy as np
import os
import tempfile
import time
import warnings
import cv2
import tqdm
from detectron2.config import get_cfg
from detectron2.data.detection_utils import read_image
from detectron2.utils.logger import setup_logger
from predictor import VisualizationDemo
# constants
WINDOW_NAME = "COCO detections"
def setup_cfg(args):
# load config from file and command-line arguments
cfg = get_cfg()
# To use demo for Panoptic-DeepLab, please uncomment the following two lines.
# from detectron2.projects.panoptic_deeplab import add_panoptic_deeplab_config # noqa
# add_panoptic_deeplab_config(cfg)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
# Set score_threshold for builtin models
cfg.MODEL.RETINANET.SCORE_THRESH_TEST = args.confidence_threshold
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = args.confidence_threshold
cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = args.confidence_threshold
cfg.freeze()
return cfg
def get_parser():
parser = argparse.ArgumentParser(description="Detectron2 demo for builtin configs")
parser.add_argument(
"--config-file",
default="configs/quick_schedules/mask_rcnn_R_50_FPN_inference_acc_test.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--webcam", action="store_true", help="Take inputs from webcam.")
parser.add_argument("--video-input", help="Path to video file.")
parser.add_argument(
"--input",
nargs="+",
help="A list of space separated input images; "
"or a single glob pattern such as 'directory/*.jpg'",
)
parser.add_argument(
"--output",
help="A file or directory to save output visualizations. "
"If not given, will show output in an OpenCV window.",
)
parser.add_argument(
"--confidence-threshold",
type=float,
default=0.5,
help="Minimum score for instance predictions to be shown",
)
parser.add_argument(
"--opts",
help="Modify config options using the command-line 'KEY VALUE' pairs",
default=[],
nargs=argparse.REMAINDER,
)
return parser
def test_opencv_video_format(codec, file_ext):
with tempfile.TemporaryDirectory(prefix="video_format_test") as dir:
filename = os.path.join(dir, "test_file" + file_ext)
writer = cv2.VideoWriter(
filename=filename,
fourcc=cv2.VideoWriter_fourcc(*codec),
fps=float(30),
frameSize=(10, 10),
isColor=True,
)
[writer.write(np.zeros((10, 10, 3), np.uint8)) for _ in range(30)]
writer.release()
if os.path.isfile(filename):
return True
return False
if __name__ == "__main__":
mp.set_start_method("spawn", force=True)
args = get_parser().parse_args()
setup_logger(name="fvcore")
logger = setup_logger()
logger.info("Arguments: " + str(args))
cfg = setup_cfg(args)
demo = VisualizationDemo(cfg)
if args.input:
if len(args.input) == 1:
args.input = glob.glob(os.path.expanduser(args.input[0]))
assert args.input, "The input path(s) was not found"
for path in tqdm.tqdm(args.input, disable=not args.output):
# use PIL, to be consistent with evaluation
img = read_image(path, format="BGR")
start_time = time.time()
predictions, visualized_output = demo.run_on_image(img)
logger.info(
"{}: {} in {:.2f}s".format(
path,
"detected {} instances".format(len(predictions["instances"]))
if "instances" in predictions
else "finished",
time.time() - start_time,
)
)
if args.output:
if os.path.isdir(args.output):
assert os.path.isdir(args.output), args.output
out_filename = os.path.join(args.output, os.path.basename(path))
else:
assert len(args.input) == 1, "Please specify a directory with args.output"
out_filename = args.output
visualized_output.save(out_filename)
else:
cv2.namedWindow(WINDOW_NAME, cv2.WINDOW_NORMAL)
cv2.imshow(WINDOW_NAME, visualized_output.get_image()[:, :, ::-1])
if cv2.waitKey(0) == 27:
break # esc to quit
elif args.webcam:
assert args.input is None, "Cannot have both --input and --webcam!"
assert args.output is None, "output not yet supported with --webcam!"
cam = cv2.VideoCapture(0)
for vis in tqdm.tqdm(demo.run_on_video(cam)):
cv2.namedWindow(WINDOW_NAME, cv2.WINDOW_NORMAL)
cv2.imshow(WINDOW_NAME, vis)
if cv2.waitKey(1) == 27:
break # esc to quit
cam.release()
cv2.destroyAllWindows()
elif args.video_input:
video = cv2.VideoCapture(args.video_input)
width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
frames_per_second = video.get(cv2.CAP_PROP_FPS)
num_frames = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
basename = os.path.basename(args.video_input)
codec, file_ext = (
("x264", ".mkv") if test_opencv_video_format("x264", ".mkv") else ("mp4v", ".mp4")
)
if codec == ".mp4v":
warnings.warn("x264 codec not available, switching to mp4v")
if args.output:
if os.path.isdir(args.output):
output_fname = os.path.join(args.output, basename)
output_fname = os.path.splitext(output_fname)[0] + file_ext
else:
output_fname = args.output
assert not os.path.isfile(output_fname), output_fname
output_file = cv2.VideoWriter(
filename=output_fname,
# some installation of opencv may not support x264 (due to its license),
# you can try other format (e.g. MPEG)
fourcc=cv2.VideoWriter_fourcc(*codec),
fps=float(frames_per_second),
frameSize=(width, height),
isColor=True,
)
assert os.path.isfile(args.video_input)
for vis_frame in tqdm.tqdm(demo.run_on_video(video), total=num_frames):
if args.output:
output_file.write(vis_frame)
else:
cv2.namedWindow(basename, cv2.WINDOW_NORMAL)
cv2.imshow(basename, vis_frame)
if cv2.waitKey(1) == 27:
break # esc to quit
video.release()
if args.output:
output_file.release()
else:
cv2.destroyAllWindows()
demo/predictor.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import atexit
import bisect
import multiprocessing as mp
from collections import deque
import cv2
import torch
from detectron2.data import MetadataCatalog
from detectron2.engine.defaults import DefaultPredictor
from detectron2.utils.video_visualizer import VideoVisualizer
from detectron2.utils.visualizer import ColorMode, Visualizer
class VisualizationDemo(object):
def __init__(self, cfg, instance_mode=ColorMode.IMAGE, parallel=False):
"""
Args:
cfg (CfgNode):
instance_mode (ColorMode):
parallel (bool): whether to run the model in different processes from visualization.
Useful since the visualization logic can be slow.
"""
self.metadata = MetadataCatalog.get(
cfg.DATASETS.TEST[0] if len(cfg.DATASETS.TEST) else "__unused"
)
self.cpu_device = torch.device("cpu")
self.instance_mode = instance_mode
self.parallel = parallel
if parallel:
num_gpu = torch.cuda.device_count()
self.predictor = AsyncPredictor(cfg, num_gpus=num_gpu)
else:
self.predictor = DefaultPredictor(cfg)
def run_on_image(self, image):
"""
Args:
image (np.ndarray): an image of shape (H, W, C) (in BGR order).
This is the format used by OpenCV.
Returns:
predictions (dict): the output of the model.
vis_output (VisImage): the visualized image output.
"""
vis_output = None
predictions = self.predictor(image)
# Convert image from OpenCV BGR format to Matplotlib RGB format.
image = image[:, :, ::-1]
visualizer = Visualizer(image, self.metadata, instance_mode=self.instance_mode)
if "panoptic_seg" in predictions:
panoptic_seg, segments_info = predictions["panoptic_seg"]
vis_output = visualizer.draw_panoptic_seg_predictions(
panoptic_seg.to(self.cpu_device), segments_info
)
else:
if "sem_seg" in predictions:
vis_output = visualizer.draw_sem_seg(
predictions["sem_seg"].argmax(dim=0).to(self.cpu_device)
)
if "instances" in predictions:
instances = predictions["instances"].to(self.cpu_device)
vis_output = visualizer.draw_instance_predictions(predictions=instances)
return predictions, vis_output
def _frame_from_video(self, video):
while video.isOpened():
success, frame = video.read()
if success:
yield frame
else:
break
def run_on_video(self, video):
"""
Visualizes predictions on frames of the input video.
Args:
video (cv2.VideoCapture): a :class:`VideoCapture` object, whose source can be
either a webcam or a video file.
Yields:
ndarray: BGR visualizations of each video frame.
"""
video_visualizer = VideoVisualizer(self.metadata, self.instance_mode)
def process_predictions(frame, predictions):
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
if "panoptic_seg" in predictions:
panoptic_seg, segments_info = predictions["panoptic_seg"]
vis_frame = video_visualizer.draw_panoptic_seg_predictions(
frame, panoptic_seg.to(self.cpu_device), segments_info
)
elif "instances" in predictions:
predictions = predictions["instances"].to(self.cpu_device)
vis_frame = video_visualizer.draw_instance_predictions(frame, predictions)
elif "sem_seg" in predictions:
vis_frame = video_visualizer.draw_sem_seg(
frame, predictions["sem_seg"].argmax(dim=0).to(self.cpu_device)
)
# Converts Matplotlib RGB format to OpenCV BGR format
vis_frame = cv2.cvtColor(vis_frame.get_image(), cv2.COLOR_RGB2BGR)
return vis_frame
frame_gen = self._frame_from_video(video)
if self.parallel:
buffer_size = self.predictor.default_buffer_size
frame_data = deque()
for cnt, frame in enumerate(frame_gen):
frame_data.append(frame)
self.predictor.put(frame)
if cnt >= buffer_size:
frame = frame_data.popleft()
predictions = self.predictor.get()
yield process_predictions(frame, predictions)
while len(frame_data):
frame = frame_data.popleft()
predictions = self.predictor.get()
yield process_predictions(frame, predictions)
else:
for frame in frame_gen:
yield process_predictions(frame, self.predictor(frame))
class AsyncPredictor:
"""
A predictor that runs the model asynchronously, possibly on >1 GPUs.
Because rendering the visualization takes considerably amount of time,
this helps improve throughput a little bit when rendering videos.
"""
class _StopToken:
pass
class _PredictWorker(mp.Process):
def __init__(self, cfg, task_queue, result_queue):
self.cfg = cfg
self.task_queue = task_queue
self.result_queue = result_queue
super().__init__()
def run(self):
predictor = DefaultPredictor(self.cfg)
while True:
task = self.task_queue.get()
if isinstance(task, AsyncPredictor._StopToken):
break
idx, data = task
result = predictor(data)
self.result_queue.put((idx, result))
def __init__(self, cfg, num_gpus: int = 1):
"""
Args:
cfg (CfgNode):
num_gpus (int): if 0, will run on CPU
"""
num_workers = max(num_gpus, 1)
self.task_queue = mp.Queue(maxsize=num_workers * 3)
self.result_queue = mp.Queue(maxsize=num_workers * 3)
self.procs = []
for gpuid in range(max(num_gpus, 1)):
cfg = cfg.clone()
cfg.defrost()
cfg.MODEL.DEVICE = "cuda:{}".format(gpuid) if num_gpus > 0 else "cpu"
self.procs.append(
AsyncPredictor._PredictWorker(cfg, self.task_queue, self.result_queue)
)
self.put_idx = 0
self.get_idx = 0
self.result_rank = []
self.result_data = []
for p in self.procs:
p.start()
atexit.register(self.shutdown)
def put(self, image):
self.put_idx += 1
self.task_queue.put((self.put_idx, image))
def get(self):
self.get_idx += 1 # the index needed for this request
if len(self.result_rank) and self.result_rank[0] == self.get_idx:
res = self.result_data[0]
del self.result_data[0], self.result_rank[0]
return res
while True:
# make sure the results are returned in the correct order
idx, res = self.result_queue.get()
if idx == self.get_idx:
return res
insert = bisect.bisect(self.result_rank, idx)
self.result_rank.insert(insert, idx)
self.result_data.insert(insert, res)
def __len__(self):
return self.put_idx - self.get_idx
def __call__(self, image):
self.put(image)
return self.get()
def shutdown(self):
for _ in self.procs:
self.task_queue.put(AsyncPredictor._StopToken())
@property
def default_buffer_size(self):
return len(self.procs) * 5
detectron2/__init__.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
from .utils.env import setup_environment
setup_environment()
# This line will be programatically read/write by setup.py.
# Leave them at the bottom of this file and don't touch them.
__version__ = "0.6"
detectron2/checkpoint/__init__.py 0 → 100644
View file @ 63e10e00
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
# File:
from . import catalog as _UNUSED # register the handler
from .detection_checkpoint import DetectionCheckpointer
from fvcore.common.checkpoint import Checkpointer, PeriodicCheckpointer
__all__ = ["Checkpointer", "PeriodicCheckpointer", "DetectionCheckpointer"]
detectron2/checkpoint/c2_model_loading.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import copy
import logging
import re
from typing import Dict, List
import torch
from tabulate import tabulate
def convert_basic_c2_names(original_keys):
"""
Apply some basic name conversion to names in C2 weights.
It only deals with typical backbone models.
Args:
original_keys (list[str]):
Returns:
list[str]: The same number of strings matching those in original_keys.
"""
layer_keys = copy.deepcopy(original_keys)
layer_keys = [
{"pred_b": "linear_b", "pred_w": "linear_w"}.get(k, k) for k in layer_keys
] # some hard-coded mappings
layer_keys = [k.replace("_", ".") for k in layer_keys]
layer_keys = [re.sub("\\.b$", ".bias", k) for k in layer_keys]
layer_keys = [re.sub("\\.w$", ".weight", k) for k in layer_keys]
# Uniform both bn and gn names to "norm"
layer_keys = [re.sub("bn\\.s$", "norm.weight", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.bias$", "norm.bias", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.rm", "norm.running_mean", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.running.mean$", "norm.running_mean", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.riv$", "norm.running_var", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.running.var$", "norm.running_var", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.gamma$", "norm.weight", k) for k in layer_keys]
layer_keys = [re.sub("bn\\.beta$", "norm.bias", k) for k in layer_keys]
layer_keys = [re.sub("gn\\.s$", "norm.weight", k) for k in layer_keys]
layer_keys = [re.sub("gn\\.bias$", "norm.bias", k) for k in layer_keys]
# stem
layer_keys = [re.sub("^res\\.conv1\\.norm\\.", "conv1.norm.", k) for k in layer_keys]
# to avoid mis-matching with "conv1" in other components (e.g. detection head)
layer_keys = [re.sub("^conv1\\.", "stem.conv1.", k) for k in layer_keys]
# layer1-4 is used by torchvision, however we follow the C2 naming strategy (res2-5)
# layer_keys = [re.sub("^res2.", "layer1.", k) for k in layer_keys]
# layer_keys = [re.sub("^res3.", "layer2.", k) for k in layer_keys]
# layer_keys = [re.sub("^res4.", "layer3.", k) for k in layer_keys]
# layer_keys = [re.sub("^res5.", "layer4.", k) for k in layer_keys]
# blocks
layer_keys = [k.replace(".branch1.", ".shortcut.") for k in layer_keys]
layer_keys = [k.replace(".branch2a.", ".conv1.") for k in layer_keys]
layer_keys = [k.replace(".branch2b.", ".conv2.") for k in layer_keys]
layer_keys = [k.replace(".branch2c.", ".conv3.") for k in layer_keys]
# DensePose substitutions
layer_keys = [re.sub("^body.conv.fcn", "body_conv_fcn", k) for k in layer_keys]
layer_keys = [k.replace("AnnIndex.lowres", "ann_index_lowres") for k in layer_keys]
layer_keys = [k.replace("Index.UV.lowres", "index_uv_lowres") for k in layer_keys]
layer_keys = [k.replace("U.lowres", "u_lowres") for k in layer_keys]
layer_keys = [k.replace("V.lowres", "v_lowres") for k in layer_keys]
return layer_keys
def convert_c2_detectron_names(weights):
"""
Map Caffe2 Detectron weight names to Detectron2 names.
Args:
weights (dict): name -> tensor
Returns:
dict: detectron2 names -> tensor
dict: detectron2 names -> C2 names
"""
logger = logging.getLogger(__name__)
logger.info("Renaming Caffe2 weights ......")
original_keys = sorted(weights.keys())
layer_keys = copy.deepcopy(original_keys)
layer_keys = convert_basic_c2_names(layer_keys)
# --------------------------------------------------------------------------
# RPN hidden representation conv
# --------------------------------------------------------------------------
# FPN case
# In the C2 model, the RPN hidden layer conv is defined for FPN level 2 and then
# shared for all other levels, hence the appearance of "fpn2"
layer_keys = [
k.replace("conv.rpn.fpn2", "proposal_generator.rpn_head.conv") for k in layer_keys
]
# Non-FPN case
layer_keys = [k.replace("conv.rpn", "proposal_generator.rpn_head.conv") for k in layer_keys]
# --------------------------------------------------------------------------
# RPN box transformation conv
# --------------------------------------------------------------------------
# FPN case (see note above about "fpn2")
layer_keys = [
k.replace("rpn.bbox.pred.fpn2", "proposal_generator.rpn_head.anchor_deltas")
for k in layer_keys
]
layer_keys = [
k.replace("rpn.cls.logits.fpn2", "proposal_generator.rpn_head.objectness_logits")
for k in layer_keys
]
# Non-FPN case
layer_keys = [
k.replace("rpn.bbox.pred", "proposal_generator.rpn_head.anchor_deltas") for k in layer_keys
]
layer_keys = [
k.replace("rpn.cls.logits", "proposal_generator.rpn_head.objectness_logits")
for k in layer_keys
]
# --------------------------------------------------------------------------
# Fast R-CNN box head
# --------------------------------------------------------------------------
layer_keys = [re.sub("^bbox\\.pred", "bbox_pred", k) for k in layer_keys]
layer_keys = [re.sub("^cls\\.score", "cls_score", k) for k in layer_keys]
layer_keys = [re.sub("^fc6\\.", "box_head.fc1.", k) for k in layer_keys]
layer_keys = [re.sub("^fc7\\.", "box_head.fc2.", k) for k in layer_keys]
# 4conv1fc head tensor names: head_conv1_w, head_conv1_gn_s
layer_keys = [re.sub("^head\\.conv", "box_head.conv", k) for k in layer_keys]
# --------------------------------------------------------------------------
# FPN lateral and output convolutions
# --------------------------------------------------------------------------
def fpn_map(name):
"""
Look for keys with the following patterns:
1) Starts with "fpn.inner."
Example: "fpn.inner.res2.2.sum.lateral.weight"
Meaning: These are lateral pathway convolutions
2) Starts with "fpn.res"
Example: "fpn.res2.2.sum.weight"
Meaning: These are FPN output convolutions
"""
splits = name.split(".")
norm = ".norm" if "norm" in splits else ""
if name.startswith("fpn.inner."):
# splits example: ['fpn', 'inner', 'res2', '2', 'sum', 'lateral', 'weight']
stage = int(splits[2][len("res") :])
return "fpn_lateral{}{}.{}".format(stage, norm, splits[-1])
elif name.startswith("fpn.res"):
# splits example: ['fpn', 'res2', '2', 'sum', 'weight']
stage = int(splits[1][len("res") :])
return "fpn_output{}{}.{}".format(stage, norm, splits[-1])
return name
layer_keys = [fpn_map(k) for k in layer_keys]
# --------------------------------------------------------------------------
# Mask R-CNN mask head
# --------------------------------------------------------------------------
# roi_heads.StandardROIHeads case
layer_keys = [k.replace(".[mask].fcn", "mask_head.mask_fcn") for k in layer_keys]
layer_keys = [re.sub("^\\.mask\\.fcn", "mask_head.mask_fcn", k) for k in layer_keys]
layer_keys = [k.replace("mask.fcn.logits", "mask_head.predictor") for k in layer_keys]
# roi_heads.Res5ROIHeads case
layer_keys = [k.replace("conv5.mask", "mask_head.deconv") for k in layer_keys]
# --------------------------------------------------------------------------
# Keypoint R-CNN head
# --------------------------------------------------------------------------
# interestingly, the keypoint head convs have blob names that are simply "conv_fcnX"
layer_keys = [k.replace("conv.fcn", "roi_heads.keypoint_head.conv_fcn") for k in layer_keys]
layer_keys = [
k.replace("kps.score.lowres", "roi_heads.keypoint_head.score_lowres") for k in layer_keys
]
layer_keys = [k.replace("kps.score.", "roi_heads.keypoint_head.score.") for k in layer_keys]
# --------------------------------------------------------------------------
# Done with replacements
# --------------------------------------------------------------------------
assert len(set(layer_keys)) == len(layer_keys)
assert len(original_keys) == len(layer_keys)
new_weights = {}
new_keys_to_original_keys = {}
for orig, renamed in zip(original_keys, layer_keys):
new_keys_to_original_keys[renamed] = orig
if renamed.startswith("bbox_pred.") or renamed.startswith("mask_head.predictor."):
# remove the meaningless prediction weight for background class
new_start_idx = 4 if renamed.startswith("bbox_pred.") else 1
new_weights[renamed] = weights[orig][new_start_idx:]
logger.info(
"Remove prediction weight for background class in {}. The shape changes from "
"{} to {}.".format(
renamed, tuple(weights[orig].shape), tuple(new_weights[renamed].shape)
)
)
elif renamed.startswith("cls_score."):
# move weights of bg class from original index 0 to last index
logger.info(
"Move classification weights for background class in {} from index 0 to "
"index {}.".format(renamed, weights[orig].shape[0] - 1)
)
new_weights[renamed] = torch.cat([weights[orig][1:], weights[orig][:1]])
else:
new_weights[renamed] = weights[orig]
return new_weights, new_keys_to_original_keys
# Note the current matching is not symmetric.
# it assumes model_state_dict will have longer names.
def align_and_update_state_dicts(model_state_dict, ckpt_state_dict, c2_conversion=True):
"""
Match names between the two state-dict, and returns a new chkpt_state_dict with names
converted to match model_state_dict with heuristics. The returned dict can be later
loaded with fvcore checkpointer.
If `c2_conversion==True`, `ckpt_state_dict` is assumed to be a Caffe2
model and will be renamed at first.
Strategy: suppose that the models that we will create will have prefixes appended
to each of its keys, for example due to an extra level of nesting that the original
pre-trained weights from ImageNet won't contain. For example, model.state_dict()
might return backbone[0].body.res2.conv1.weight, while the pre-trained model contains
res2.conv1.weight. We thus want to match both parameters together.
For that, we look for each model weight, look among all loaded keys if there is one
that is a suffix of the current weight name, and use it if that's the case.
If multiple matches exist, take the one with longest size
of the corresponding name. For example, for the same model as before, the pretrained
weight file can contain both res2.conv1.weight, as well as conv1.weight. In this case,
we want to match backbone[0].body.conv1.weight to conv1.weight, and
backbone[0].body.res2.conv1.weight to res2.conv1.weight.
"""
model_keys = sorted(model_state_dict.keys())
if c2_conversion:
ckpt_state_dict, original_keys = convert_c2_detectron_names(ckpt_state_dict)
# original_keys: the name in the original dict (before renaming)
else:
original_keys = {x: x for x in ckpt_state_dict.keys()}
ckpt_keys = sorted(ckpt_state_dict.keys())
def match(a, b):
# Matched ckpt_key should be a complete (starts with '.') suffix.
# For example, roi_heads.mesh_head.whatever_conv1 does not match conv1,
# but matches whatever_conv1 or mesh_head.whatever_conv1.
return a == b or a.endswith("." + b)
# get a matrix of string matches, where each (i, j) entry correspond to the size of the
# ckpt_key string, if it matches
match_matrix = [len(j) if match(i, j) else 0 for i in model_keys for j in ckpt_keys]
match_matrix = torch.as_tensor(match_matrix).view(len(model_keys), len(ckpt_keys))
# use the matched one with longest size in case of multiple matches
max_match_size, idxs = match_matrix.max(1)
# remove indices that correspond to no-match
idxs[max_match_size == 0] = -1
logger = logging.getLogger(__name__)
# matched_pairs (matched checkpoint key --> matched model key)
matched_keys = {}
result_state_dict = {}
for idx_model, idx_ckpt in enumerate(idxs.tolist()):
if idx_ckpt == -1:
continue
key_model = model_keys[idx_model]
key_ckpt = ckpt_keys[idx_ckpt]
value_ckpt = ckpt_state_dict[key_ckpt]
shape_in_model = model_state_dict[key_model].shape
if shape_in_model != value_ckpt.shape:
logger.warning(
"Shape of {} in checkpoint is {}, while shape of {} in model is {}.".format(
key_ckpt, value_ckpt.shape, key_model, shape_in_model
)
)
logger.warning(
"{} will not be loaded. Please double check and see if this is desired.".format(
key_ckpt
)
)
continue
assert key_model not in result_state_dict
result_state_dict[key_model] = value_ckpt
if key_ckpt in matched_keys: # already added to matched_keys
logger.error(
"Ambiguity found for {} in checkpoint!"
"It matches at least two keys in the model ({} and {}).".format(
key_ckpt, key_model, matched_keys[key_ckpt]
)
)
raise ValueError("Cannot match one checkpoint key to multiple keys in the model.")
matched_keys[key_ckpt] = key_model
# logging:
matched_model_keys = sorted(matched_keys.values())
if len(matched_model_keys) == 0:
logger.warning("No weights in checkpoint matched with model.")
return ckpt_state_dict
common_prefix = _longest_common_prefix(matched_model_keys)
rev_matched_keys = {v: k for k, v in matched_keys.items()}
original_keys = {k: original_keys[rev_matched_keys[k]] for k in matched_model_keys}
model_key_groups = _group_keys_by_module(matched_model_keys, original_keys)
table = []
memo = set()
for key_model in matched_model_keys:
if key_model in memo:
continue
if key_model in model_key_groups:
group = model_key_groups[key_model]
memo |= set(group)
shapes = [tuple(model_state_dict[k].shape) for k in group]
table.append(
(
_longest_common_prefix([k[len(common_prefix) :] for k in group]) + "*",
_group_str([original_keys[k] for k in group]),
" ".join([str(x).replace(" ", "") for x in shapes]),
)
)
else:
key_checkpoint = original_keys[key_model]
shape = str(tuple(model_state_dict[key_model].shape))
table.append((key_model[len(common_prefix) :], key_checkpoint, shape))
table_str = tabulate(
table, tablefmt="pipe", headers=["Names in Model", "Names in Checkpoint", "Shapes"]
)
logger.info(
"Following weights matched with "
+ (f"submodule {common_prefix[:-1]}" if common_prefix else "model")
+ ":\n"
+ table_str
)
unmatched_ckpt_keys = [k for k in ckpt_keys if k not in set(matched_keys.keys())]
for k in unmatched_ckpt_keys:
result_state_dict[k] = ckpt_state_dict[k]
return result_state_dict
def _group_keys_by_module(keys: List[str], original_names: Dict[str, str]):
"""
Params in the same submodule are grouped together.
Args:
keys: names of all parameters
original_names: mapping from parameter name to their name in the checkpoint
Returns:
dict[name -> all other names in the same group]
"""
def _submodule_name(key):
pos = key.rfind(".")
if pos < 0:
return None
prefix = key[: pos + 1]
return prefix
all_submodules = [_submodule_name(k) for k in keys]
all_submodules = [x for x in all_submodules if x]
all_submodules = sorted(all_submodules, key=len)
ret = {}
for prefix in all_submodules:
group = [k for k in keys if k.startswith(prefix)]
if len(group) <= 1:
continue
original_name_lcp = _longest_common_prefix_str([original_names[k] for k in group])
if len(original_name_lcp) == 0:
# don't group weights if original names don't share prefix
continue
for k in group:
if k in ret:
continue
ret[k] = group
return ret
def _longest_common_prefix(names: List[str]) -> str:
"""
["abc.zfg", "abc.zef"] -> "abc."
"""
names = [n.split(".") for n in names]
m1, m2 = min(names), max(names)
ret = [a for a, b in zip(m1, m2) if a == b]
ret = ".".join(ret) + "." if len(ret) else ""
return ret
def _longest_common_prefix_str(names: List[str]) -> str:
m1, m2 = min(names), max(names)
lcp = [a for a, b in zip(m1, m2) if a == b]
lcp = "".join(lcp)
return lcp
def _group_str(names: List[str]) -> str:
"""
Turn "common1", "common2", "common3" into "common{1,2,3}"
"""
lcp = _longest_common_prefix_str(names)
rest = [x[len(lcp) :] for x in names]
rest = "{" + ",".join(rest) + "}"
ret = lcp + rest
# add some simplification for BN specifically
ret = ret.replace("bn_{beta,running_mean,running_var,gamma}", "bn_*")
ret = ret.replace("bn_beta,bn_running_mean,bn_running_var,bn_gamma", "bn_*")
return ret
detectron2/checkpoint/catalog.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import logging
from detectron2.utils.file_io import PathHandler, PathManager
class ModelCatalog(object):
"""
Store mappings from names to third-party models.
"""
S3_C2_DETECTRON_PREFIX = "https://dl.fbaipublicfiles.com/detectron"
# MSRA models have STRIDE_IN_1X1=True. False otherwise.
# NOTE: all BN models here have fused BN into an affine layer.
# As a result, you should only load them to a model with "FrozenBN".
# Loading them to a model with regular BN or SyncBN is wrong.
# Even when loaded to FrozenBN, it is still different from affine by an epsilon,
# which should be negligible for training.
# NOTE: all models here uses PIXEL_STD=[1,1,1]
# NOTE: Most of the BN models here are no longer used. We use the
# re-converted pre-trained models under detectron2 model zoo instead.
C2_IMAGENET_MODELS = {
"MSRA/R-50": "ImageNetPretrained/MSRA/R-50.pkl",
"MSRA/R-101": "ImageNetPretrained/MSRA/R-101.pkl",
"FAIR/R-50-GN": "ImageNetPretrained/47261647/R-50-GN.pkl",
"FAIR/R-101-GN": "ImageNetPretrained/47592356/R-101-GN.pkl",
"FAIR/X-101-32x8d": "ImageNetPretrained/20171220/X-101-32x8d.pkl",
"FAIR/X-101-64x4d": "ImageNetPretrained/FBResNeXt/X-101-64x4d.pkl",
"FAIR/X-152-32x8d-IN5k": "ImageNetPretrained/25093814/X-152-32x8d-IN5k.pkl",
}
C2_DETECTRON_PATH_FORMAT = (
"{prefix}/{url}/output/train/{dataset}/{type}/model_final.pkl" # noqa B950
)
C2_DATASET_COCO = "coco_2014_train%3Acoco_2014_valminusminival"
C2_DATASET_COCO_KEYPOINTS = "keypoints_coco_2014_train%3Akeypoints_coco_2014_valminusminival"
# format: {model_name} -> part of the url
C2_DETECTRON_MODELS = {
"35857197/e2e_faster_rcnn_R-50-C4_1x": "35857197/12_2017_baselines/e2e_faster_rcnn_R-50-C4_1x.yaml.01_33_49.iAX0mXvW", # noqa B950
"35857345/e2e_faster_rcnn_R-50-FPN_1x": "35857345/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_1x.yaml.01_36_30.cUF7QR7I", # noqa B950
"35857890/e2e_faster_rcnn_R-101-FPN_1x": "35857890/12_2017_baselines/e2e_faster_rcnn_R-101-FPN_1x.yaml.01_38_50.sNxI7sX7", # noqa B950
"36761737/e2e_faster_rcnn_X-101-32x8d-FPN_1x": "36761737/12_2017_baselines/e2e_faster_rcnn_X-101-32x8d-FPN_1x.yaml.06_31_39.5MIHi1fZ", # noqa B950
"35858791/e2e_mask_rcnn_R-50-C4_1x": "35858791/12_2017_baselines/e2e_mask_rcnn_R-50-C4_1x.yaml.01_45_57.ZgkA7hPB", # noqa B950
"35858933/e2e_mask_rcnn_R-50-FPN_1x": "35858933/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml.01_48_14.DzEQe4wC", # noqa B950
"35861795/e2e_mask_rcnn_R-101-FPN_1x": "35861795/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_1x.yaml.02_31_37.KqyEK4tT", # noqa B950
"36761843/e2e_mask_rcnn_X-101-32x8d-FPN_1x": "36761843/12_2017_baselines/e2e_mask_rcnn_X-101-32x8d-FPN_1x.yaml.06_35_59.RZotkLKI", # noqa B950
"48616381/e2e_mask_rcnn_R-50-FPN_2x_gn": "GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q", # noqa B950
"37697547/e2e_keypoint_rcnn_R-50-FPN_1x": "37697547/12_2017_baselines/e2e_keypoint_rcnn_R-50-FPN_1x.yaml.08_42_54.kdzV35ao", # noqa B950
"35998355/rpn_R-50-C4_1x": "35998355/12_2017_baselines/rpn_R-50-C4_1x.yaml.08_00_43.njH5oD9L", # noqa B950
"35998814/rpn_R-50-FPN_1x": "35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179", # noqa B950
"36225147/fast_R-50-FPN_1x": "36225147/12_2017_baselines/fast_rcnn_R-50-FPN_1x.yaml.08_39_09.L3obSdQ2", # noqa B950
}
@staticmethod
def get(name):
if name.startswith("Caffe2Detectron/COCO"):
return ModelCatalog._get_c2_detectron_baseline(name)
if name.startswith("ImageNetPretrained/"):
return ModelCatalog._get_c2_imagenet_pretrained(name)
raise RuntimeError("model not present in the catalog: {}".format(name))
@staticmethod
def _get_c2_imagenet_pretrained(name):
prefix = ModelCatalog.S3_C2_DETECTRON_PREFIX
name = name[len("ImageNetPretrained/") :]
name = ModelCatalog.C2_IMAGENET_MODELS[name]
url = "/".join([prefix, name])
return url
@staticmethod
def _get_c2_detectron_baseline(name):
name = name[len("Caffe2Detectron/COCO/") :]
url = ModelCatalog.C2_DETECTRON_MODELS[name]
if "keypoint_rcnn" in name:
dataset = ModelCatalog.C2_DATASET_COCO_KEYPOINTS
else:
dataset = ModelCatalog.C2_DATASET_COCO
if "35998355/rpn_R-50-C4_1x" in name:
# this one model is somehow different from others ..
type = "rpn"
else:
type = "generalized_rcnn"
# Detectron C2 models are stored in the structure defined in `C2_DETECTRON_PATH_FORMAT`.
url = ModelCatalog.C2_DETECTRON_PATH_FORMAT.format(
prefix=ModelCatalog.S3_C2_DETECTRON_PREFIX, url=url, type=type, dataset=dataset
)
return url
class ModelCatalogHandler(PathHandler):
"""
Resolve URL like catalog://.
"""
PREFIX = "catalog://"
def _get_supported_prefixes(self):
return [self.PREFIX]
def _get_local_path(self, path, **kwargs):
logger = logging.getLogger(__name__)
catalog_path = ModelCatalog.get(path[len(self.PREFIX) :])
logger.info("Catalog entry {} points to {}".format(path, catalog_path))
return PathManager.get_local_path(catalog_path, **kwargs)
def _open(self, path, mode="r", **kwargs):
return PathManager.open(self._get_local_path(path), mode, **kwargs)
PathManager.register_handler(ModelCatalogHandler())
detectron2/checkpoint/detection_checkpoint.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import logging
import os
import pickle
import torch
from fvcore.common.checkpoint import Checkpointer
from torch.nn.parallel import DistributedDataParallel
import detectron2.utils.comm as comm
from detectron2.utils.file_io import PathManager
from .c2_model_loading import align_and_update_state_dicts
class DetectionCheckpointer(Checkpointer):
"""
Same as :class:`Checkpointer`, but is able to:
1. handle models in detectron & detectron2 model zoo, and apply conversions for legacy models.
2. correctly load checkpoints that are only available on the master worker
"""
def __init__(self, model, save_dir="", *, save_to_disk=None, **checkpointables):
is_main_process = comm.is_main_process()
super().__init__(
model,
save_dir,
save_to_disk=is_main_process if save_to_disk is None else save_to_disk,
**checkpointables,
)
self.path_manager = PathManager
def load(self, path, *args, **kwargs):
need_sync = False
if path and isinstance(self.model, DistributedDataParallel):
logger = logging.getLogger(__name__)
path = self.path_manager.get_local_path(path)
has_file = os.path.isfile(path)
all_has_file = comm.all_gather(has_file)
if not all_has_file[0]:
raise OSError(f"File {path} not found on main worker.")
if not all(all_has_file):
logger.warning(
f"Not all workers can read checkpoint {path}. "
"Training may fail to fully resume."
)
# TODO: broadcast the checkpoint file contents from main
# worker, and load from it instead.
need_sync = True
if not has_file:
path = None # don't load if not readable
ret = super().load(path, *args, **kwargs)
if need_sync:
logger.info("Broadcasting model states from main worker ...")
self.model._sync_params_and_buffers()
return ret
def _load_file(self, filename):
if filename.endswith(".pkl"):
with PathManager.open(filename, "rb") as f:
data = pickle.load(f, encoding="latin1")
if "model" in data and "__author__" in data:
# file is in Detectron2 model zoo format
self.logger.info("Reading a file from '{}'".format(data["__author__"]))
return data
else:
# assume file is from Caffe2 / Detectron1 model zoo
if "blobs" in data:
# Detection models have "blobs", but ImageNet models don't
data = data["blobs"]
data = {k: v for k, v in data.items() if not k.endswith("_momentum")}
return {"model": data, "__author__": "Caffe2", "matching_heuristics": True}
elif filename.endswith(".pyth"):
# assume file is from pycls; no one else seems to use the ".pyth" extension
with PathManager.open(filename, "rb") as f:
data = torch.load(f)
assert (
"model_state" in data
), f"Cannot load .pyth file {filename}; pycls checkpoints must contain 'model_state'."
model_state = {
k: v
for k, v in data["model_state"].items()
if not k.endswith("num_batches_tracked")
}
return {"model": model_state, "__author__": "pycls", "matching_heuristics": True}
loaded = super()._load_file(filename) # load native pth checkpoint
if "model" not in loaded:
loaded = {"model": loaded}
return loaded
def _load_model(self, checkpoint):
if checkpoint.get("matching_heuristics", False):
self._convert_ndarray_to_tensor(checkpoint["model"])
# convert weights by name-matching heuristics
checkpoint["model"] = align_and_update_state_dicts(
self.model.state_dict(),
checkpoint["model"],
c2_conversion=checkpoint.get("__author__", None) == "Caffe2",
)
# for non-caffe2 models, use standard ways to load it
incompatible = super()._load_model(checkpoint)
model_buffers = dict(self.model.named_buffers(recurse=False))
for k in ["pixel_mean", "pixel_std"]:
# Ignore missing key message about pixel_mean/std.
# Though they may be missing in old checkpoints, they will be correctly
# initialized from config anyway.
if k in model_buffers:
try:
incompatible.missing_keys.remove(k)
except ValueError:
pass
for k in incompatible.unexpected_keys[:]:
# Ignore unexpected keys about cell anchors. They exist in old checkpoints
# but now they are non-persistent buffers and will not be in new checkpoints.
if "anchor_generator.cell_anchors" in k:
incompatible.unexpected_keys.remove(k)
return incompatible
detectron2/config/__init__.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
from .compat import downgrade_config, upgrade_config
from .config import CfgNode, get_cfg, global_cfg, set_global_cfg, configurable
from .instantiate import instantiate
from .lazy import LazyCall, LazyConfig
__all__ = [
"CfgNode",
"get_cfg",
"global_cfg",
"set_global_cfg",
"downgrade_config",
"upgrade_config",
"configurable",
"instantiate",
"LazyCall",
"LazyConfig",
]
from detectron2.utils.env import fixup_module_metadata
fixup_module_metadata(__name__, globals(), __all__)
del fixup_module_metadata
detectron2/config/compat.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
"""
Backward compatibility of configs.
Instructions to bump version:
+ It's not needed to bump version if new keys are added.
It's only needed when backward-incompatible changes happen
(i.e., some existing keys disappear, or the meaning of a key changes)
+ To bump version, do the following:
1. Increment _C.VERSION in defaults.py
2. Add a converter in this file.
Each ConverterVX has a function "upgrade" which in-place upgrades config from X-1 to X,
and a function "downgrade" which in-place downgrades config from X to X-1
In each function, VERSION is left unchanged.
Each converter assumes that its input has the relevant keys
(i.e., the input is not a partial config).
3. Run the tests (test_config.py) to make sure the upgrade & downgrade
functions are consistent.
"""
import logging
from typing import List, Optional, Tuple
from .config import CfgNode as CN
from .defaults import _C
__all__ = ["upgrade_config", "downgrade_config"]
def upgrade_config(cfg: CN, to_version: Optional[int] = None) -> CN:
"""
Upgrade a config from its current version to a newer version.
Args:
cfg (CfgNode):
to_version (int): defaults to the latest version.
"""
cfg = cfg.clone()
if to_version is None:
to_version = _C.VERSION
assert cfg.VERSION <= to_version, "Cannot upgrade from v{} to v{}!".format(
cfg.VERSION, to_version
)
for k in range(cfg.VERSION, to_version):
converter = globals()["ConverterV" + str(k + 1)]
converter.upgrade(cfg)
cfg.VERSION = k + 1
return cfg
def downgrade_config(cfg: CN, to_version: int) -> CN:
"""
Downgrade a config from its current version to an older version.
Args:
cfg (CfgNode):
to_version (int):
Note:
A general downgrade of arbitrary configs is not always possible due to the
different functionalities in different versions.
The purpose of downgrade is only to recover the defaults in old versions,
allowing it to load an old partial yaml config.
Therefore, the implementation only needs to fill in the default values
in the old version when a general downgrade is not possible.
"""
cfg = cfg.clone()
assert cfg.VERSION >= to_version, "Cannot downgrade from v{} to v{}!".format(
cfg.VERSION, to_version
)
for k in range(cfg.VERSION, to_version, -1):
converter = globals()["ConverterV" + str(k)]
converter.downgrade(cfg)
cfg.VERSION = k - 1
return cfg
def guess_version(cfg: CN, filename: str) -> int:
"""
Guess the version of a partial config where the VERSION field is not specified.
Returns the version, or the latest if cannot make a guess.
This makes it easier for users to migrate.
"""
logger = logging.getLogger(__name__)
def _has(name: str) -> bool:
cur = cfg
for n in name.split("."):
if n not in cur:
return False
cur = cur[n]
return True
# Most users' partial configs have "MODEL.WEIGHT", so guess on it
ret = None
if _has("MODEL.WEIGHT") or _has("TEST.AUG_ON"):
ret = 1
if ret is not None:
logger.warning("Config '{}' has no VERSION. Assuming it to be v{}.".format(filename, ret))
else:
ret = _C.VERSION
logger.warning(
"Config '{}' has no VERSION. Assuming it to be compatible with latest v{}.".format(
filename, ret
)
)
return ret
def _rename(cfg: CN, old: str, new: str) -> None:
old_keys = old.split(".")
new_keys = new.split(".")
def _set(key_seq: List[str], val: str) -> None:
cur = cfg
for k in key_seq[:-1]:
if k not in cur:
cur[k] = CN()
cur = cur[k]
cur[key_seq[-1]] = val
def _get(key_seq: List[str]) -> CN:
cur = cfg
for k in key_seq:
cur = cur[k]
return cur
def _del(key_seq: List[str]) -> None:
cur = cfg
for k in key_seq[:-1]:
cur = cur[k]
del cur[key_seq[-1]]
if len(cur) == 0 and len(key_seq) > 1:
_del(key_seq[:-1])
_set(new_keys, _get(old_keys))
_del(old_keys)
class _RenameConverter:
"""
A converter that handles simple rename.
"""
RENAME: List[Tuple[str, str]] = [] # list of tuples of (old name, new name)
@classmethod
def upgrade(cls, cfg: CN) -> None:
for old, new in cls.RENAME:
_rename(cfg, old, new)
@classmethod
def downgrade(cls, cfg: CN) -> None:
for old, new in cls.RENAME[::-1]:
_rename(cfg, new, old)
class ConverterV1(_RenameConverter):
RENAME = [("MODEL.RPN_HEAD.NAME", "MODEL.RPN.HEAD_NAME")]
class ConverterV2(_RenameConverter):
"""
A large bulk of rename, before public release.
"""
RENAME = [
("MODEL.WEIGHT", "MODEL.WEIGHTS"),
("MODEL.PANOPTIC_FPN.SEMANTIC_LOSS_SCALE", "MODEL.SEM_SEG_HEAD.LOSS_WEIGHT"),
("MODEL.PANOPTIC_FPN.RPN_LOSS_SCALE", "MODEL.RPN.LOSS_WEIGHT"),
("MODEL.PANOPTIC_FPN.INSTANCE_LOSS_SCALE", "MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT"),
("MODEL.PANOPTIC_FPN.COMBINE_ON", "MODEL.PANOPTIC_FPN.COMBINE.ENABLED"),
(
"MODEL.PANOPTIC_FPN.COMBINE_OVERLAP_THRESHOLD",
"MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH",
),
(
"MODEL.PANOPTIC_FPN.COMBINE_STUFF_AREA_LIMIT",
"MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT",
),
(
"MODEL.PANOPTIC_FPN.COMBINE_INSTANCES_CONFIDENCE_THRESHOLD",
"MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH",
),
("MODEL.ROI_HEADS.SCORE_THRESH", "MODEL.ROI_HEADS.SCORE_THRESH_TEST"),
("MODEL.ROI_HEADS.NMS", "MODEL.ROI_HEADS.NMS_THRESH_TEST"),
("MODEL.RETINANET.INFERENCE_SCORE_THRESHOLD", "MODEL.RETINANET.SCORE_THRESH_TEST"),
("MODEL.RETINANET.INFERENCE_TOPK_CANDIDATES", "MODEL.RETINANET.TOPK_CANDIDATES_TEST"),
("MODEL.RETINANET.INFERENCE_NMS_THRESHOLD", "MODEL.RETINANET.NMS_THRESH_TEST"),
("TEST.DETECTIONS_PER_IMG", "TEST.DETECTIONS_PER_IMAGE"),
("TEST.AUG_ON", "TEST.AUG.ENABLED"),
("TEST.AUG_MIN_SIZES", "TEST.AUG.MIN_SIZES"),
("TEST.AUG_MAX_SIZE", "TEST.AUG.MAX_SIZE"),
("TEST.AUG_FLIP", "TEST.AUG.FLIP"),
]
@classmethod
def upgrade(cls, cfg: CN) -> None:
super().upgrade(cfg)
if cfg.MODEL.META_ARCHITECTURE == "RetinaNet":
_rename(
cfg, "MODEL.RETINANET.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS"
)
_rename(cfg, "MODEL.RETINANET.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
del cfg["MODEL"]["RPN"]["ANCHOR_SIZES"]
del cfg["MODEL"]["RPN"]["ANCHOR_ASPECT_RATIOS"]
else:
_rename(cfg, "MODEL.RPN.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS")
_rename(cfg, "MODEL.RPN.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
del cfg["MODEL"]["RETINANET"]["ANCHOR_SIZES"]
del cfg["MODEL"]["RETINANET"]["ANCHOR_ASPECT_RATIOS"]
del cfg["MODEL"]["RETINANET"]["ANCHOR_STRIDES"]
@classmethod
def downgrade(cls, cfg: CN) -> None:
super().downgrade(cfg)
_rename(cfg, "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS", "MODEL.RPN.ANCHOR_ASPECT_RATIOS")
_rename(cfg, "MODEL.ANCHOR_GENERATOR.SIZES", "MODEL.RPN.ANCHOR_SIZES")
cfg.MODEL.RETINANET.ANCHOR_ASPECT_RATIOS = cfg.MODEL.RPN.ANCHOR_ASPECT_RATIOS
cfg.MODEL.RETINANET.ANCHOR_SIZES = cfg.MODEL.RPN.ANCHOR_SIZES
cfg.MODEL.RETINANET.ANCHOR_STRIDES = [] # this is not used anywhere in any version
detectron2/config/config.py 0 → 100644
View file @ 63e10e00
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
import functools
import inspect
import logging
from fvcore.common.config import CfgNode as _CfgNode
from detectron2.utils.file_io import PathManager
class CfgNode(_CfgNode):
"""
The same as `fvcore.common.config.CfgNode`, but different in:
1. Use unsafe yaml loading by default.
Note that this may lead to arbitrary code execution: you must not
load a config file from untrusted sources before manually inspecting
the content of the file.
2. Support config versioning.
When attempting to merge an old config, it will convert the old config automatically.
.. automethod:: clone
.. automethod:: freeze
.. automethod:: defrost
.. automethod:: is_frozen
.. automethod:: load_yaml_with_base
.. automethod:: merge_from_list
.. automethod:: merge_from_other_cfg
"""
@classmethod
def _open_cfg(cls, filename):
return PathManager.open(filename, "r")
# Note that the default value of allow_unsafe is changed to True
def merge_from_file(self, cfg_filename: str, allow_unsafe: bool = True) -> None:
"""
Load content from the given config file and merge it into self.
Args:
cfg_filename: config filename
allow_unsafe: allow unsafe yaml syntax
"""
assert PathManager.isfile(cfg_filename), f"Config file '{cfg_filename}' does not exist!"
loaded_cfg = self.load_yaml_with_base(cfg_filename, allow_unsafe=allow_unsafe)
loaded_cfg = type(self)(loaded_cfg)
# defaults.py needs to import CfgNode
from .defaults import _C
latest_ver = _C.VERSION
assert (
latest_ver == self.VERSION
), "CfgNode.merge_from_file is only allowed on a config object of latest version!"
logger = logging.getLogger(__name__)
loaded_ver = loaded_cfg.get("VERSION", None)
if loaded_ver is None:
from .compat import guess_version
loaded_ver = guess_version(loaded_cfg, cfg_filename)
assert loaded_ver <= self.VERSION, "Cannot merge a v{} config into a v{} config.".format(
loaded_ver, self.VERSION
)
if loaded_ver == self.VERSION:
self.merge_from_other_cfg(loaded_cfg)
else:
# compat.py needs to import CfgNode
from .compat import upgrade_config, downgrade_config
logger.warning(
"Loading an old v{} config file '{}' by automatically upgrading to v{}. "
"See docs/CHANGELOG.md for instructions to update your files.".format(
loaded_ver, cfg_filename, self.VERSION
)
)
# To convert, first obtain a full config at an old version
old_self = downgrade_config(self, to_version=loaded_ver)
old_self.merge_from_other_cfg(loaded_cfg)
new_config = upgrade_config(old_self)
self.clear()
self.update(new_config)
def dump(self, *args, **kwargs):
"""
Returns:
str: a yaml string representation of the config
"""
# to make it show up in docs
return super().dump(*args, **kwargs)
global_cfg = CfgNode()
def get_cfg() -> CfgNode:
"""
Get a copy of the default config.
Returns:
a detectron2 CfgNode instance.
"""
from .defaults import _C
return _C.clone()
def set_global_cfg(cfg: CfgNode) -> None:
"""
Let the global config point to the given cfg.
Assume that the given "cfg" has the key "KEY", after calling
`set_global_cfg(cfg)`, the key can be accessed by:
::
from detectron2.config import global_cfg
print(global_cfg.KEY)
By using a hacky global config, you can access these configs anywhere,
without having to pass the config object or the values deep into the code.
This is a hacky feature introduced for quick prototyping / research exploration.
"""
global global_cfg
global_cfg.clear()
global_cfg.update(cfg)
def configurable(init_func=None, *, from_config=None):
"""
Decorate a function or a class's __init__ method so that it can be called
with a :class:`CfgNode` object using a :func:`from_config` function that translates
:class:`CfgNode` to arguments.
Examples:
::
# Usage 1: Decorator on __init__:
class A:
@configurable
def __init__(self, a, b=2, c=3):
pass
@classmethod
def from_config(cls, cfg): # 'cfg' must be the first argument
# Returns kwargs to be passed to __init__
return {"a": cfg.A, "b": cfg.B}
a1 = A(a=1, b=2) # regular construction
a2 = A(cfg) # construct with a cfg
a3 = A(cfg, b=3, c=4) # construct with extra overwrite
# Usage 2: Decorator on any function. Needs an extra from_config argument:
@configurable(from_config=lambda cfg: {"a: cfg.A, "b": cfg.B})
def a_func(a, b=2, c=3):
pass
a1 = a_func(a=1, b=2) # regular call
a2 = a_func(cfg) # call with a cfg
a3 = a_func(cfg, b=3, c=4) # call with extra overwrite
Args:
init_func (callable): a class's ``__init__`` method in usage 1. The
class must have a ``from_config`` classmethod which takes `cfg` as
the first argument.
from_config (callable): the from_config function in usage 2. It must take `cfg`
as its first argument.
"""
if init_func is not None:
assert (
inspect.isfunction(init_func)
and from_config is None
and init_func.__name__ == "__init__"
), "Incorrect use of @configurable. Check API documentation for examples."
@functools.wraps(init_func)
def wrapped(self, *args, **kwargs):
try:
from_config_func = type(self).from_config
except AttributeError as e:
raise AttributeError(
"Class with @configurable must have a 'from_config' classmethod."
) from e
if not inspect.ismethod(from_config_func):
raise TypeError("Class with @configurable must have a 'from_config' classmethod.")
if _called_with_cfg(*args, **kwargs):
explicit_args = _get_args_from_config(from_config_func, *args, **kwargs)
init_func(self, **explicit_args)
else:
init_func(self, *args, **kwargs)
return wrapped
else:
if from_config is None:
return configurable # @configurable() is made equivalent to @configurable
assert inspect.isfunction(
from_config
), "from_config argument of configurable must be a function!"
def wrapper(orig_func):
@functools.wraps(orig_func)
def wrapped(*args, **kwargs):
if _called_with_cfg(*args, **kwargs):
explicit_args = _get_args_from_config(from_config, *args, **kwargs)
return orig_func(**explicit_args)
else:
return orig_func(*args, **kwargs)
wrapped.from_config = from_config
return wrapped
return wrapper
def _get_args_from_config(from_config_func, *args, **kwargs):
"""
Use `from_config` to obtain explicit arguments.
Returns:
dict: arguments to be used for cls.__init__
"""
signature = inspect.signature(from_config_func)
if list(signature.parameters.keys())[0] != "cfg":
if inspect.isfunction(from_config_func):
name = from_config_func.__name__
else:
name = f"{from_config_func.__self__}.from_config"
raise TypeError(f"{name} must take 'cfg' as the first argument!")
support_var_arg = any(
param.kind in [param.VAR_POSITIONAL, param.VAR_KEYWORD]
for param in signature.parameters.values()
)
if support_var_arg: # forward all arguments to from_config, if from_config accepts them
ret = from_config_func(*args, **kwargs)
else:
# forward supported arguments to from_config
supported_arg_names = set(signature.parameters.keys())
extra_kwargs = {}
for name in list(kwargs.keys()):
if name not in supported_arg_names:
extra_kwargs[name] = kwargs.pop(name)
ret = from_config_func(*args, **kwargs)
# forward the other arguments to __init__
ret.update(extra_kwargs)
return ret
def _called_with_cfg(*args, **kwargs):
"""
Returns:
bool: whether the arguments contain CfgNode and should be considered
forwarded to from_config.
"""
from omegaconf import DictConfig
if len(args) and isinstance(args[0], (_CfgNode, DictConfig)):
return True
if isinstance(kwargs.pop("cfg", None), (_CfgNode, DictConfig)):
return True
# `from_config`'s first argument is forced to be "cfg".
# So the above check covers all cases.
return False
detectron2/config/defaults.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
from .config import CfgNode as CN
# NOTE: given the new config system
# (https://detectron2.readthedocs.io/en/latest/tutorials/lazyconfigs.html),
# we will stop adding new functionalities to default CfgNode.
# -----------------------------------------------------------------------------
# Convention about Training / Test specific parameters
# -----------------------------------------------------------------------------
# Whenever an argument can be either used for training or for testing, the
# corresponding name will be post-fixed by a _TRAIN for a training parameter,
# or _TEST for a test-specific parameter.
# For example, the number of images during training will be
# IMAGES_PER_BATCH_TRAIN, while the number of images for testing will be
# IMAGES_PER_BATCH_TEST
# -----------------------------------------------------------------------------
# Config definition
# -----------------------------------------------------------------------------
_C = CN()
# The version number, to upgrade from old configs to new ones if any
# changes happen. It's recommended to keep a VERSION in your config file.
_C.VERSION = 2
_C.MODEL = CN()
_C.MODEL.LOAD_PROPOSALS = False
_C.MODEL.MASK_ON = False
_C.MODEL.KEYPOINT_ON = False
_C.MODEL.DEVICE = "cuda"
_C.MODEL.META_ARCHITECTURE = "GeneralizedRCNN"
# Path (a file path, or URL like detectron2://.., https://..) to a checkpoint file
# to be loaded to the model. You can find available models in the model zoo.
_C.MODEL.WEIGHTS = ""
# Values to be used for image normalization (BGR order, since INPUT.FORMAT defaults to BGR).
# To train on images of different number of channels, just set different mean & std.
# Default values are the mean pixel value from ImageNet: [103.53, 116.28, 123.675]
_C.MODEL.PIXEL_MEAN = [103.530, 116.280, 123.675]
# When using pre-trained models in Detectron1 or any MSRA models,
# std has been absorbed into its conv1 weights, so the std needs to be set 1.
# Otherwise, you can use [57.375, 57.120, 58.395] (ImageNet std)
_C.MODEL.PIXEL_STD = [1.0, 1.0, 1.0]
# -----------------------------------------------------------------------------
# INPUT
# -----------------------------------------------------------------------------
_C.INPUT = CN()
# By default, {MIN,MAX}_SIZE options are used in transforms.ResizeShortestEdge.
# Please refer to ResizeShortestEdge for detailed definition.
# Size of the smallest side of the image during training
_C.INPUT.MIN_SIZE_TRAIN = (800,)
# Sample size of smallest side by choice or random selection from range give by
# INPUT.MIN_SIZE_TRAIN
_C.INPUT.MIN_SIZE_TRAIN_SAMPLING = "choice"
# Maximum size of the side of the image during training
_C.INPUT.MAX_SIZE_TRAIN = 1333
# Size of the smallest side of the image during testing. Set to zero to disable resize in testing.
_C.INPUT.MIN_SIZE_TEST = 800
# Maximum size of the side of the image during testing
_C.INPUT.MAX_SIZE_TEST = 1333
# Mode for flipping images used in data augmentation during training
# choose one of ["horizontal, "vertical", "none"]
_C.INPUT.RANDOM_FLIP = "horizontal"
# `True` if cropping is used for data augmentation during training
_C.INPUT.CROP = CN({"ENABLED": False})
# Cropping type. See documentation of `detectron2.data.transforms.RandomCrop` for explanation.
_C.INPUT.CROP.TYPE = "relative_range"
# Size of crop in range (0, 1] if CROP.TYPE is "relative" or "relative_range" and in number of
# pixels if CROP.TYPE is "absolute"
_C.INPUT.CROP.SIZE = [0.9, 0.9]
# Whether the model needs RGB, YUV, HSV etc.
# Should be one of the modes defined here, as we use PIL to read the image:
# https://pillow.readthedocs.io/en/stable/handbook/concepts.html#concept-modes
# with BGR being the one exception. One can set image format to BGR, we will
# internally use RGB for conversion and flip the channels over
_C.INPUT.FORMAT = "BGR"
# The ground truth mask format that the model will use.
# Mask R-CNN supports either "polygon" or "bitmask" as ground truth.
_C.INPUT.MASK_FORMAT = "polygon" # alternative: "bitmask"
# -----------------------------------------------------------------------------
# Dataset
# -----------------------------------------------------------------------------
_C.DATASETS = CN()
# List of the dataset names for training. Must be registered in DatasetCatalog
# Samples from these datasets will be merged and used as one dataset.
_C.DATASETS.TRAIN = ()
# List of the pre-computed proposal files for training, which must be consistent
# with datasets listed in DATASETS.TRAIN.
_C.DATASETS.PROPOSAL_FILES_TRAIN = ()
# Number of top scoring precomputed proposals to keep for training
_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN = 2000
# List of the dataset names for testing. Must be registered in DatasetCatalog
_C.DATASETS.TEST = ()
# List of the pre-computed proposal files for test, which must be consistent
# with datasets listed in DATASETS.TEST.
_C.DATASETS.PROPOSAL_FILES_TEST = ()
# Number of top scoring precomputed proposals to keep for test
_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST = 1000
# -----------------------------------------------------------------------------
# DataLoader
# -----------------------------------------------------------------------------
_C.DATALOADER = CN()
# Number of data loading threads
_C.DATALOADER.NUM_WORKERS = 4
# If True, each batch should contain only images for which the aspect ratio
# is compatible. This groups portrait images together, and landscape images
# are not batched with portrait images.
_C.DATALOADER.ASPECT_RATIO_GROUPING = True
# Options: TrainingSampler, RepeatFactorTrainingSampler
_C.DATALOADER.SAMPLER_TRAIN = "TrainingSampler"
# Repeat threshold for RepeatFactorTrainingSampler
_C.DATALOADER.REPEAT_THRESHOLD = 0.0
# Tf True, when working on datasets that have instance annotations, the
# training dataloader will filter out images without associated annotations
_C.DATALOADER.FILTER_EMPTY_ANNOTATIONS = True
# ---------------------------------------------------------------------------- #
# Backbone options
# ---------------------------------------------------------------------------- #
_C.MODEL.BACKBONE = CN()
_C.MODEL.BACKBONE.NAME = "build_resnet_backbone"
# Freeze the first several stages so they are not trained.
# There are 5 stages in ResNet. The first is a convolution, and the following
# stages are each group of residual blocks.
_C.MODEL.BACKBONE.FREEZE_AT = 2
# ---------------------------------------------------------------------------- #
# FPN options
# ---------------------------------------------------------------------------- #
_C.MODEL.FPN = CN()
# Names of the input feature maps to be used by FPN
# They must have contiguous power of 2 strides
# e.g., ["res2", "res3", "res4", "res5"]
_C.MODEL.FPN.IN_FEATURES = []
_C.MODEL.FPN.OUT_CHANNELS = 256
# Options: "" (no norm), "GN"
_C.MODEL.FPN.NORM = ""
# Types for fusing the FPN top-down and lateral features. Can be either "sum" or "avg"
_C.MODEL.FPN.FUSE_TYPE = "sum"
# ---------------------------------------------------------------------------- #
# Proposal generator options
# ---------------------------------------------------------------------------- #
_C.MODEL.PROPOSAL_GENERATOR = CN()
# Current proposal generators include "RPN", "RRPN" and "PrecomputedProposals"
_C.MODEL.PROPOSAL_GENERATOR.NAME = "RPN"
# Proposal height and width both need to be greater than MIN_SIZE
# (a the scale used during training or inference)
_C.MODEL.PROPOSAL_GENERATOR.MIN_SIZE = 0
# ---------------------------------------------------------------------------- #
# Anchor generator options
# ---------------------------------------------------------------------------- #
_C.MODEL.ANCHOR_GENERATOR = CN()
# The generator can be any name in the ANCHOR_GENERATOR registry
_C.MODEL.ANCHOR_GENERATOR.NAME = "DefaultAnchorGenerator"
# Anchor sizes (i.e. sqrt of area) in absolute pixels w.r.t. the network input.
# Format: list[list[float]]. SIZES[i] specifies the list of sizes to use for
# IN_FEATURES[i]; len(SIZES) must be equal to len(IN_FEATURES) or 1.
# When len(SIZES) == 1, SIZES[0] is used for all IN_FEATURES.
_C.MODEL.ANCHOR_GENERATOR.SIZES = [[32, 64, 128, 256, 512]]
# Anchor aspect ratios. For each area given in `SIZES`, anchors with different aspect
# ratios are generated by an anchor generator.
# Format: list[list[float]]. ASPECT_RATIOS[i] specifies the list of aspect ratios (H/W)
# to use for IN_FEATURES[i]; len(ASPECT_RATIOS) == len(IN_FEATURES) must be true,
# or len(ASPECT_RATIOS) == 1 is true and aspect ratio list ASPECT_RATIOS[0] is used
# for all IN_FEATURES.
_C.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS = [[0.5, 1.0, 2.0]]
# Anchor angles.
# list[list[float]], the angle in degrees, for each input feature map.
# ANGLES[i] specifies the list of angles for IN_FEATURES[i].
_C.MODEL.ANCHOR_GENERATOR.ANGLES = [[-90, 0, 90]]
# Relative offset between the center of the first anchor and the top-left corner of the image
# Value has to be in [0, 1). Recommend to use 0.5, which means half stride.
# The value is not expected to affect model accuracy.
_C.MODEL.ANCHOR_GENERATOR.OFFSET = 0.0
# ---------------------------------------------------------------------------- #
# RPN options
# ---------------------------------------------------------------------------- #
_C.MODEL.RPN = CN()
_C.MODEL.RPN.HEAD_NAME = "StandardRPNHead" # used by RPN_HEAD_REGISTRY
# Names of the input feature maps to be used by RPN
# e.g., ["p2", "p3", "p4", "p5", "p6"] for FPN
_C.MODEL.RPN.IN_FEATURES = ["res4"]
# Remove RPN anchors that go outside the image by BOUNDARY_THRESH pixels
# Set to -1 or a large value, e.g. 100000, to disable pruning anchors
_C.MODEL.RPN.BOUNDARY_THRESH = -1
# IOU overlap ratios [BG_IOU_THRESHOLD, FG_IOU_THRESHOLD]
# Minimum overlap required between an anchor and ground-truth box for the
# (anchor, gt box) pair to be a positive example (IoU >= FG_IOU_THRESHOLD
# ==> positive RPN example: 1)
# Maximum overlap allowed between an anchor and ground-truth box for the
# (anchor, gt box) pair to be a negative examples (IoU < BG_IOU_THRESHOLD
# ==> negative RPN example: 0)
# Anchors with overlap in between (BG_IOU_THRESHOLD <= IoU < FG_IOU_THRESHOLD)
# are ignored (-1)
_C.MODEL.RPN.IOU_THRESHOLDS = [0.3, 0.7]
_C.MODEL.RPN.IOU_LABELS = [0, -1, 1]
# Number of regions per image used to train RPN
_C.MODEL.RPN.BATCH_SIZE_PER_IMAGE = 256
# Target fraction of foreground (positive) examples per RPN minibatch
_C.MODEL.RPN.POSITIVE_FRACTION = 0.5
# Options are: "smooth_l1", "giou", "diou", "ciou"
_C.MODEL.RPN.BBOX_REG_LOSS_TYPE = "smooth_l1"
_C.MODEL.RPN.BBOX_REG_LOSS_WEIGHT = 1.0
# Weights on (dx, dy, dw, dh) for normalizing RPN anchor regression targets
_C.MODEL.RPN.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
_C.MODEL.RPN.SMOOTH_L1_BETA = 0.0
_C.MODEL.RPN.LOSS_WEIGHT = 1.0
# Number of top scoring RPN proposals to keep before applying NMS
# When FPN is used, this is *per FPN level* (not total)
_C.MODEL.RPN.PRE_NMS_TOPK_TRAIN = 12000
_C.MODEL.RPN.PRE_NMS_TOPK_TEST = 6000
# Number of top scoring RPN proposals to keep after applying NMS
# When FPN is used, this limit is applied per level and then again to the union
# of proposals from all levels
# NOTE: When FPN is used, the meaning of this config is different from Detectron1.
# It means per-batch topk in Detectron1, but per-image topk here.
# See the "find_top_rpn_proposals" function for details.
_C.MODEL.RPN.POST_NMS_TOPK_TRAIN = 2000
_C.MODEL.RPN.POST_NMS_TOPK_TEST = 1000
# NMS threshold used on RPN proposals
_C.MODEL.RPN.NMS_THRESH = 0.7
# Set this to -1 to use the same number of output channels as input channels.
_C.MODEL.RPN.CONV_DIMS = [-1]
# ---------------------------------------------------------------------------- #
# ROI HEADS options
# ---------------------------------------------------------------------------- #
_C.MODEL.ROI_HEADS = CN()
_C.MODEL.ROI_HEADS.NAME = "Res5ROIHeads"
# Number of foreground classes
_C.MODEL.ROI_HEADS.NUM_CLASSES = 80
# Names of the input feature maps to be used by ROI heads
# Currently all heads (box, mask, ...) use the same input feature map list
# e.g., ["p2", "p3", "p4", "p5"] is commonly used for FPN
_C.MODEL.ROI_HEADS.IN_FEATURES = ["res4"]
# IOU overlap ratios [IOU_THRESHOLD]
# Overlap threshold for an RoI to be considered background (if < IOU_THRESHOLD)
# Overlap threshold for an RoI to be considered foreground (if >= IOU_THRESHOLD)
_C.MODEL.ROI_HEADS.IOU_THRESHOLDS = [0.5]
_C.MODEL.ROI_HEADS.IOU_LABELS = [0, 1]
# RoI minibatch size *per image* (number of regions of interest [ROIs]) during training
# Total number of RoIs per training minibatch =
# ROI_HEADS.BATCH_SIZE_PER_IMAGE * SOLVER.IMS_PER_BATCH
# E.g., a common configuration is: 512 * 16 = 8192
_C.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512
# Target fraction of RoI minibatch that is labeled foreground (i.e. class > 0)
_C.MODEL.ROI_HEADS.POSITIVE_FRACTION = 0.25
# Only used on test mode
# Minimum score threshold (assuming scores in a [0, 1] range); a value chosen to
# balance obtaining high recall with not having too many low precision
# detections that will slow down inference post processing steps (like NMS)
# A default threshold of 0.0 increases AP by ~0.2-0.3 but significantly slows down
# inference.
_C.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.05
# Overlap threshold used for non-maximum suppression (suppress boxes with
# IoU >= this threshold)
_C.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.5
# If True, augment proposals with ground-truth boxes before sampling proposals to
# train ROI heads.
_C.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT = True
# ---------------------------------------------------------------------------- #
# Box Head
# ---------------------------------------------------------------------------- #
_C.MODEL.ROI_BOX_HEAD = CN()
# C4 don't use head name option
# Options for non-C4 models: FastRCNNConvFCHead,
_C.MODEL.ROI_BOX_HEAD.NAME = ""
# Options are: "smooth_l1", "giou", "diou", "ciou"
_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_TYPE = "smooth_l1"
# The final scaling coefficient on the box regression loss, used to balance the magnitude of its
# gradients with other losses in the model. See also `MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT`.
_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_WEIGHT = 1.0
# Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets
# These are empirically chosen to approximately lead to unit variance targets
_C.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS = (10.0, 10.0, 5.0, 5.0)
# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
_C.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA = 0.0
_C.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 14
_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0
# Type of pooling operation applied to the incoming feature map for each RoI
_C.MODEL.ROI_BOX_HEAD.POOLER_TYPE = "ROIAlignV2"
_C.MODEL.ROI_BOX_HEAD.NUM_FC = 0
# Hidden layer dimension for FC layers in the RoI box head
_C.MODEL.ROI_BOX_HEAD.FC_DIM = 1024
_C.MODEL.ROI_BOX_HEAD.NUM_CONV = 0
# Channel dimension for Conv layers in the RoI box head
_C.MODEL.ROI_BOX_HEAD.CONV_DIM = 256
# Normalization method for the convolution layers.
# Options: "" (no norm), "GN", "SyncBN".
_C.MODEL.ROI_BOX_HEAD.NORM = ""
# Whether to use class agnostic for bbox regression
_C.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG = False
# If true, RoI heads use bounding boxes predicted by the box head rather than proposal boxes.
_C.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES = False
# ---------------------------------------------------------------------------- #
# Cascaded Box Head
# ---------------------------------------------------------------------------- #
_C.MODEL.ROI_BOX_CASCADE_HEAD = CN()
# The number of cascade stages is implicitly defined by the length of the following two configs.
_C.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS = (
(10.0, 10.0, 5.0, 5.0),
(20.0, 20.0, 10.0, 10.0),
(30.0, 30.0, 15.0, 15.0),
)
_C.MODEL.ROI_BOX_CASCADE_HEAD.IOUS = (0.5, 0.6, 0.7)
# ---------------------------------------------------------------------------- #
# Mask Head
# ---------------------------------------------------------------------------- #
_C.MODEL.ROI_MASK_HEAD = CN()
_C.MODEL.ROI_MASK_HEAD.NAME = "MaskRCNNConvUpsampleHead"
_C.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION = 14
_C.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO = 0
_C.MODEL.ROI_MASK_HEAD.NUM_CONV = 0 # The number of convs in the mask head
_C.MODEL.ROI_MASK_HEAD.CONV_DIM = 256
# Normalization method for the convolution layers.
# Options: "" (no norm), "GN", "SyncBN".
_C.MODEL.ROI_MASK_HEAD.NORM = ""
# Whether to use class agnostic for mask prediction
_C.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK = False
# Type of pooling operation applied to the incoming feature map for each RoI
_C.MODEL.ROI_MASK_HEAD.POOLER_TYPE = "ROIAlignV2"
# ---------------------------------------------------------------------------- #
# Keypoint Head
# ---------------------------------------------------------------------------- #
_C.MODEL.ROI_KEYPOINT_HEAD = CN()
_C.MODEL.ROI_KEYPOINT_HEAD.NAME = "KRCNNConvDeconvUpsampleHead"
_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION = 14
_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO = 0
_C.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS = tuple(512 for _ in range(8))
_C.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = 17 # 17 is the number of keypoints in COCO.
# Images with too few (or no) keypoints are excluded from training.
_C.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE = 1
# Normalize by the total number of visible keypoints in the minibatch if True.
# Otherwise, normalize by the total number of keypoints that could ever exist
# in the minibatch.
# The keypoint softmax loss is only calculated on visible keypoints.
# Since the number of visible keypoints can vary significantly between
# minibatches, this has the effect of up-weighting the importance of
# minibatches with few visible keypoints. (Imagine the extreme case of
# only one visible keypoint versus N: in the case of N, each one
# contributes 1/N to the gradient compared to the single keypoint
# determining the gradient direction). Instead, we can normalize the
# loss by the total number of keypoints, if it were the case that all
# keypoints were visible in a full minibatch. (Returning to the example,
# this means that the one visible keypoint contributes as much as each
# of the N keypoints.)
_C.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS = True
# Multi-task loss weight to use for keypoints
# Recommended values:
# - use 1.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is True
# - use 4.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is False
_C.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT = 1.0
# Type of pooling operation applied to the incoming feature map for each RoI
_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE = "ROIAlignV2"
# ---------------------------------------------------------------------------- #
# Semantic Segmentation Head
# ---------------------------------------------------------------------------- #
_C.MODEL.SEM_SEG_HEAD = CN()
_C.MODEL.SEM_SEG_HEAD.NAME = "SemSegFPNHead"
_C.MODEL.SEM_SEG_HEAD.IN_FEATURES = ["p2", "p3", "p4", "p5"]
# Label in the semantic segmentation ground truth that is ignored, i.e., no loss is calculated for
# the correposnding pixel.
_C.MODEL.SEM_SEG_HEAD.IGNORE_VALUE = 255
# Number of classes in the semantic segmentation head
_C.MODEL.SEM_SEG_HEAD.NUM_CLASSES = 54
# Number of channels in the 3x3 convs inside semantic-FPN heads.
_C.MODEL.SEM_SEG_HEAD.CONVS_DIM = 128
# Outputs from semantic-FPN heads are up-scaled to the COMMON_STRIDE stride.
_C.MODEL.SEM_SEG_HEAD.COMMON_STRIDE = 4
# Normalization method for the convolution layers. Options: "" (no norm), "GN".
_C.MODEL.SEM_SEG_HEAD.NORM = "GN"
_C.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT = 1.0
_C.MODEL.PANOPTIC_FPN = CN()
# Scaling of all losses from instance detection / segmentation head.
_C.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT = 1.0
# options when combining instance & semantic segmentation outputs
_C.MODEL.PANOPTIC_FPN.COMBINE = CN({"ENABLED": True}) # "COMBINE.ENABLED" is deprecated & not used
_C.MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH = 0.5
_C.MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT = 4096
_C.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = 0.5
# ---------------------------------------------------------------------------- #
# RetinaNet Head
# ---------------------------------------------------------------------------- #
_C.MODEL.RETINANET = CN()
# This is the number of foreground classes.
_C.MODEL.RETINANET.NUM_CLASSES = 80
_C.MODEL.RETINANET.IN_FEATURES = ["p3", "p4", "p5", "p6", "p7"]
# Convolutions to use in the cls and bbox tower
# NOTE: this doesn't include the last conv for logits
_C.MODEL.RETINANET.NUM_CONVS = 4
# IoU overlap ratio [bg, fg] for labeling anchors.
# Anchors with < bg are labeled negative (0)
# Anchors with >= bg and < fg are ignored (-1)
# Anchors with >= fg are labeled positive (1)
_C.MODEL.RETINANET.IOU_THRESHOLDS = [0.4, 0.5]
_C.MODEL.RETINANET.IOU_LABELS = [0, -1, 1]
# Prior prob for rare case (i.e. foreground) at the beginning of training.
# This is used to set the bias for the logits layer of the classifier subnet.
# This improves training stability in the case of heavy class imbalance.
_C.MODEL.RETINANET.PRIOR_PROB = 0.01
# Inference cls score threshold, only anchors with score > INFERENCE_TH are
# considered for inference (to improve speed)
_C.MODEL.RETINANET.SCORE_THRESH_TEST = 0.05
# Select topk candidates before NMS
_C.MODEL.RETINANET.TOPK_CANDIDATES_TEST = 1000
_C.MODEL.RETINANET.NMS_THRESH_TEST = 0.5
# Weights on (dx, dy, dw, dh) for normalizing Retinanet anchor regression targets
_C.MODEL.RETINANET.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
# Loss parameters
_C.MODEL.RETINANET.FOCAL_LOSS_GAMMA = 2.0
_C.MODEL.RETINANET.FOCAL_LOSS_ALPHA = 0.25
_C.MODEL.RETINANET.SMOOTH_L1_LOSS_BETA = 0.1
# Options are: "smooth_l1", "giou", "diou", "ciou"
_C.MODEL.RETINANET.BBOX_REG_LOSS_TYPE = "smooth_l1"
# One of BN, SyncBN, FrozenBN, GN
# Only supports GN until unshared norm is implemented
_C.MODEL.RETINANET.NORM = ""
# ---------------------------------------------------------------------------- #
# ResNe[X]t options (ResNets = {ResNet, ResNeXt}
# Note that parts of a resnet may be used for both the backbone and the head
# These options apply to both
# ---------------------------------------------------------------------------- #
_C.MODEL.RESNETS = CN()
_C.MODEL.RESNETS.DEPTH = 50
_C.MODEL.RESNETS.OUT_FEATURES = ["res4"] # res4 for C4 backbone, res2..5 for FPN backbone
# Number of groups to use; 1 ==> ResNet; > 1 ==> ResNeXt
_C.MODEL.RESNETS.NUM_GROUPS = 1
# Options: FrozenBN, GN, "SyncBN", "BN"
_C.MODEL.RESNETS.NORM = "FrozenBN"
# Baseline width of each group.
# Scaling this parameters will scale the width of all bottleneck layers.
_C.MODEL.RESNETS.WIDTH_PER_GROUP = 64
# Place the stride 2 conv on the 1x1 filter
# Use True only for the original MSRA ResNet; use False for C2 and Torch models
_C.MODEL.RESNETS.STRIDE_IN_1X1 = True
# Apply dilation in stage "res5"
_C.MODEL.RESNETS.RES5_DILATION = 1
# Output width of res2. Scaling this parameters will scale the width of all 1x1 convs in ResNet
# For R18 and R34, this needs to be set to 64
_C.MODEL.RESNETS.RES2_OUT_CHANNELS = 256
_C.MODEL.RESNETS.STEM_OUT_CHANNELS = 64
# Apply Deformable Convolution in stages
# Specify if apply deform_conv on Res2, Res3, Res4, Res5
_C.MODEL.RESNETS.DEFORM_ON_PER_STAGE = [False, False, False, False]
# Use True to use modulated deform_conv (DeformableV2, https://arxiv.org/abs/1811.11168);
# Use False for DeformableV1.
_C.MODEL.RESNETS.DEFORM_MODULATED = False
# Number of groups in deformable conv.
_C.MODEL.RESNETS.DEFORM_NUM_GROUPS = 1
# ---------------------------------------------------------------------------- #
# Solver
# ---------------------------------------------------------------------------- #
_C.SOLVER = CN()
# Options: WarmupMultiStepLR, WarmupCosineLR.
# See detectron2/solver/build.py for definition.
_C.SOLVER.LR_SCHEDULER_NAME = "WarmupMultiStepLR"
_C.SOLVER.MAX_ITER = 40000
_C.SOLVER.BASE_LR = 0.001
_C.SOLVER.MOMENTUM = 0.9
_C.SOLVER.NESTEROV = False
_C.SOLVER.WEIGHT_DECAY = 0.0001
# The weight decay that's applied to parameters of normalization layers
# (typically the affine transformation)
_C.SOLVER.WEIGHT_DECAY_NORM = 0.0
_C.SOLVER.GAMMA = 0.1
# The iteration number to decrease learning rate by GAMMA.
_C.SOLVER.STEPS = (30000,)
_C.SOLVER.WARMUP_FACTOR = 1.0 / 1000
_C.SOLVER.WARMUP_ITERS = 1000
_C.SOLVER.WARMUP_METHOD = "linear"
# Save a checkpoint after every this number of iterations
_C.SOLVER.CHECKPOINT_PERIOD = 5000
# Number of images per batch across all machines. This is also the number
# of training images per step (i.e. per iteration). If we use 16 GPUs
# and IMS_PER_BATCH = 32, each GPU will see 2 images per batch.
# May be adjusted automatically if REFERENCE_WORLD_SIZE is set.
_C.SOLVER.IMS_PER_BATCH = 16
# The reference number of workers (GPUs) this config is meant to train with.
# It takes no effect when set to 0.
# With a non-zero value, it will be used by DefaultTrainer to compute a desired
# per-worker batch size, and then scale the other related configs (total batch size,
# learning rate, etc) to match the per-worker batch size.
# See documentation of `DefaultTrainer.auto_scale_workers` for details:
_C.SOLVER.REFERENCE_WORLD_SIZE = 0
# Detectron v1 (and previous detection code) used a 2x higher LR and 0 WD for
# biases. This is not useful (at least for recent models). You should avoid
# changing these and they exist only to reproduce Detectron v1 training if
# desired.
_C.SOLVER.BIAS_LR_FACTOR = 1.0
_C.SOLVER.WEIGHT_DECAY_BIAS = None # None means following WEIGHT_DECAY
# Gradient clipping
_C.SOLVER.CLIP_GRADIENTS = CN({"ENABLED": False})
# Type of gradient clipping, currently 2 values are supported:
# - "value": the absolute values of elements of each gradients are clipped
# - "norm": the norm of the gradient for each parameter is clipped thus
# affecting all elements in the parameter
_C.SOLVER.CLIP_GRADIENTS.CLIP_TYPE = "value"
# Maximum absolute value used for clipping gradients
_C.SOLVER.CLIP_GRADIENTS.CLIP_VALUE = 1.0
# Floating point number p for L-p norm to be used with the "norm"
# gradient clipping type; for L-inf, please specify .inf
_C.SOLVER.CLIP_GRADIENTS.NORM_TYPE = 2.0
# Enable automatic mixed precision for training
# Note that this does not change model's inference behavior.
# To use AMP in inference, run inference under autocast()
_C.SOLVER.AMP = CN({"ENABLED": False})
# ---------------------------------------------------------------------------- #
# Specific test options
# ---------------------------------------------------------------------------- #
_C.TEST = CN()
# For end-to-end tests to verify the expected accuracy.
# Each item is [task, metric, value, tolerance]
# e.g.: [['bbox', 'AP', 38.5, 0.2]]
_C.TEST.EXPECTED_RESULTS = []
# The period (in terms of steps) to evaluate the model during training.
# Set to 0 to disable.
_C.TEST.EVAL_PERIOD = 0
# The sigmas used to calculate keypoint OKS. See http://cocodataset.org/#keypoints-eval
# When empty, it will use the defaults in COCO.
# Otherwise it should be a list[float] with the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS.
_C.TEST.KEYPOINT_OKS_SIGMAS = []
# Maximum number of detections to return per image during inference (100 is
# based on the limit established for the COCO dataset).
_C.TEST.DETECTIONS_PER_IMAGE = 100
_C.TEST.AUG = CN({"ENABLED": False})
_C.TEST.AUG.MIN_SIZES = (400, 500, 600, 700, 800, 900, 1000, 1100, 1200)
_C.TEST.AUG.MAX_SIZE = 4000
_C.TEST.AUG.FLIP = True
_C.TEST.PRECISE_BN = CN({"ENABLED": False})
_C.TEST.PRECISE_BN.NUM_ITER = 200
# ---------------------------------------------------------------------------- #
# Misc options
# ---------------------------------------------------------------------------- #
# Directory where output files are written
_C.OUTPUT_DIR = "./output"
# Set seed to negative to fully randomize everything.
# Set seed to positive to use a fixed seed. Note that a fixed seed increases
# reproducibility but does not guarantee fully deterministic behavior.
# Disabling all parallelism further increases reproducibility.
_C.SEED = -1
# Benchmark different cudnn algorithms.
# If input images have very different sizes, this option will have large overhead
# for about 10k iterations. It usually hurts total time, but can benefit for certain models.
# If input images have the same or similar sizes, benchmark is often helpful.
_C.CUDNN_BENCHMARK = False
# The period (in terms of steps) for minibatch visualization at train time.
# Set to 0 to disable.
_C.VIS_PERIOD = 0
# global config is for quick hack purposes.
# You can set them in command line or config files,
# and access it with:
#
# from detectron2.config import global_cfg
# print(global_cfg.HACK)
#
# Do not commit any configs into it.
_C.GLOBAL = CN()
_C.GLOBAL.HACK = 1.0
detectron2/config/instantiate.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import dataclasses
import logging
from collections import abc
from typing import Any
from detectron2.utils.registry import _convert_target_to_string, locate
__all__ = ["dump_dataclass", "instantiate"]
def dump_dataclass(obj: Any):
"""
Dump a dataclass recursively into a dict that can be later instantiated.
Args:
obj: a dataclass object
Returns:
dict
"""
assert dataclasses.is_dataclass(obj) and not isinstance(
obj, type
), "dump_dataclass() requires an instance of a dataclass."
ret = {"_target_": _convert_target_to_string(type(obj))}
for f in dataclasses.fields(obj):
v = getattr(obj, f.name)
if dataclasses.is_dataclass(v):
v = dump_dataclass(v)
if isinstance(v, (list, tuple)):
v = [dump_dataclass(x) if dataclasses.is_dataclass(x) else x for x in v]
ret[f.name] = v
return ret
def instantiate(cfg):
"""
Recursively instantiate objects defined in dictionaries by
"_target_" and arguments.
Args:
cfg: a dict-like object with "_target_" that defines the caller, and
other keys that define the arguments
Returns:
object instantiated by cfg
"""
from omegaconf import ListConfig
if isinstance(cfg, ListConfig):
lst = [instantiate(x) for x in cfg]
return ListConfig(lst, flags={"allow_objects": True})
if isinstance(cfg, list):
# Specialize for list, because many classes take
# list[objects] as arguments, such as ResNet, DatasetMapper
return [instantiate(x) for x in cfg]
if isinstance(cfg, abc.Mapping) and "_target_" in cfg:
# conceptually equivalent to hydra.utils.instantiate(cfg) with _convert_=all,
# but faster: https://github.com/facebookresearch/hydra/issues/1200
cfg = {k: instantiate(v) for k, v in cfg.items()}
cls = cfg.pop("_target_")
cls = instantiate(cls)
if isinstance(cls, str):
cls_name = cls
cls = locate(cls_name)
assert cls is not None, cls_name
else:
try:
cls_name = cls.__module__ + "." + cls.__qualname__
except Exception:
# target could be anything, so the above could fail
cls_name = str(cls)
assert callable(cls), f"_target_ {cls} does not define a callable object"
try:
return cls(**cfg)
except TypeError:
logger = logging.getLogger(__name__)
logger.error(f"Error when instantiating {cls_name}!")
raise
return cfg # return as-is if don't know what to do
detectron2/config/lazy.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import ast
import builtins
import importlib
import inspect
import logging
import os
import uuid
from collections import abc
from contextlib import contextmanager
from copy import deepcopy
from dataclasses import is_dataclass
from typing import List, Tuple, Union
import cloudpickle
import yaml
from omegaconf import DictConfig, ListConfig, OmegaConf
from detectron2.utils.file_io import PathManager
from detectron2.utils.registry import _convert_target_to_string
__all__ = ["LazyCall", "LazyConfig"]
class LazyCall:
"""
Wrap a callable so that when it's called, the call will not be executed,
but returns a dict that describes the call.
LazyCall object has to be called with only keyword arguments. Positional
arguments are not yet supported.
Examples:
::
from detectron2.config import instantiate, LazyCall
layer_cfg = LazyCall(nn.Conv2d)(in_channels=32, out_channels=32)
layer_cfg.out_channels = 64 # can edit it afterwards
layer = instantiate(layer_cfg)
"""
def __init__(self, target):
if not (callable(target) or isinstance(target, (str, abc.Mapping))):
raise TypeError(
f"target of LazyCall must be a callable or defines a callable! Got {target}"
)
self._target = target
def __call__(self, **kwargs):
if is_dataclass(self._target):
# omegaconf object cannot hold dataclass type
# https://github.com/omry/omegaconf/issues/784
target = _convert_target_to_string(self._target)
else:
target = self._target
kwargs["_target_"] = target
return DictConfig(content=kwargs, flags={"allow_objects": True})
def _visit_dict_config(cfg, func):
"""
Apply func recursively to all DictConfig in cfg.
"""
if isinstance(cfg, DictConfig):
func(cfg)
for v in cfg.values():
_visit_dict_config(v, func)
elif isinstance(cfg, ListConfig):
for v in cfg:
_visit_dict_config(v, func)
def _validate_py_syntax(filename):
# see also https://github.com/open-mmlab/mmcv/blob/master/mmcv/utils/config.py
with PathManager.open(filename, "r") as f:
content = f.read()
try:
ast.parse(content)
except SyntaxError as e:
raise SyntaxError(f"Config file {filename} has syntax error!") from e
def _cast_to_config(obj):
# if given a dict, return DictConfig instead
if isinstance(obj, dict):
return DictConfig(obj, flags={"allow_objects": True})
return obj
_CFG_PACKAGE_NAME = "detectron2._cfg_loader"
"""
A namespace to put all imported config into.
"""
def _random_package_name(filename):
# generate a random package name when loading config files
return _CFG_PACKAGE_NAME + str(uuid.uuid4())[:4] + "." + os.path.basename(filename)
@contextmanager
def _patch_import():
"""
Enhance relative import statements in config files, so that they:
1. locate files purely based on relative location, regardless of packages.
e.g. you can import file without having __init__
2. do not cache modules globally; modifications of module states has no side effect
3. support other storage system through PathManager
4. imported dict are turned into omegaconf.DictConfig automatically
"""
old_import = builtins.__import__
def find_relative_file(original_file, relative_import_path, level):
cur_file = os.path.dirname(original_file)
for _ in range(level - 1):
cur_file = os.path.dirname(cur_file)
cur_name = relative_import_path.lstrip(".")
for part in cur_name.split("."):
cur_file = os.path.join(cur_file, part)
# NOTE: directory import is not handled. Because then it's unclear
# if such import should produce python module or DictConfig. This can
# be discussed further if needed.
if not cur_file.endswith(".py"):
cur_file += ".py"
if not PathManager.isfile(cur_file):
raise ImportError(
f"Cannot import name {relative_import_path} from "
f"{original_file}: {cur_file} has to exist."
)
return cur_file
def new_import(name, globals=None, locals=None, fromlist=(), level=0):
if (
# Only deal with relative imports inside config files
level != 0
and globals is not None
and (globals.get("__package__", "") or "").startswith(_CFG_PACKAGE_NAME)
):
cur_file = find_relative_file(globals["__file__"], name, level)
_validate_py_syntax(cur_file)
spec = importlib.machinery.ModuleSpec(
_random_package_name(cur_file), None, origin=cur_file
)
module = importlib.util.module_from_spec(spec)
module.__file__ = cur_file
with PathManager.open(cur_file) as f:
content = f.read()
exec(compile(content, cur_file, "exec"), module.__dict__)
for name in fromlist: # turn imported dict into DictConfig automatically
val = _cast_to_config(module.__dict__[name])
module.__dict__[name] = val
return module
return old_import(name, globals, locals, fromlist=fromlist, level=level)
builtins.__import__ = new_import
yield new_import
builtins.__import__ = old_import
class LazyConfig:
"""
Provide methods to save, load, and overrides an omegaconf config object
which may contain definition of lazily-constructed objects.
"""
@staticmethod
def load_rel(filename: str, keys: Union[None, str, Tuple[str, ...]] = None):
"""
Similar to :meth:`load()`, but load path relative to the caller's
source file.
This has the same functionality as a relative import, except that this method
accepts filename as a string, so more characters are allowed in the filename.
"""
caller_frame = inspect.stack()[1]
caller_fname = caller_frame[0].f_code.co_filename
assert caller_fname != "<string>", "load_rel Unable to find caller"
caller_dir = os.path.dirname(caller_fname)
filename = os.path.join(caller_dir, filename)
return LazyConfig.load(filename, keys)
@staticmethod
def load(filename: str, keys: Union[None, str, Tuple[str, ...]] = None):
"""
Load a config file.
Args:
filename: absolute path or relative path w.r.t. the current working directory
keys: keys to load and return. If not given, return all keys
(whose values are config objects) in a dict.
"""
has_keys = keys is not None
filename = filename.replace("/./", "/") # redundant
if os.path.splitext(filename)[1] not in [".py", ".yaml", ".yml"]:
raise ValueError(f"Config file {filename} has to be a python or yaml file.")
if filename.endswith(".py"):
_validate_py_syntax(filename)
with _patch_import():
# Record the filename
module_namespace = {
"__file__": filename,
"__package__": _random_package_name(filename),
}
with PathManager.open(filename) as f:
content = f.read()
# Compile first with filename to:
# 1. make filename appears in stacktrace
# 2. make load_rel able to find its parent's (possibly remote) location
exec(compile(content, filename, "exec"), module_namespace)
ret = module_namespace
else:
with PathManager.open(filename) as f:
obj = yaml.unsafe_load(f)
ret = OmegaConf.create(obj, flags={"allow_objects": True})
if has_keys:
if isinstance(keys, str):
return _cast_to_config(ret[keys])
else:
return tuple(_cast_to_config(ret[a]) for a in keys)
else:
if filename.endswith(".py"):
# when not specified, only load those that are config objects
ret = DictConfig(
{
name: _cast_to_config(value)
for name, value in ret.items()
if isinstance(value, (DictConfig, ListConfig, dict))
and not name.startswith("_")
},
flags={"allow_objects": True},
)
return ret
@staticmethod
def save(cfg, filename: str):
"""
Save a config object to a yaml file.
Note that when the config dictionary contains complex objects (e.g. lambda),
it can't be saved to yaml. In that case we will print an error and
attempt to save to a pkl file instead.
Args:
cfg: an omegaconf config object
filename: yaml file name to save the config file
"""
logger = logging.getLogger(__name__)
try:
cfg = deepcopy(cfg)
except Exception:
pass
else:
# if it's deep-copyable, then...
def _replace_type_by_name(x):
if "_target_" in x and callable(x._target_):
try:
x._target_ = _convert_target_to_string(x._target_)
except AttributeError:
pass
# not necessary, but makes yaml looks nicer
_visit_dict_config(cfg, _replace_type_by_name)
save_pkl = False
try:
dict = OmegaConf.to_container(cfg, resolve=False)
dumped = yaml.dump(dict, default_flow_style=None, allow_unicode=True, width=9999)
with PathManager.open(filename, "w") as f:
f.write(dumped)
try:
_ = yaml.unsafe_load(dumped) # test that it is loadable
except Exception:
logger.warning(
"The config contains objects that cannot serialize to a valid yaml. "
f"{filename} is human-readable but cannot be loaded."
)
save_pkl = True
except Exception:
logger.exception("Unable to serialize the config to yaml. Error:")
save_pkl = True
if save_pkl:
new_filename = filename + ".pkl"
try:
# retry by pickle
with PathManager.open(new_filename, "wb") as f:
cloudpickle.dump(cfg, f)
logger.warning(f"Config is saved using cloudpickle at {new_filename}.")
except Exception:
pass
@staticmethod
def apply_overrides(cfg, overrides: List[str]):
"""
In-place override contents of cfg.
Args:
cfg: an omegaconf config object
overrides: list of strings in the format of "a=b" to override configs.
See https://hydra.cc/docs/next/advanced/override_grammar/basic/
for syntax.
Returns:
the cfg object
"""
def safe_update(cfg, key, value):
parts = key.split(".")
for idx in range(1, len(parts)):
prefix = ".".join(parts[:idx])
v = OmegaConf.select(cfg, prefix, default=None)
if v is None:
break
if not OmegaConf.is_config(v):
raise KeyError(
f"Trying to update key {key}, but {prefix} "
f"is not a config, but has type {type(v)}."
)
OmegaConf.update(cfg, key, value, merge=True)
from hydra.core.override_parser.overrides_parser import OverridesParser
parser = OverridesParser.create()
overrides = parser.parse_overrides(overrides)
for o in overrides:
key = o.key_or_group
value = o.value()
if o.is_delete():
# TODO support this
raise NotImplementedError("deletion is not yet a supported override")
safe_update(cfg, key, value)
return cfg
@staticmethod
def to_py(cfg, prefix: str = "cfg."):
"""
Try to convert a config object into Python-like psuedo code.
Note that perfect conversion is not always possible. So the returned
results are mainly meant to be human-readable, and not meant to be executed.
Args:
cfg: an omegaconf config object
prefix: root name for the resulting code (default: "cfg.")
Returns:
str of formatted Python code
"""
import black
cfg = OmegaConf.to_container(cfg, resolve=True)
def _to_str(obj, prefix=None, inside_call=False):
if prefix is None:
prefix = []
if isinstance(obj, abc.Mapping) and "_target_" in obj:
# Dict representing a function call
target = _convert_target_to_string(obj.pop("_target_"))
args = []
for k, v in sorted(obj.items()):
args.append(f"{k}={_to_str(v, inside_call=True)}")
args = ", ".join(args)
call = f"{target}({args})"
return "".join(prefix) + call
elif isinstance(obj, abc.Mapping) and not inside_call:
# Dict that is not inside a call is a list of top-level config objects that we
# render as one object per line with dot separated prefixes
key_list = []
for k, v in sorted(obj.items()):
if isinstance(v, abc.Mapping) and "_target_" not in v:
key_list.append(_to_str(v, prefix=prefix + [k + "."]))
else:
key = "".join(prefix) + k
key_list.append(f"{key}={_to_str(v)}")
return "\n".join(key_list)
elif isinstance(obj, abc.Mapping):
# Dict that is inside a call is rendered as a regular dict
return (
"{"
+ ",".join(
f"{repr(k)}: {_to_str(v, inside_call=inside_call)}"
for k, v in sorted(obj.items())
)
+ "}"
)
elif isinstance(obj, list):
return "[" + ",".join(_to_str(x, inside_call=inside_call) for x in obj) + "]"
else:
return repr(obj)
py_str = _to_str(cfg, prefix=[prefix])
try:
return black.format_str(py_str, mode=black.Mode())
except black.InvalidInput:
return py_str
detectron2/data/__init__.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
from . import transforms # isort:skip
from .build import (
build_batch_data_loader,
build_detection_test_loader,
build_detection_train_loader,
get_detection_dataset_dicts,
load_proposals_into_dataset,
print_instances_class_histogram,
)
from .catalog import DatasetCatalog, MetadataCatalog, Metadata
from .common import DatasetFromList, MapDataset, ToIterableDataset
from .dataset_mapper import DatasetMapper
# ensure the builtin datasets are registered
from . import datasets, samplers # isort:skip
__all__ = [k for k in globals().keys() if not k.startswith("_")]
detectron2/data/benchmark.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import logging
import numpy as np
from itertools import count
from typing import List, Tuple
import torch
import tqdm
from fvcore.common.timer import Timer
from detectron2.utils import comm
from .build import build_batch_data_loader
from .common import DatasetFromList, MapDataset
from .samplers import TrainingSampler
logger = logging.getLogger(__name__)
class _EmptyMapDataset(torch.utils.data.Dataset):
"""
Map anything to emptiness.
"""
def __init__(self, dataset):
self.ds = dataset
def __len__(self):
return len(self.ds)
def __getitem__(self, idx):
_ = self.ds[idx]
return [0]
def iter_benchmark(
iterator, num_iter: int, warmup: int = 5, max_time_seconds: float = 60
) -> Tuple[float, List[float]]:
"""
Benchmark an iterator/iterable for `num_iter` iterations with an extra
`warmup` iterations of warmup.
End early if `max_time_seconds` time is spent on iterations.
Returns:
float: average time (seconds) per iteration
list[float]: time spent on each iteration. Sometimes useful for further analysis.
"""
num_iter, warmup = int(num_iter), int(warmup)
iterator = iter(iterator)
for _ in range(warmup):
next(iterator)
timer = Timer()
all_times = []
for curr_iter in tqdm.trange(num_iter):
start = timer.seconds()
if start > max_time_seconds:
num_iter = curr_iter
break
next(iterator)
all_times.append(timer.seconds() - start)
avg = timer.seconds() / num_iter
return avg, all_times
class DataLoaderBenchmark:
"""
Some common benchmarks that help understand perf bottleneck of a standard dataloader
made of dataset, mapper and sampler.
"""
def __init__(
self,
dataset,
*,
mapper,
sampler=None,
total_batch_size,
num_workers=0,
max_time_seconds: int = 90,
):
"""
Args:
max_time_seconds (int): maximum time to spent for each benchmark
other args: same as in `build.py:build_detection_train_loader`
"""
if isinstance(dataset, list):
dataset = DatasetFromList(dataset, copy=False, serialize=True)
if sampler is None:
sampler = TrainingSampler(len(dataset))
self.dataset = dataset
self.mapper = mapper
self.sampler = sampler
self.total_batch_size = total_batch_size
self.num_workers = num_workers
self.per_gpu_batch_size = self.total_batch_size // comm.get_world_size()
self.max_time_seconds = max_time_seconds
def _benchmark(self, iterator, num_iter, warmup, msg=None):
avg, all_times = iter_benchmark(iterator, num_iter, warmup, self.max_time_seconds)
if msg is not None:
self._log_time(msg, avg, all_times)
return avg, all_times
def _log_time(self, msg, avg, all_times, distributed=False):
percentiles = [np.percentile(all_times, k, interpolation="nearest") for k in [1, 5, 95, 99]]
if not distributed:
logger.info(
f"{msg}: avg={1.0/avg:.1f} it/s, "
f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "
f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."
)
return
avg_per_gpu = comm.all_gather(avg)
percentiles_per_gpu = comm.all_gather(percentiles)
if comm.get_rank() > 0:
return
for idx, avg, percentiles in zip(count(), avg_per_gpu, percentiles_per_gpu):
logger.info(
f"GPU{idx} {msg}: avg={1.0/avg:.1f} it/s, "
f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "
f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."
)
def benchmark_dataset(self, num_iter, warmup=5):
"""
Benchmark the speed of taking raw samples from the dataset.
"""
def loader():
while True:
for k in self.sampler:
yield self.dataset[k]
self._benchmark(loader(), num_iter, warmup, "Dataset Alone")
def benchmark_mapper(self, num_iter, warmup=5):
"""
Benchmark the speed of taking raw samples from the dataset and map
them in a single process.
"""
def loader():
while True:
for k in self.sampler:
yield self.mapper(self.dataset[k])
self._benchmark(loader(), num_iter, warmup, "Single Process Mapper (sec/sample)")
def benchmark_workers(self, num_iter, warmup=10):
"""
Benchmark the dataloader by tuning num_workers to [0, 1, self.num_workers].
"""
candidates = [0, 1]
if self.num_workers not in candidates:
candidates.append(self.num_workers)
dataset = MapDataset(self.dataset, self.mapper)
for n in candidates:
loader = build_batch_data_loader(
dataset,
self.sampler,
self.total_batch_size,
num_workers=n,
)
self._benchmark(
iter(loader),
num_iter * max(n, 1),
warmup * max(n, 1),
f"DataLoader ({n} workers, bs={self.per_gpu_batch_size})",
)
del loader
def benchmark_IPC(self, num_iter, warmup=10):
"""
Benchmark the dataloader where each worker outputs nothing. This
eliminates the IPC overhead compared to the regular dataloader.
PyTorch multiprocessing's IPC only optimizes for torch tensors.
Large numpy arrays or other data structure may incur large IPC overhead.
"""
n = self.num_workers
dataset = _EmptyMapDataset(MapDataset(self.dataset, self.mapper))
loader = build_batch_data_loader(
dataset, self.sampler, self.total_batch_size, num_workers=n
)
self._benchmark(
iter(loader),
num_iter * max(n, 1),
warmup * max(n, 1),
f"DataLoader ({n} workers, bs={self.per_gpu_batch_size}) w/o comm",
)
def benchmark_distributed(self, num_iter, warmup=10):
"""
Benchmark the dataloader in each distributed worker, and log results of
all workers. This helps understand the final performance as well as
the variances among workers.
It also prints startup time (first iter) of the dataloader.
"""
gpu = comm.get_world_size()
dataset = MapDataset(self.dataset, self.mapper)
n = self.num_workers
loader = build_batch_data_loader(
dataset, self.sampler, self.total_batch_size, num_workers=n
)
timer = Timer()
loader = iter(loader)
next(loader)
startup_time = timer.seconds()
logger.info("Dataloader startup time: {:.2f} seconds".format(startup_time))
comm.synchronize()
avg, all_times = self._benchmark(loader, num_iter * max(n, 1), warmup * max(n, 1))
del loader
self._log_time(
f"DataLoader ({gpu} GPUs x {n} workers, total bs={self.total_batch_size})",
avg,
all_times,
True,
)
detectron2/data/build.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import itertools
import logging
import numpy as np
import operator
import pickle
import torch
import torch.utils.data as torchdata
from tabulate import tabulate
from termcolor import colored
from detectron2.config import configurable
from detectron2.structures import BoxMode
from detectron2.utils.comm import get_world_size
from detectron2.utils.env import seed_all_rng
from detectron2.utils.file_io import PathManager
from detectron2.utils.logger import _log_api_usage, log_first_n
from .catalog import DatasetCatalog, MetadataCatalog
from .common import AspectRatioGroupedDataset, DatasetFromList, MapDataset, ToIterableDataset
from .dataset_mapper import DatasetMapper
from .detection_utils import check_metadata_consistency
from .samplers import (
InferenceSampler,
RandomSubsetTrainingSampler,
RepeatFactorTrainingSampler,
TrainingSampler,
)
"""
This file contains the default logic to build a dataloader for training or testing.
"""
__all__ = [
"build_batch_data_loader",
"build_detection_train_loader",
"build_detection_test_loader",
"get_detection_dataset_dicts",
"load_proposals_into_dataset",
"print_instances_class_histogram",
]
def filter_images_with_only_crowd_annotations(dataset_dicts):
"""
Filter out images with none annotations or only crowd annotations
(i.e., images without non-crowd annotations).
A common training-time preprocessing on COCO dataset.
Args:
dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
Returns:
list[dict]: the same format, but filtered.
"""
num_before = len(dataset_dicts)
def valid(anns):
for ann in anns:
if ann.get("iscrowd", 0) == 0:
return True
return False
dataset_dicts = [x for x in dataset_dicts if valid(x["annotations"])]
num_after = len(dataset_dicts)
logger = logging.getLogger(__name__)
logger.info(
"Removed {} images with no usable annotations. {} images left.".format(
num_before - num_after, num_after
)
)
return dataset_dicts
def filter_images_with_few_keypoints(dataset_dicts, min_keypoints_per_image):
"""
Filter out images with too few number of keypoints.
Args:
dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
Returns:
list[dict]: the same format as dataset_dicts, but filtered.
"""
num_before = len(dataset_dicts)
def visible_keypoints_in_image(dic):
# Each keypoints field has the format [x1, y1, v1, ...], where v is visibility
annotations = dic["annotations"]
return sum(
(np.array(ann["keypoints"][2::3]) > 0).sum()
for ann in annotations
if "keypoints" in ann
)
dataset_dicts = [
x for x in dataset_dicts if visible_keypoints_in_image(x) >= min_keypoints_per_image
]
num_after = len(dataset_dicts)
logger = logging.getLogger(__name__)
logger.info(
"Removed {} images with fewer than {} keypoints.".format(
num_before - num_after, min_keypoints_per_image
)
)
return dataset_dicts
def load_proposals_into_dataset(dataset_dicts, proposal_file):
"""
Load precomputed object proposals into the dataset.
The proposal file should be a pickled dict with the following keys:
- "ids": list[int] or list[str], the image ids
- "boxes": list[np.ndarray], each is an Nx4 array of boxes corresponding to the image id
- "objectness_logits": list[np.ndarray], each is an N sized array of objectness scores
corresponding to the boxes.
- "bbox_mode": the BoxMode of the boxes array. Defaults to ``BoxMode.XYXY_ABS``.
Args:
dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
proposal_file (str): file path of pre-computed proposals, in pkl format.
Returns:
list[dict]: the same format as dataset_dicts, but added proposal field.
"""
logger = logging.getLogger(__name__)
logger.info("Loading proposals from: {}".format(proposal_file))
with PathManager.open(proposal_file, "rb") as f:
proposals = pickle.load(f, encoding="latin1")
# Rename the key names in D1 proposal files
rename_keys = {"indexes": "ids", "scores": "objectness_logits"}
for key in rename_keys:
if key in proposals:
proposals[rename_keys[key]] = proposals.pop(key)
# Fetch the indexes of all proposals that are in the dataset
# Convert image_id to str since they could be int.
img_ids = set({str(record["image_id"]) for record in dataset_dicts})
id_to_index = {str(id): i for i, id in enumerate(proposals["ids"]) if str(id) in img_ids}
# Assuming default bbox_mode of precomputed proposals are 'XYXY_ABS'
bbox_mode = BoxMode(proposals["bbox_mode"]) if "bbox_mode" in proposals else BoxMode.XYXY_ABS
for record in dataset_dicts:
# Get the index of the proposal
i = id_to_index[str(record["image_id"])]
boxes = proposals["boxes"][i]
objectness_logits = proposals["objectness_logits"][i]
# Sort the proposals in descending order of the scores
inds = objectness_logits.argsort()[::-1]
record["proposal_boxes"] = boxes[inds]
record["proposal_objectness_logits"] = objectness_logits[inds]
record["proposal_bbox_mode"] = bbox_mode
return dataset_dicts
def print_instances_class_histogram(dataset_dicts, class_names):
"""
Args:
dataset_dicts (list[dict]): list of dataset dicts.
class_names (list[str]): list of class names (zero-indexed).
"""
num_classes = len(class_names)
hist_bins = np.arange(num_classes + 1)
histogram = np.zeros((num_classes,), dtype=np.int)
for entry in dataset_dicts:
annos = entry["annotations"]
classes = np.asarray(
[x["category_id"] for x in annos if not x.get("iscrowd", 0)], dtype=np.int
)
if len(classes):
assert classes.min() >= 0, f"Got an invalid category_id={classes.min()}"
assert (
classes.max() < num_classes
), f"Got an invalid category_id={classes.max()} for a dataset of {num_classes} classes"
histogram += np.histogram(classes, bins=hist_bins)[0]
N_COLS = min(6, len(class_names) * 2)
def short_name(x):
# make long class names shorter. useful for lvis
if len(x) > 13:
return x[:11] + ".."
return x
data = list(
itertools.chain(*[[short_name(class_names[i]), int(v)] for i, v in enumerate(histogram)])
)
total_num_instances = sum(data[1::2])
data.extend([None] * (N_COLS - (len(data) % N_COLS)))
if num_classes > 1:
data.extend(["total", total_num_instances])
data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
table = tabulate(
data,
headers=["category", "#instances"] * (N_COLS // 2),
tablefmt="pipe",
numalign="left",
stralign="center",
)
log_first_n(
logging.INFO,
"Distribution of instances among all {} categories:\n".format(num_classes)
+ colored(table, "cyan"),
key="message",
)
def get_detection_dataset_dicts(
names,
filter_empty=True,
min_keypoints=0,
proposal_files=None,
check_consistency=True,
):
"""
Load and prepare dataset dicts for instance detection/segmentation and semantic segmentation.
Args:
names (str or list[str]): a dataset name or a list of dataset names
filter_empty (bool): whether to filter out images without instance annotations
min_keypoints (int): filter out images with fewer keypoints than
`min_keypoints`. Set to 0 to do nothing.
proposal_files (list[str]): if given, a list of object proposal files
that match each dataset in `names`.
check_consistency (bool): whether to check if datasets have consistent metadata.
Returns:
list[dict]: a list of dicts following the standard dataset dict format.
"""
if isinstance(names, str):
names = [names]
assert len(names), names
dataset_dicts = [DatasetCatalog.get(dataset_name) for dataset_name in names]
for dataset_name, dicts in zip(names, dataset_dicts):
assert len(dicts), "Dataset '{}' is empty!".format(dataset_name)
if proposal_files is not None:
assert len(names) == len(proposal_files)
# load precomputed proposals from proposal files
dataset_dicts = [
load_proposals_into_dataset(dataset_i_dicts, proposal_file)
for dataset_i_dicts, proposal_file in zip(dataset_dicts, proposal_files)
]
if isinstance(dataset_dicts[0], torchdata.Dataset):
return torchdata.ConcatDataset(dataset_dicts)
dataset_dicts = list(itertools.chain.from_iterable(dataset_dicts))
has_instances = "annotations" in dataset_dicts[0]
if filter_empty and has_instances:
dataset_dicts = filter_images_with_only_crowd_annotations(dataset_dicts)
if min_keypoints > 0 and has_instances:
dataset_dicts = filter_images_with_few_keypoints(dataset_dicts, min_keypoints)
if check_consistency and has_instances:
try:
class_names = MetadataCatalog.get(names[0]).thing_classes
check_metadata_consistency("thing_classes", names)
print_instances_class_histogram(dataset_dicts, class_names)
except AttributeError: # class names are not available for this dataset
pass
assert len(dataset_dicts), "No valid data found in {}.".format(",".join(names))
return dataset_dicts
def build_batch_data_loader(
dataset,
sampler,
total_batch_size,
*,
aspect_ratio_grouping=False,
num_workers=0,
collate_fn=None,
):
"""
Build a batched dataloader. The main differences from `torch.utils.data.DataLoader` are:
1. support aspect ratio grouping options
2. use no "batch collation", because this is common for detection training
Args:
dataset (torch.utils.data.Dataset): a pytorch map-style or iterable dataset.
sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces indices.
Must be provided iff. ``dataset`` is a map-style dataset.
total_batch_size, aspect_ratio_grouping, num_workers, collate_fn: see
:func:`build_detection_train_loader`.
Returns:
iterable[list]. Length of each list is the batch size of the current
GPU. Each element in the list comes from the dataset.
"""
world_size = get_world_size()
assert (
total_batch_size > 0 and total_batch_size % world_size == 0
), "Total batch size ({}) must be divisible by the number of gpus ({}).".format(
total_batch_size, world_size
)
batch_size = total_batch_size // world_size
if isinstance(dataset, torchdata.IterableDataset):
assert sampler is None, "sampler must be None if dataset is IterableDataset"
else:
dataset = ToIterableDataset(dataset, sampler)
if aspect_ratio_grouping:
data_loader = torchdata.DataLoader(
dataset,
num_workers=num_workers,
collate_fn=operator.itemgetter(0), # don't batch, but yield individual elements
worker_init_fn=worker_init_reset_seed,
) # yield individual mapped dict
data_loader = AspectRatioGroupedDataset(data_loader, batch_size)
if collate_fn is None:
return data_loader
return MapDataset(data_loader, collate_fn)
else:
return torchdata.DataLoader(
dataset,
batch_size=batch_size,
drop_last=True,
num_workers=num_workers,
collate_fn=trivial_batch_collator if collate_fn is None else collate_fn,
worker_init_fn=worker_init_reset_seed,
)
def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
if dataset is None:
dataset = get_detection_dataset_dicts(
cfg.DATASETS.TRAIN,
filter_empty=cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS,
min_keypoints=cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE
if cfg.MODEL.KEYPOINT_ON
else 0,
proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None,
)
_log_api_usage("dataset." + cfg.DATASETS.TRAIN[0])
if mapper is None:
mapper = DatasetMapper(cfg, True)
if sampler is None:
sampler_name = cfg.DATALOADER.SAMPLER_TRAIN
logger = logging.getLogger(__name__)
logger.info("Using training sampler {}".format(sampler_name))
if sampler_name == "TrainingSampler":
sampler = TrainingSampler(len(dataset))
elif sampler_name == "RepeatFactorTrainingSampler":
repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
dataset, cfg.DATALOADER.REPEAT_THRESHOLD
)
sampler = RepeatFactorTrainingSampler(repeat_factors)
elif sampler_name == "RandomSubsetTrainingSampler":
sampler = RandomSubsetTrainingSampler(len(dataset), cfg.DATALOADER.RANDOM_SUBSET_RATIO)
else:
raise ValueError("Unknown training sampler: {}".format(sampler_name))
return {
"dataset": dataset,
"sampler": sampler,
"mapper": mapper,
"total_batch_size": cfg.SOLVER.IMS_PER_BATCH,
"aspect_ratio_grouping": cfg.DATALOADER.ASPECT_RATIO_GROUPING,
"num_workers": cfg.DATALOADER.NUM_WORKERS,
}
@configurable(from_config=_train_loader_from_config)
def build_detection_train_loader(
dataset,
*,
mapper,
sampler=None,
total_batch_size,
aspect_ratio_grouping=True,
num_workers=0,
collate_fn=None,
):
"""
Build a dataloader for object detection with some default features.
This interface is experimental.
Args:
dataset (list or torch.utils.data.Dataset): a list of dataset dicts,
or a pytorch dataset (either map-style or iterable). It can be obtained
by using :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`.
mapper (callable): a callable which takes a sample (dict) from dataset and
returns the format to be consumed by the model.
When using cfg, the default choice is ``DatasetMapper(cfg, is_train=True)``.
sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces
indices to be applied on ``dataset``.
If ``dataset`` is map-style, the default sampler is a :class:`TrainingSampler`,
which coordinates an infinite random shuffle sequence across all workers.
Sampler must be None if ``dataset`` is iterable.
total_batch_size (int): total batch size across all workers. Batching
simply puts data into a list.
aspect_ratio_grouping (bool): whether to group images with similar
aspect ratio for efficiency. When enabled, it requires each
element in dataset be a dict with keys "width" and "height".
num_workers (int): number of parallel data loading workers
collate_fn: same as the argument of `torch.utils.data.DataLoader`.
Defaults to do no collation and return a list of data.
No collation is OK for small batch size and simple data structures.
If your batch size is large and each sample contains too many small tensors,
it's more efficient to collate them in data loader.
Returns:
torch.utils.data.DataLoader:
a dataloader. Each output from it is a ``list[mapped_element]`` of length
``total_batch_size / num_workers``, where ``mapped_element`` is produced
by the ``mapper``.
"""
if isinstance(dataset, list):
dataset = DatasetFromList(dataset, copy=False)
if mapper is not None:
dataset = MapDataset(dataset, mapper)
if isinstance(dataset, torchdata.IterableDataset):
assert sampler is None, "sampler must be None if dataset is IterableDataset"
else:
if sampler is None:
sampler = TrainingSampler(len(dataset))
assert isinstance(sampler, torchdata.Sampler), f"Expect a Sampler but got {type(sampler)}"
return build_batch_data_loader(
dataset,
sampler,
total_batch_size,
aspect_ratio_grouping=aspect_ratio_grouping,
num_workers=num_workers,
collate_fn=collate_fn,
)
def _test_loader_from_config(cfg, dataset_name, mapper=None):
"""
Uses the given `dataset_name` argument (instead of the names in cfg), because the
standard practice is to evaluate each test set individually (not combining them).
"""
if isinstance(dataset_name, str):
dataset_name = [dataset_name]
dataset = get_detection_dataset_dicts(
dataset_name,
filter_empty=False,
proposal_files=[
cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(x)] for x in dataset_name
]
if cfg.MODEL.LOAD_PROPOSALS
else None,
)
if mapper is None:
mapper = DatasetMapper(cfg, False)
return {"dataset": dataset, "mapper": mapper, "num_workers": cfg.DATALOADER.NUM_WORKERS}
@configurable(from_config=_test_loader_from_config)
def build_detection_test_loader(dataset, *, mapper, sampler=None, num_workers=0, collate_fn=None):
"""
Similar to `build_detection_train_loader`, but uses a batch size of 1,
and :class:`InferenceSampler`. This sampler coordinates all workers to
produce the exact set of all samples.
This interface is experimental.
Args:
dataset (list or torch.utils.data.Dataset): a list of dataset dicts,
or a pytorch dataset (either map-style or iterable). They can be obtained
by using :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`.
mapper (callable): a callable which takes a sample (dict) from dataset
and returns the format to be consumed by the model.
When using cfg, the default choice is ``DatasetMapper(cfg, is_train=False)``.
sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces
indices to be applied on ``dataset``. Default to :class:`InferenceSampler`,
which splits the dataset across all workers. Sampler must be None
if `dataset` is iterable.
num_workers (int): number of parallel data loading workers
collate_fn: same as the argument of `torch.utils.data.DataLoader`.
Defaults to do no collation and return a list of data.
Returns:
DataLoader: a torch DataLoader, that loads the given detection
dataset, with test-time transformation and batching.
Examples:
::
data_loader = build_detection_test_loader(
DatasetRegistry.get("my_test"),
mapper=DatasetMapper(...))
# or, instantiate with a CfgNode:
data_loader = build_detection_test_loader(cfg, "my_test")
"""
if isinstance(dataset, list):
dataset = DatasetFromList(dataset, copy=False)
if mapper is not None:
dataset = MapDataset(dataset, mapper)
if isinstance(dataset, torchdata.IterableDataset):
assert sampler is None, "sampler must be None if dataset is IterableDataset"
else:
if sampler is None:
sampler = InferenceSampler(len(dataset))
# Always use 1 image per worker during inference since this is the
# standard when reporting inference time in papers.
return torchdata.DataLoader(
dataset,
batch_size=1,
sampler=sampler,
num_workers=num_workers,
collate_fn=trivial_batch_collator if collate_fn is None else collate_fn,
)
def trivial_batch_collator(batch):
"""
A batch collator that does nothing.
"""
return batch
def worker_init_reset_seed(worker_id):
initial_seed = torch.initial_seed() % 2 ** 31
seed_all_rng(initial_seed + worker_id)
detectron2/data/catalog.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import copy
import logging
import types
from collections import UserDict
from typing import List
from detectron2.utils.logger import log_first_n
__all__ = ["DatasetCatalog", "MetadataCatalog", "Metadata"]
class _DatasetCatalog(UserDict):
"""
A global dictionary that stores information about the datasets and how to obtain them.
It contains a mapping from strings
(which are names that identify a dataset, e.g. "coco_2014_train")
to a function which parses the dataset and returns the samples in the
format of `list[dict]`.
The returned dicts should be in Detectron2 Dataset format (See DATASETS.md for details)
if used with the data loader functionalities in `data/build.py,data/detection_transform.py`.
The purpose of having this catalog is to make it easy to choose
different datasets, by just using the strings in the config.
"""
def register(self, name, func):
"""
Args:
name (str): the name that identifies a dataset, e.g. "coco_2014_train".
func (callable): a callable which takes no arguments and returns a list of dicts.
It must return the same results if called multiple times.
"""
assert callable(func), "You must register a function with `DatasetCatalog.register`!"
assert name not in self, "Dataset '{}' is already registered!".format(name)
self[name] = func
def get(self, name):
"""
Call the registered function and return its results.
Args:
name (str): the name that identifies a dataset, e.g. "coco_2014_train".
Returns:
list[dict]: dataset annotations.
"""
try:
f = self[name]
except KeyError as e:
raise KeyError(
"Dataset '{}' is not registered! Available datasets are: {}".format(
name, ", ".join(list(self.keys()))
)
) from e
return f()
def list(self) -> List[str]:
"""
List all registered datasets.
Returns:
list[str]
"""
return list(self.keys())
def remove(self, name):
"""
Alias of ``pop``.
"""
self.pop(name)
def __str__(self):
return "DatasetCatalog(registered datasets: {})".format(", ".join(self.keys()))
__repr__ = __str__
DatasetCatalog = _DatasetCatalog()
DatasetCatalog.__doc__ = (
_DatasetCatalog.__doc__
+ """
.. automethod:: detectron2.data.catalog.DatasetCatalog.register
.. automethod:: detectron2.data.catalog.DatasetCatalog.get
"""
)
class Metadata(types.SimpleNamespace):
"""
A class that supports simple attribute setter/getter.
It is intended for storing metadata of a dataset and make it accessible globally.
Examples:
::
# somewhere when you load the data:
MetadataCatalog.get("mydataset").thing_classes = ["person", "dog"]
# somewhere when you print statistics or visualize:
classes = MetadataCatalog.get("mydataset").thing_classes
"""
# the name of the dataset
# set default to N/A so that `self.name` in the errors will not trigger getattr again
name: str = "N/A"
_RENAMED = {
"class_names": "thing_classes",
"dataset_id_to_contiguous_id": "thing_dataset_id_to_contiguous_id",
"stuff_class_names": "stuff_classes",
}
def __getattr__(self, key):
if key in self._RENAMED:
log_first_n(
logging.WARNING,
"Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
n=10,
)
return getattr(self, self._RENAMED[key])
# "name" exists in every metadata
if len(self.__dict__) > 1:
raise AttributeError(
"Attribute '{}' does not exist in the metadata of dataset '{}'. Available "
"keys are {}.".format(key, self.name, str(self.__dict__.keys()))
)
else:
raise AttributeError(
f"Attribute '{key}' does not exist in the metadata of dataset '{self.name}': "
"metadata is empty."
)
def __setattr__(self, key, val):
if key in self._RENAMED:
log_first_n(
logging.WARNING,
"Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
n=10,
)
setattr(self, self._RENAMED[key], val)
# Ensure that metadata of the same name stays consistent
try:
oldval = getattr(self, key)
assert oldval == val, (
"Attribute '{}' in the metadata of '{}' cannot be set "
"to a different value!\n{} != {}".format(key, self.name, oldval, val)
)
except AttributeError:
super().__setattr__(key, val)
def as_dict(self):
"""
Returns all the metadata as a dict.
Note that modifications to the returned dict will not reflect on the Metadata object.
"""
return copy.copy(self.__dict__)
def set(self, **kwargs):
"""
Set multiple metadata with kwargs.
"""
for k, v in kwargs.items():
setattr(self, k, v)
return self
def get(self, key, default=None):
"""
Access an attribute and return its value if exists.
Otherwise return default.
"""
try:
return getattr(self, key)
except AttributeError:
return default
class _MetadataCatalog(UserDict):
"""
MetadataCatalog is a global dictionary that provides access to
:class:`Metadata` of a given dataset.
The metadata associated with a certain name is a singleton: once created, the
metadata will stay alive and will be returned by future calls to ``get(name)``.
It's like global variables, so don't abuse it.
It's meant for storing knowledge that's constant and shared across the execution
of the program, e.g.: the class names in COCO.
"""
def get(self, name):
"""
Args:
name (str): name of a dataset (e.g. coco_2014_train).
Returns:
Metadata: The :class:`Metadata` instance associated with this name,
or create an empty one if none is available.
"""
assert len(name)
r = super().get(name, None)
if r is None:
r = self[name] = Metadata(name=name)
return r
def list(self):
"""
List all registered metadata.
Returns:
list[str]: keys (names of datasets) of all registered metadata
"""
return list(self.keys())
def remove(self, name):
"""
Alias of ``pop``.
"""
self.pop(name)
def __str__(self):
return "MetadataCatalog(registered metadata: {})".format(", ".join(self.keys()))
__repr__ = __str__
MetadataCatalog = _MetadataCatalog()
MetadataCatalog.__doc__ = (
_MetadataCatalog.__doc__
+ """
.. automethod:: detectron2.data.catalog.MetadataCatalog.get
"""
)
detectron2/data/common.py 0 → 100644
View file @ 63e10e00
# Copyright (c) Facebook, Inc. and its affiliates.
import copy
import itertools
import logging
import numpy as np
import pickle
import random
import torch.utils.data as data
from torch.utils.data.sampler import Sampler
from detectron2.utils.serialize import PicklableWrapper
__all__ = ["MapDataset", "DatasetFromList", "AspectRatioGroupedDataset", "ToIterableDataset"]
def _shard_iterator_dataloader_worker(iterable):
# Shard the iterable if we're currently inside pytorch dataloader worker.
worker_info = data.get_worker_info()
if worker_info is None or worker_info.num_workers == 1:
# do nothing
yield from iterable
else:
yield from itertools.islice(iterable, worker_info.id, None, worker_info.num_workers)
class _MapIterableDataset(data.IterableDataset):
"""
Map a function over elements in an IterableDataset.
Similar to pytorch's MapIterDataPipe, but support filtering when map_func
returns None.
This class is not public-facing. Will be called by `MapDataset`.
"""
def __init__(self, dataset, map_func):
self._dataset = dataset
self._map_func = PicklableWrapper(map_func) # wrap so that a lambda will work
def __len__(self):
return len(self._dataset)
def __iter__(self):
for x in map(self._map_func, self._dataset):
if x is not None:
yield x
class MapDataset(data.Dataset):
"""
Map a function over the elements in a dataset.
"""
def __init__(self, dataset, map_func):
"""
Args:
dataset: a dataset where map function is applied. Can be either
map-style or iterable dataset. When given an iterable dataset,
the returned object will also be an iterable dataset.
map_func: a callable which maps the element in dataset. map_func can
return None to skip the data (e.g. in case of errors).
How None is handled depends on the style of `dataset`.
If `dataset` is map-style, it randomly tries other elements.
If `dataset` is iterable, it skips the data and tries the next.
"""
self._dataset = dataset
self._map_func = PicklableWrapper(map_func) # wrap so that a lambda will work
self._rng = random.Random(42)
self._fallback_candidates = set(range(len(dataset)))
def __new__(cls, dataset, map_func):
is_iterable = isinstance(dataset, data.IterableDataset)
if is_iterable:
return _MapIterableDataset(dataset, map_func)
else:
return super().__new__(cls)
def __getnewargs__(self):
return self._dataset, self._map_func
def __len__(self):
return len(self._dataset)
def __getitem__(self, idx):
retry_count = 0
cur_idx = int(idx)
while True:
data = self._map_func(self._dataset[cur_idx])
if data is not None:
self._fallback_candidates.add(cur_idx)
return data
# _map_func fails for this idx, use a random new index from the pool
retry_count += 1
self._fallback_candidates.discard(cur_idx)
cur_idx = self._rng.sample(self._fallback_candidates, k=1)[0]
if retry_count >= 3:
logger = logging.getLogger(__name__)
logger.warning(
"Failed to apply `_map_func` for idx: {}, retry count: {}".format(
idx, retry_count
)
)
class DatasetFromList(data.Dataset):
"""
Wrap a list to a torch Dataset. It produces elements of the list as data.
"""
def __init__(self, lst: list, copy: bool = True, serialize: bool = True):
"""
Args:
lst (list): a list which contains elements to produce.
copy (bool): whether to deepcopy the element when producing it,
so that the result can be modified in place without affecting the
source in the list.
serialize (bool): whether to hold memory using serialized objects, when
enabled, data loader workers can use shared RAM from master
process instead of making a copy.
"""
self._lst = lst
self._copy = copy
self._serialize = serialize
def _serialize(data):
buffer = pickle.dumps(data, protocol=-1)
return np.frombuffer(buffer, dtype=np.uint8)
if self._serialize:
logger = logging.getLogger(__name__)
logger.info(
"Serializing {} elements to byte tensors and concatenating them all ...".format(
len(self._lst)
)
)
self._lst = [_serialize(x) for x in self._lst]
self._addr = np.asarray([len(x) for x in self._lst], dtype=np.int64)
self._addr = np.cumsum(self._addr)
self._lst = np.concatenate(self._lst)
logger.info("Serialized dataset takes {:.2f} MiB".format(len(self._lst) / 1024 ** 2))
def __len__(self):
if self._serialize:
return len(self._addr)
else:
return len(self._lst)
def __getitem__(self, idx):
if self._serialize:
start_addr = 0 if idx == 0 else self._addr[idx - 1].item()
end_addr = self._addr[idx].item()
bytes = memoryview(self._lst[start_addr:end_addr])
return pickle.loads(bytes)
elif self._copy:
return copy.deepcopy(self._lst[idx])
else:
return self._lst[idx]
class ToIterableDataset(data.IterableDataset):
"""
Convert an old indices-based (also called map-style) dataset
to an iterable-style dataset.
"""
def __init__(self, dataset: data.Dataset, sampler: Sampler, shard_sampler: bool = True):
"""
Args:
dataset: an old-style dataset with ``__getitem__``
sampler: a cheap iterable that produces indices to be applied on ``dataset``.
shard_sampler: whether to shard the sampler based on the current pytorch data loader
worker id. When an IterableDataset is forked by pytorch's DataLoader into multiple
workers, it is responsible for sharding its data based on worker id so that workers
don't produce identical data.
Most samplers (like our TrainingSampler) do not shard based on dataloader worker id
and this argument should be set to True. But certain samplers may be already
sharded, in that case this argument should be set to False.
"""
assert not isinstance(dataset, data.IterableDataset), dataset
assert isinstance(sampler, Sampler), sampler
self.dataset = dataset
self.sampler = sampler
self.shard_sampler = shard_sampler
def __iter__(self):
if not self.shard_sampler:
sampler = self.sampler
else:
# With map-style dataset, `DataLoader(dataset, sampler)` runs the
# sampler in main process only. But `DataLoader(ToIterableDataset(dataset, sampler))`
# will run sampler in every of the N worker. So we should only keep 1/N of the ids on
# each worker. The assumption is that sampler is cheap to iterate so it's fine to
# discard ids in workers.
sampler = _shard_iterator_dataloader_worker(self.sampler)
for idx in sampler:
yield self.dataset[idx]
def __len__(self):
return len(self.sampler)
class AspectRatioGroupedDataset(data.IterableDataset):
"""
Batch data that have similar aspect ratio together.
In this implementation, images whose aspect ratio < (or >) 1 will
be batched together.
This improves training speed because the images then need less padding
to form a batch.
It assumes the underlying dataset produces dicts with "width" and "height" keys.
It will then produce a list of original dicts with length = batch_size,
all with similar aspect ratios.
"""
def __init__(self, dataset, batch_size):
"""
Args:
dataset: an iterable. Each element must be a dict with keys
"width" and "height", which will be used to batch data.
batch_size (int):
"""
self.dataset = dataset
self.batch_size = batch_size
self._buckets = [[] for _ in range(2)]
# Hard-coded two aspect ratio groups: w > h and w < h.
# Can add support for more aspect ratio groups, but doesn't seem useful
def __iter__(self):
for d in self.dataset:
w, h = d["width"], d["height"]
bucket_id = 0 if w > h else 1
bucket = self._buckets[bucket_id]
bucket.append(d)
if len(bucket) == self.batch_size:
yield bucket[:]
del bucket[:]
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