Skip to content

GitLab

"...git@developer.sourcefind.cn:chenpangpang/transformers.git" did not exist on "fd5320bb574ac86745d3aeb60ba2372abc4204ba"
Unverified Commit 5962d62b authored by Pavel Iakubovskii's avatar Pavel Iakubovskii Committed by GitHub
Browse files

Update object detection guide (#30683) 



* Object detection guide

* Minor update

* Minor updates, links

* Fix typo

* Wording, add albu space

* Add missing part

* Update docs/source/en/tasks/object_detection.md
Co-authored-by: default avataramyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/tasks/object_detection.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/tasks/object_detection.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Fix device, add imports for inference

---------
Co-authored-by: default avataramyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>
parent e7a5f45e
Hide whitespace changes
Inline Side-by-side
Showing with 1238 additions and 281 deletions
docs/source/en/tasks/object_detection.md
View file @ 5962d62b
...@@ -41,7 +41,8 @@ To see all architectures and checkpoints compatible with this task, we recommend ...@@ -41,7 +41,8 @@ To see all architectures and checkpoints compatible with this task, we recommend
Before you begin, make sure you have all the necessary libraries installed: Before you begin, make sure you have all the necessary libraries installed:
```bash ```bash
pip install -q datasets transformers accelerate evaluate albumentations pip install -q datasets transformers accelerate timm
pip install -q -U albumentations>=1.4.5 torchmetrics pycocotools
``` ```
You'll use 🤗 Datasets to load a dataset from the Hugging Face Hub, 🤗 Transformers to train your model, You'll use 🤗 Datasets to load a dataset from the Hugging Face Hub, 🤗 Transformers to train your model,
...@@ -56,47 +57,67 @@ When prompted, enter your token to log in: ...@@ -56,47 +57,67 @@ When prompted, enter your token to log in:
>>> notebook_login() >>> notebook_login()
``` ```
To get started, we'll define global constants, namely the model name and image size. For this tutorial, we'll use the conditional DETR model due to its faster convergence. Feel free to select any object detection model available in the `transformers` library.
```py
>>> MODEL_NAME = "microsoft/conditional-detr-resnet-50" # or "facebook/detr-resnet-50"
>>> IMAGE_SIZE = 480
```
## Load the CPPE-5 dataset ## Load the CPPE-5 dataset
The [CPPE-5 dataset](https://huggingface.co/datasets/cppe-5) contains images with The [CPPE-5 dataset](https://huggingface.co/datasets/cppe-5) contains images with
annotations identifying medical personal protective equipment (PPE) in the context of the COVID-19 pandemic. annotations identifying medical personal protective equipment (PPE) in the context of the COVID-19 pandemic.
Start by loading the dataset: Start by loading the dataset and creating a `validation` split from `train`:
```py ```py
>>> from datasets import load_dataset >>> from datasets import load_dataset
>>> cppe5 = load_dataset("cppe-5") >>> cppe5 = load_dataset("cppe-5")
>>> if "validation" not in cppe5:
... split = cppe5["train"].train_test_split(0.15, seed=1337)
... cppe5["train"] = split["train"]
... cppe5["validation"] = split["test"]
>>> cppe5 >>> cppe5
DatasetDict({ DatasetDict({
train: Dataset({ train: Dataset({
features: ['image_id', 'image', 'width', 'height', 'objects'], features: ['image_id', 'image', 'width', 'height', 'objects'],
num_rows: 1000 num_rows: 850
}) })
test: Dataset({ test: Dataset({
features: ['image_id', 'image', 'width', 'height', 'objects'], features: ['image_id', 'image', 'width', 'height', 'objects'],
num_rows: 29 num_rows: 29
}) })
validation: Dataset({
features: ['image_id', 'image', 'width', 'height', 'objects'],
num_rows: 150
})
}) })
``` ```
You'll see that this dataset already comes with a training set containing 1000 images and a test set with 29 images. You'll see that this dataset has 1000 images for train and validation sets and a test set with 29 images.
To get familiar with the data, explore what the examples look like. To get familiar with the data, explore what the examples look like.
```py ```py
>>> cppe5["train"][0] >>> cppe5["train"][0]
{'image_id': 15, {
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=943x663 at 0x7F9EC9E77C10>, 'image_id': 366,
'width': 943, 'image': <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=500x290>,
'height': 663, 'width': 500,
'objects': {'id': [114, 115, 116, 117], 'height': 500,
'area': [3796, 1596, 152768, 81002], 'objects': {
'bbox': [[302.0, 109.0, 73.0, 52.0], 'id': [1932, 1933, 1934],
[810.0, 100.0, 57.0, 28.0], 'area': [27063, 34200, 32431],
[160.0, 31.0, 248.0, 616.0], 'bbox': [[29.0, 11.0, 97.0, 279.0],
[741.0, 68.0, 202.0, 401.0]], [201.0, 1.0, 120.0, 285.0],
'category': [4, 4, 0, 0]}} [382.0, 0.0, 113.0, 287.0]],
'category': [0, 0, 0]
}
}
``` ```
The examples in the dataset have the following fields: The examples in the dataset have the following fields:
...@@ -121,8 +142,8 @@ To get an even better understanding of the data, visualize an example in the dat ...@@ -121,8 +142,8 @@ To get an even better understanding of the data, visualize an example in the dat
>>> import os >>> import os
>>> from PIL import Image, ImageDraw >>> from PIL import Image, ImageDraw
>>> image = cppe5["train"][0]["image"] >>> image = cppe5["train"][2]["image"]
>>> annotations = cppe5["train"][0]["objects"] >>> annotations = cppe5["train"][2]["objects"]
>>> draw = ImageDraw.Draw(image) >>> draw = ImageDraw.Draw(image)
>>> categories = cppe5["train"].features["objects"].feature["category"].names >>> categories = cppe5["train"].features["objects"].feature["category"].names
...@@ -150,27 +171,21 @@ To get an even better understanding of the data, visualize an example in the dat ...@@ -150,27 +171,21 @@ To get an even better understanding of the data, visualize an example in the dat
>>> image >>> image
``` ```
<div class="flex justify-center"> <div class="flex justify-center">
<img src="https://i.imgur.com/TdaqPJO.png" alt="CPPE-5 Image Example"/> <img src="https://i.imgur.com/oVQb9SF.png" alt="CPPE-5 Image Example"/>
</div> </div>
To visualize the bounding boxes with associated labels, you can get the labels from the dataset's metadata, specifically To visualize the bounding boxes with associated labels, you can get the labels from the dataset's metadata, specifically
the `category` field. the `category` field.
You'll also want to create dictionaries that map a label id to a label class (`id2label`) and the other way around (`label2id`). You'll also want to create dictionaries that map a label id to a label class (`id2label`) and the other way around (`label2id`).
You can use them later when setting up the model. Including these maps will make your model reusable by others if you share You can use them later when setting up the model. Including these maps will make your model reusable by others if you share
it on the Hugging Face Hub. Please note that, the part of above code that draws the bounding boxes assume that it is in `XYWH` (x,y co-ordinates and width and height of the box) format. It might not work for other formats like `(x1, y1, x2, y2)`. it on the Hugging Face Hub. Please note that, the part of above code that draws the bounding boxes assume that it is in `COCO` format `(x_min, y_min, width, height)`. It has to be adjusted to work for other formats like `(x_min, y_min, x_max, y_max)`.
As a final step of getting familiar with the data, explore it for potential issues. One common problem with datasets for As a final step of getting familiar with the data, explore it for potential issues. One common problem with datasets for
object detection is bounding boxes that "stretch" beyond the edge of the image. Such "runaway" bounding boxes can raise object detection is bounding boxes that "stretch" beyond the edge of the image. Such "runaway" bounding boxes can raise
errors during training and should be addressed at this stage. There are a few examples with this issue in this dataset. errors during training and should be addressed. There are a few examples with this issue in this dataset.
To keep things simple in this guide, we remove these images from the data. To keep things simple in this guide, we will set `clip=True` for `BboxParams` in transformations below.
```py
>>> remove_idx = [590, 821, 822, 875, 876, 878, 879]
>>> keep = [i for i in range(len(cppe5["train"])) if i not in remove_idx]
>>> cppe5["train"] = cppe5["train"].select(keep)
```
## Preprocess the data ## Preprocess the data
...@@ -189,8 +204,13 @@ Instantiate the image processor from the same checkpoint as the model you want t ...@@ -189,8 +204,13 @@ Instantiate the image processor from the same checkpoint as the model you want t
```py ```py
>>> from transformers import AutoImageProcessor >>> from transformers import AutoImageProcessor
>>> checkpoint = "facebook/detr-resnet-50" >>> image_processor = AutoImageProcessor.from_pretrained(
>>> image_processor = AutoImageProcessor.from_pretrained(checkpoint) ... MODEL_NAME,
... # At this moment we recommend using external transform to pad and resize images.
... # It`s faster and yields better results for object-detection models.
... do_pad=False,
... do_resize=False,
... )
``` ```
Before passing the images to the `image_processor`, apply two preprocessing transformations to the dataset: Before passing the images to the `image_processor`, apply two preprocessing transformations to the dataset:
...@@ -201,20 +221,35 @@ First, to make sure the model does not overfit on the training data, you can app ...@@ -201,20 +221,35 @@ First, to make sure the model does not overfit on the training data, you can app
This library ensures that transformations affect the image and update the bounding boxes accordingly. This library ensures that transformations affect the image and update the bounding boxes accordingly.
The 🤗 Datasets library documentation has a detailed [guide on how to augment images for object detection](https://huggingface.co/docs/datasets/object_detection), The 🤗 Datasets library documentation has a detailed [guide on how to augment images for object detection](https://huggingface.co/docs/datasets/object_detection),
and it uses the exact same dataset as an example. Apply the same approach here, resize each image to (480, 480), and it uses the exact same dataset as an example. Apply the same approach here, resize each image to (480, 480),
flip it horizontally, and brighten it: flip it horizontally, and brighten it. For additional augmentation options, explore the [Albumentations Demo Space](https://huggingface.co/spaces/qubvel-hf/albumentations-demo).
```py ```py
>>> import albumentations >>> import albumentations as A
>>> import numpy as np
>>> import torch >>> max_size = IMAGE_SIZE
>>> # Resize image longest edge to 480 and then pad image to square 480x480.
>>> # This padding and resizing strategy give better results, see
>>> # https://github.com/huggingface/transformers/pull/30422#discussion_r1584647408
>>> basic_transforms = [
... A.LongestMaxSize(max_size=max_size),
... A.PadIfNeeded(max_size, max_size, border_mode=0, value=(128, 128, 128), position="top_left"),
... ]
>>> transform = albumentations.Compose( >>> train_augment_and_transform = A.Compose(
... [ ... [
... albumentations.Resize(480, 480), ... A.Perspective(p=0.1),
... albumentations.HorizontalFlip(p=1.0), ... A.HorizontalFlip(p=0.5),
... albumentations.RandomBrightnessContrast(p=1.0), ... A.RandomBrightnessContrast(p=0.5),
... A.HueSaturationValue(p=0.1),
... *basic_transforms,
... ], ... ],
... bbox_params=albumentations.BboxParams(format="coco", label_fields=["category"]), ... bbox_params=A.BboxParams(format="coco", label_fields=["category"], clip=True, min_area=25),
... )
>>> validation_transform = A.Compose(
... basic_transforms,
... bbox_params=A.BboxParams(format="coco", label_fields=["category"], clip=True),
... ) ... )
``` ```
...@@ -222,43 +257,65 @@ The `image_processor` expects the annotations to be in the following format: `{' ...@@ -222,43 +257,65 @@ The `image_processor` expects the annotations to be in the following format: `{'
where each dictionary is a COCO object annotation. Let's add a function to reformat annotations for a single example: where each dictionary is a COCO object annotation. Let's add a function to reformat annotations for a single example:
```py ```py
>>> def formatted_anns(image_id, category, area, bbox): >>> def format_image_annotations_as_coco(image_id, categories, areas, bboxes):
... """Format one set of image annotations to the COCO format
... Args:
... image_id (str): image id. e.g. "0001"
... categories (List[int]): list of categories/class labels corresponding to provided bounding boxes
... areas (List[float]): list of corresponding areas to provided bounding boxes
... bboxes (List[Tuple[float]]): list of bounding boxes provided in COCO format
... ([center_x, center_y, width, height] in absolute coordinates)
... Returns:
... dict: {
... "image_id": image id,
... "annotations": list of formatted annotations
... }
... """
... annotations = [] ... annotations = []
... for i in range(0, len(category)): ... for category, area, bbox in zip(categories, areas, bboxes):
... new_ann = { ... formatted_annotation = {
... "image_id": image_id, ... "image_id": image_id,
... "category_id": category[i], ... "category_id": category,
... "isCrowd": 0, ... "iscrowd": 0,
... "area": area[i], ... "area": area,
... "bbox": list(bbox[i]), ... "bbox": list(bbox),
... } ... }
... annotations.append(new_ann) ... annotations.append(formatted_annotation)
... return {
... "image_id": image_id,
... "annotations": annotations,
... }
... return annotations
``` ```
Now you can combine the image and annotation transformations to use on a batch of examples: Now you can combine the image and annotation transformations to use on a batch of examples:
```py ```py
>>> # transforming a batch >>> def augment_and_transform_batch(examples, transform, image_processor):
>>> def transform_aug_ann(examples): ... """Apply augmentations and format annotations in COCO format for object detection task"""
... image_ids = examples["image_id"]
... images, bboxes, area, categories = [], [], [], [] ... images = []
... for image, objects in zip(examples["image"], examples["objects"]): ... annotations = []
... image = np.array(image.convert("RGB"))[:, :, ::-1] ... for image_id, image, objects in zip(examples["image_id"], examples["image"], examples["objects"]):
... out = transform(image=image, bboxes=objects["bbox"], category=objects["category"]) ... image = np.array(image.convert("RGB"))
... area.append(objects["area"]) ... # apply augmentations
... images.append(out["image"]) ... output = transform(image=image, bboxes=objects["bbox"], category=objects["category"])
... bboxes.append(out["bboxes"]) ... images.append(output["image"])
... categories.append(out["category"])
... # format annotations in COCO format
... targets = [ ... formatted_annotations = format_image_annotations_as_coco(
... {"image_id": id_, "annotations": formatted_anns(id_, cat_, ar_, box_)} ... image_id, output["category"], objects["area"], output["bboxes"]
... for id_, cat_, ar_, box_ in zip(image_ids, categories, area, bboxes) ... )
... ] ... annotations.append(formatted_annotations)
... return image_processor(images=images, annotations=targets, return_tensors="pt") ... # Apply the image processor transformations: resizing, rescaling, normalization
... result = image_processor(images=images, annotations=annotations, return_tensors="pt")
... return result
``` ```
Apply this preprocessing function to the entire dataset using 🤗 Datasets [`~datasets.Dataset.with_transform`] method. This method applies Apply this preprocessing function to the entire dataset using 🤗 Datasets [`~datasets.Dataset.with_transform`] method. This method applies
...@@ -268,39 +325,49 @@ At this point, you can check what an example from the dataset looks like after t ...@@ -268,39 +325,49 @@ At this point, you can check what an example from the dataset looks like after t
with `pixel_values`, a tensor with `pixel_mask`, and `labels`. with `pixel_values`, a tensor with `pixel_mask`, and `labels`.
```py ```py
>>> cppe5["train"] = cppe5["train"].with_transform(transform_aug_ann) >>> from functools import partial
>>> # Make transform functions for batch and apply for dataset splits
>>> train_transform_batch = partial(
... augment_and_transform_batch, transform=train_augment_and_transform, image_processor=image_processor
... )
>>> validation_transform_batch = partial(
... augment_and_transform_batch, transform=validation_transform, image_processor=image_processor
... )
>>> cppe5["train"] = cppe5["train"].with_transform(train_transform_batch)
>>> cppe5["validation"] = cppe5["validation"].with_transform(validation_transform_batch)
>>> cppe5["test"] = cppe5["test"].with_transform(validation_transform_batch)
>>> cppe5["train"][15] >>> cppe5["train"][15]
{'pixel_values': tensor([[[ 0.9132, 0.9132, 0.9132, ..., -1.9809, -1.9809, -1.9809], {'pixel_values': tensor([[[ 1.9235, 1.9407, 1.9749, ..., -0.7822, -0.7479, -0.6965],
[ 0.9132, 0.9132, 0.9132, ..., -1.9809, -1.9809, -1.9809], [ 1.9578, 1.9749, 1.9920, ..., -0.7993, -0.7650, -0.7308],
[ 0.9132, 0.9132, 0.9132, ..., -1.9638, -1.9638, -1.9638], [ 2.0092, 2.0092, 2.0263, ..., -0.8507, -0.8164, -0.7822],
..., ...,
[-1.5699, -1.5699, -1.5699, ..., -1.9980, -1.9980, -1.9980], [ 0.0741, 0.0741, 0.0741, ..., 0.0741, 0.0741, 0.0741],
[-1.5528, -1.5528, -1.5528, ..., -1.9980, -1.9809, -1.9809], [ 0.0741, 0.0741, 0.0741, ..., 0.0741, 0.0741, 0.0741],
[-1.5528, -1.5528, -1.5528, ..., -1.9980, -1.9809, -1.9809]], [ 0.0741, 0.0741, 0.0741, ..., 0.0741, 0.0741, 0.0741]],
[[ 1.3081, 1.3081, 1.3081, ..., -1.8431, -1.8431, -1.8431], [[ 1.6232, 1.6408, 1.6583, ..., 0.8704, 1.0105, 1.1331],
[ 1.3081, 1.3081, 1.3081, ..., -1.8431, -1.8431, -1.8431], [ 1.6408, 1.6583, 1.6758, ..., 0.8529, 0.9930, 1.0980],
[ 1.3081, 1.3081, 1.3081, ..., -1.8256, -1.8256, -1.8256], [ 1.6933, 1.6933, 1.7108, ..., 0.8179, 0.9580, 1.0630],
..., ...,
[-1.3179, -1.3179, -1.3179, ..., -1.8606, -1.8606, -1.8606], [ 0.2052, 0.2052, 0.2052, ..., 0.2052, 0.2052, 0.2052],
[-1.3004, -1.3004, -1.3004, ..., -1.8606, -1.8431, -1.8431], [ 0.2052, 0.2052, 0.2052, ..., 0.2052, 0.2052, 0.2052],
[-1.3004, -1.3004, -1.3004, ..., -1.8606, -1.8431, -1.8431]], [ 0.2052, 0.2052, 0.2052, ..., 0.2052, 0.2052, 0.2052]],
[[ 1.4200, 1.4200, 1.4200, ..., -1.6476, -1.6476, -1.6476], [[ 1.8905, 1.9080, 1.9428, ..., -0.1487, -0.0964, -0.0615],
[ 1.4200, 1.4200, 1.4200, ..., -1.6476, -1.6476, -1.6476], [ 1.9254, 1.9428, 1.9603, ..., -0.1661, -0.1138, -0.0790],
[ 1.4200, 1.4200, 1.4200, ..., -1.6302, -1.6302, -1.6302], [ 1.9777, 1.9777, 1.9951, ..., -0.2010, -0.1138, -0.0790],
..., ...,
[-1.0201, -1.0201, -1.0201, ..., -1.5604, -1.5604, -1.5604], [ 0.4265, 0.4265, 0.4265, ..., 0.4265, 0.4265, 0.4265],
[-1.0027, -1.0027, -1.0027, ..., -1.5604, -1.5430, -1.5430], [ 0.4265, 0.4265, 0.4265, ..., 0.4265, 0.4265, 0.4265],
[-1.0027, -1.0027, -1.0027, ..., -1.5604, -1.5430, -1.5430]]]), [ 0.4265, 0.4265, 0.4265, ..., 0.4265, 0.4265, 0.4265]]]),
'pixel_mask': tensor([[1, 1, 1, ..., 1, 1, 1], 'labels': {'image_id': tensor([688]), 'class_labels': tensor([3, 4, 2, 0, 0]), 'boxes': tensor([[0.4700, 0.1933, 0.1467, 0.0767],
[1, 1, 1, ..., 1, 1, 1], [0.4858, 0.2600, 0.1150, 0.1000],
[1, 1, 1, ..., 1, 1, 1], [0.4042, 0.4517, 0.1217, 0.1300],
..., [0.4242, 0.3217, 0.3617, 0.5567],
[1, 1, 1, ..., 1, 1, 1], [0.6617, 0.4033, 0.5400, 0.4533]]), 'area': tensor([ 4048., 4140., 5694., 72478., 88128.]), 'iscrowd': tensor([0, 0, 0, 0, 0]), 'orig_size': tensor([480, 480])}}
[1, 1, 1, ..., 1, 1, 1],
[1, 1, 1, ..., 1, 1, 1]]),
'labels': {'size': tensor([800, 800]), 'image_id': tensor([756]), 'class_labels': tensor([4]), 'boxes': tensor([[0.7340, 0.6986, 0.3414, 0.5944]]), 'area': tensor([519544.4375]), 'iscrowd': tensor([0]), 'orig_size': tensor([480, 480])}}
``` ```
You have successfully augmented the individual images and prepared their annotations. However, preprocessing isn't You have successfully augmented the individual images and prepared their annotations. However, preprocessing isn't
...@@ -309,18 +376,137 @@ Pad images (which are now `pixel_values`) to the largest image in a batch, and c ...@@ -309,18 +376,137 @@ Pad images (which are now `pixel_values`) to the largest image in a batch, and c
to indicate which pixels are real (1) and which are padding (0). to indicate which pixels are real (1) and which are padding (0).
```py ```py
>>> import torch
>>> def collate_fn(batch): >>> def collate_fn(batch):
... pixel_values = [item["pixel_values"] for item in batch] ... data = {}
... encoding = image_processor.pad(pixel_values, return_tensors="pt") ... data["pixel_values"] = torch.stack([x["pixel_values"] for x in batch])
... labels = [item["labels"] for item in batch] ... data["labels"] = [x["labels"] for x in batch]
... batch = {} ... if "pixel_mask" in batch[0]:
... batch["pixel_values"] = encoding["pixel_values"] ... data["pixel_mask"] = torch.stack([x["pixel_mask"] for x in batch])
... batch["pixel_mask"] = encoding["pixel_mask"] ... return data
... batch["labels"] = labels
... return batch ```
## Preparing function to compute mAP
Object detection models are commonly evaluated with a set of <a href="https://cocodataset.org/#detection-eval">COCO-style metrics</a>. We are going to use `torchmetrics` to compute `mAP` (mean average precision) and `mAR` (mean average recall) metrics and will wrap it to `compute_metrics` function in order to use in [`Trainer`] for evaluation.
Intermediate format of boxes used for training is `YOLO` (normalized) but we will compute metrics for boxes in `Pascal VOC` (absolute) format in order to correctly handle box areas. Let's define a function that converts bounding boxes to `Pascal VOC` format:
```py
>>> from transformers.image_transforms import center_to_corners_format
>>> def convert_bbox_yolo_to_pascal(boxes, image_size):
... """
... Convert bounding boxes from YOLO format (x_center, y_center, width, height) in range [0, 1]
... to Pascal VOC format (x_min, y_min, x_max, y_max) in absolute coordinates.
... Args:
... boxes (torch.Tensor): Bounding boxes in YOLO format
... image_size (Tuple[int, int]): Image size in format (height, width)
... Returns:
... torch.Tensor: Bounding boxes in Pascal VOC format (x_min, y_min, x_max, y_max)
... """
... # convert center to corners format
... boxes = center_to_corners_format(boxes)
... # convert to absolute coordinates
... height, width = image_size
... boxes = boxes * torch.tensor([[width, height, width, height]])
... return boxes
```
Then, in `compute_metrics` function we collect `predicted` and `target` bounding boxes, scores and labels from evaluation loop results and pass it to the scoring function.
```py
>>> import numpy as np
>>> from dataclasses import dataclass
>>> from torchmetrics.detection.mean_ap import MeanAveragePrecision
>>> @dataclass
>>> class ModelOutput:
... logits: torch.Tensor
... pred_boxes: torch.Tensor
>>> @torch.no_grad()
>>> def compute_metrics(evaluation_results, image_processor, threshold=0.0, id2label=None):
... """
... Compute mean average mAP, mAR and their variants for the object detection task.
... Args:
... evaluation_results (EvalPrediction): Predictions and targets from evaluation.
... threshold (float, optional): Threshold to filter predicted boxes by confidence. Defaults to 0.0.
... id2label (Optional[dict], optional): Mapping from class id to class name. Defaults to None.
... Returns:
... Mapping[str, float]: Metrics in a form of dictionary {<metric_name>: <metric_value>}
... """
... predictions, targets = evaluation_results.predictions, evaluation_results.label_ids
... # For metric computation we need to provide:
... # - targets in a form of list of dictionaries with keys "boxes", "labels"
... # - predictions in a form of list of dictionaries with keys "boxes", "scores", "labels"
... image_sizes = []
... post_processed_targets = []
... post_processed_predictions = []
... # Collect targets in the required format for metric computation
... for batch in targets:
... # collect image sizes, we will need them for predictions post processing
... batch_image_sizes = torch.tensor(np.array([x["orig_size"] for x in batch]))
... image_sizes.append(batch_image_sizes)
... # collect targets in the required format for metric computation
... # boxes were converted to YOLO format needed for model training
... # here we will convert them to Pascal VOC format (x_min, y_min, x_max, y_max)
... for image_target in batch:
... boxes = torch.tensor(image_target["boxes"])
... boxes = convert_bbox_yolo_to_pascal(boxes, image_target["orig_size"])
... labels = torch.tensor(image_target["class_labels"])
... post_processed_targets.append({"boxes": boxes, "labels": labels})
... # Collect predictions in the required format for metric computation,
... # model produce boxes in YOLO format, then image_processor convert them to Pascal VOC format
... for batch, target_sizes in zip(predictions, image_sizes):
... batch_logits, batch_boxes = batch[1], batch[2]
... output = ModelOutput(logits=torch.tensor(batch_logits), pred_boxes=torch.tensor(batch_boxes))
... post_processed_output = image_processor.post_process_object_detection(
... output, threshold=threshold, target_sizes=target_sizes
... )
... post_processed_predictions.extend(post_processed_output)
... # Compute metrics
... metric = MeanAveragePrecision(box_format="xyxy", class_metrics=True)
... metric.update(post_processed_predictions, post_processed_targets)
... metrics = metric.compute()
... # Replace list of per class metrics with separate metric for each class
... classes = metrics.pop("classes")
... map_per_class = metrics.pop("map_per_class")
... mar_100_per_class = metrics.pop("mar_100_per_class")
... for class_id, class_map, class_mar in zip(classes, map_per_class, mar_100_per_class):
... class_name = id2label[class_id.item()] if id2label is not None else class_id.item()
... metrics[f"map_{class_name}"] = class_map
... metrics[f"mar_100_{class_name}"] = class_mar
... metrics = {k: round(v.item(), 4) for k, v in metrics.items()}
... return metrics
>>> eval_compute_metrics_fn = partial(
... compute_metrics, image_processor=image_processor, id2label=id2label, threshold=0.0
... )
``` ```
## Training the DETR model ## Training the detection model
You have done most of the heavy lifting in the previous sections, so now you are ready to train your model! You have done most of the heavy lifting in the previous sections, so now you are ready to train your model!
The images in this dataset are still quite large, even after resizing. This means that finetuning this model will The images in this dataset are still quite large, even after resizing. This means that finetuning this model will
require at least one GPU. require at least one GPU.
...@@ -338,17 +524,20 @@ and `id2label` maps that you created earlier from the dataset's metadata. Additi ...@@ -338,17 +524,20 @@ and `id2label` maps that you created earlier from the dataset's metadata. Additi
>>> from transformers import AutoModelForObjectDetection >>> from transformers import AutoModelForObjectDetection
>>> model = AutoModelForObjectDetection.from_pretrained( >>> model = AutoModelForObjectDetection.from_pretrained(
... checkpoint, ... MODEL_NAME,
... id2label=id2label, ... id2label=id2label,
... label2id=label2id, ... label2id=label2id,
... ignore_mismatched_sizes=True, ... ignore_mismatched_sizes=True,
... revision="no_timm", # DETR models can be loaded without timm
... ) ... )
``` ```
In the [`TrainingArguments`] use `output_dir` to specify where to save your model, then configure hyperparameters as you see fit. In the [`TrainingArguments`] use `output_dir` to specify where to save your model, then configure hyperparameters as you see fit. For `num_train_epochs=30` training will take about 35 minutes in Google Colab T4 GPU, increase the number of epoch to get better results.
It is important you do not remove unused columns because this will drop the image column. Without the image column, you
Important notes:
- Do not remove unused columns because this will drop the image column. Without the image column, you
can't create `pixel_values`. For this reason, set `remove_unused_columns` to `False`. can't create `pixel_values`. For this reason, set `remove_unused_columns` to `False`.
- Set `eval_do_concat_batches=False` to get proper evaluation results. Images have different number of target boxes, if batches are concatenated we will not be able to determine which boxes belongs to particular image.
If you wish to share your model by pushing to the Hub, set `push_to_hub` to `True` (you must be signed in to Hugging If you wish to share your model by pushing to the Hub, set `push_to_hub` to `True` (you must be signed in to Hugging
Face to upload your model). Face to upload your model).
...@@ -356,16 +545,23 @@ Face to upload your model). ...@@ -356,16 +545,23 @@ Face to upload your model).
>>> from transformers import TrainingArguments >>> from transformers import TrainingArguments
>>> training_args = TrainingArguments( >>> training_args = TrainingArguments(
... output_dir="detr-resnet-50_finetuned_cppe5", ... output_dir="detr_finetuned_cppe5",
... num_train_epochs=30,
... fp16=False,
... per_device_train_batch_size=8, ... per_device_train_batch_size=8,
... num_train_epochs=100, ... dataloader_num_workers=4,
... fp16=True, ... learning_rate=5e-5,
... save_steps=200, ... lr_scheduler_type="cosine",
... logging_steps=50,
... learning_rate=1e-5,
... weight_decay=1e-4, ... weight_decay=1e-4,
... max_grad_norm=0.01,
... metric_for_best_model="eval_map",
... greater_is_better=True,
... load_best_model_at_end=True,
... eval_strategy="epoch",
... save_strategy="epoch",
... save_total_limit=2, ... save_total_limit=2,
... remove_unused_columns=False, ... remove_unused_columns=False,
... eval_do_concat_batches=False,
... push_to_hub=True, ... push_to_hub=True,
... ) ... )
``` ```
...@@ -378,13 +574,864 @@ Finally, bring everything together, and call [`~transformers.Trainer.train`]: ...@@ -378,13 +574,864 @@ Finally, bring everything together, and call [`~transformers.Trainer.train`]:
>>> trainer = Trainer( >>> trainer = Trainer(
... model=model, ... model=model,
... args=training_args, ... args=training_args,
... data_collator=collate_fn,
... train_dataset=cppe5["train"], ... train_dataset=cppe5["train"],
... eval_dataset=cppe5["validation"],
... tokenizer=image_processor, ... tokenizer=image_processor,
... data_collator=collate_fn,
... compute_metrics=eval_compute_metrics_fn,
... ) ... )
>>> trainer.train() >>> trainer.train()
``` ```
<div>
<progress value='3210' max='3210' style='width:300px; height:20px; vertical-align: middle;'></progress>
[3210/3210 26:07, Epoch 30/30]
</div>
<table border="1" class="dataframe">
<thead>
<tr style="text-align: left;">
<th>Epoch</th>
<th>Training Loss</th>
<th>Validation Loss</th>
<th>Map</th>
<th>Map 50</th>
<th>Map 75</th>
<th>Map Small</th>
<th>Map Medium</th>
<th>Map Large</th>
<th>Mar 1</th>
<th>Mar 10</th>
<th>Mar 100</th>
<th>Mar Small</th>
<th>Mar Medium</th>
<th>Mar Large</th>
<th>Map Coverall</th>
<th>Mar 100 Coverall</th>
<th>Map Face Shield</th>
<th>Mar 100 Face Shield</th>
<th>Map Gloves</th>
<th>Mar 100 Gloves</th>
<th>Map Goggles</th>
<th>Mar 100 Goggles</th>
<th>Map Mask</th>
<th>Mar 100 Mask</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td>No log</td>
<td>2.629903</td>
<td>0.008900</td>
<td>0.023200</td>
<td>0.006500</td>
<td>0.001300</td>
<td>0.002800</td>
<td>0.020500</td>
<td>0.021500</td>
<td>0.070400</td>
<td>0.101400</td>
<td>0.007600</td>
<td>0.106200</td>
<td>0.096100</td>
<td>0.036700</td>
<td>0.232000</td>
<td>0.000300</td>
<td>0.019000</td>
<td>0.003900</td>
<td>0.125400</td>
<td>0.000100</td>
<td>0.003100</td>
<td>0.003500</td>
<td>0.127600</td>
</tr>
<tr>
<td>2</td>
<td>No log</td>
<td>3.479864</td>
<td>0.014800</td>
<td>0.034600</td>
<td>0.010800</td>
<td>0.008600</td>
<td>0.011700</td>
<td>0.012500</td>
<td>0.041100</td>
<td>0.098700</td>
<td>0.130000</td>
<td>0.056000</td>
<td>0.062200</td>
<td>0.111900</td>
<td>0.053500</td>
<td>0.447300</td>
<td>0.010600</td>
<td>0.100000</td>
<td>0.000200</td>
<td>0.022800</td>
<td>0.000100</td>
<td>0.015400</td>
<td>0.009700</td>
<td>0.064400</td>
</tr>
<tr>
<td>3</td>
<td>No log</td>
<td>2.107622</td>
<td>0.041700</td>
<td>0.094000</td>
<td>0.034300</td>
<td>0.024100</td>
<td>0.026400</td>
<td>0.047400</td>
<td>0.091500</td>
<td>0.182800</td>
<td>0.225800</td>
<td>0.087200</td>
<td>0.199400</td>
<td>0.210600</td>
<td>0.150900</td>
<td>0.571200</td>
<td>0.017300</td>
<td>0.101300</td>
<td>0.007300</td>
<td>0.180400</td>
<td>0.002100</td>
<td>0.026200</td>
<td>0.031000</td>
<td>0.250200</td>
</tr>
<tr>
<td>4</td>
<td>No log</td>
<td>2.031242</td>
<td>0.055900</td>
<td>0.120600</td>
<td>0.046900</td>
<td>0.013800</td>
<td>0.038100</td>
<td>0.090300</td>
<td>0.105900</td>
<td>0.225600</td>
<td>0.266100</td>
<td>0.130200</td>
<td>0.228100</td>
<td>0.330000</td>
<td>0.191000</td>
<td>0.572100</td>
<td>0.010600</td>
<td>0.157000</td>
<td>0.014600</td>
<td>0.235300</td>
<td>0.001700</td>
<td>0.052300</td>
<td>0.061800</td>
<td>0.313800</td>
</tr>
<tr>
<td>5</td>
<td>3.889400</td>
<td>1.883433</td>
<td>0.089700</td>
<td>0.201800</td>
<td>0.067300</td>
<td>0.022800</td>
<td>0.065300</td>
<td>0.129500</td>
<td>0.136000</td>
<td>0.272200</td>
<td>0.303700</td>
<td>0.112900</td>
<td>0.312500</td>
<td>0.424600</td>
<td>0.300200</td>
<td>0.585100</td>
<td>0.032700</td>
<td>0.202500</td>
<td>0.031300</td>
<td>0.271000</td>
<td>0.008700</td>
<td>0.126200</td>
<td>0.075500</td>
<td>0.333800</td>
</tr>
<tr>
<td>6</td>
<td>3.889400</td>
<td>1.807503</td>
<td>0.118500</td>
<td>0.270900</td>
<td>0.090200</td>
<td>0.034900</td>
<td>0.076700</td>
<td>0.152500</td>
<td>0.146100</td>
<td>0.297800</td>
<td>0.325400</td>
<td>0.171700</td>
<td>0.283700</td>
<td>0.545900</td>
<td>0.396900</td>
<td>0.554500</td>
<td>0.043000</td>
<td>0.262000</td>
<td>0.054500</td>
<td>0.271900</td>
<td>0.020300</td>
<td>0.230800</td>
<td>0.077600</td>
<td>0.308000</td>
</tr>
<tr>
<td>7</td>
<td>3.889400</td>
<td>1.716169</td>
<td>0.143500</td>
<td>0.307700</td>
<td>0.123200</td>
<td>0.045800</td>
<td>0.097800</td>
<td>0.258300</td>
<td>0.165300</td>
<td>0.327700</td>
<td>0.352600</td>
<td>0.140900</td>
<td>0.336700</td>
<td>0.599400</td>
<td>0.442900</td>
<td>0.620700</td>
<td>0.069400</td>
<td>0.301300</td>
<td>0.081600</td>
<td>0.292000</td>
<td>0.011000</td>
<td>0.230800</td>
<td>0.112700</td>
<td>0.318200</td>
</tr>
<tr>
<td>8</td>
<td>3.889400</td>
<td>1.679014</td>
<td>0.153000</td>
<td>0.355800</td>
<td>0.127900</td>
<td>0.038700</td>
<td>0.115600</td>
<td>0.291600</td>
<td>0.176000</td>
<td>0.322500</td>
<td>0.349700</td>
<td>0.135600</td>
<td>0.326100</td>
<td>0.643700</td>
<td>0.431700</td>
<td>0.582900</td>
<td>0.069800</td>
<td>0.265800</td>
<td>0.088600</td>
<td>0.274600</td>
<td>0.028300</td>
<td>0.280000</td>
<td>0.146700</td>
<td>0.345300</td>
</tr>
<tr>
<td>9</td>
<td>3.889400</td>
<td>1.618239</td>
<td>0.172100</td>
<td>0.375300</td>
<td>0.137600</td>
<td>0.046100</td>
<td>0.141700</td>
<td>0.308500</td>
<td>0.194000</td>
<td>0.356200</td>
<td>0.386200</td>
<td>0.162400</td>
<td>0.359200</td>
<td>0.677700</td>
<td>0.469800</td>
<td>0.623900</td>
<td>0.102100</td>
<td>0.317700</td>
<td>0.099100</td>
<td>0.290200</td>
<td>0.029300</td>
<td>0.335400</td>
<td>0.160200</td>
<td>0.364000</td>
</tr>
<tr>
<td>10</td>
<td>1.599700</td>
<td>1.572512</td>
<td>0.179500</td>
<td>0.400400</td>
<td>0.147200</td>
<td>0.056500</td>
<td>0.141700</td>
<td>0.316700</td>
<td>0.213100</td>
<td>0.357600</td>
<td>0.381300</td>
<td>0.197900</td>
<td>0.344300</td>
<td>0.638500</td>
<td>0.466900</td>
<td>0.623900</td>
<td>0.101300</td>
<td>0.311400</td>
<td>0.104700</td>
<td>0.279500</td>
<td>0.051600</td>
<td>0.338500</td>
<td>0.173000</td>
<td>0.353300</td>
</tr>
<tr>
<td>11</td>
<td>1.599700</td>
<td>1.528889</td>
<td>0.192200</td>
<td>0.415000</td>
<td>0.160800</td>
<td>0.053700</td>
<td>0.150500</td>
<td>0.378000</td>
<td>0.211500</td>
<td>0.371700</td>
<td>0.397800</td>
<td>0.204900</td>
<td>0.374600</td>
<td>0.684800</td>
<td>0.491900</td>
<td>0.632400</td>
<td>0.131200</td>
<td>0.346800</td>
<td>0.122000</td>
<td>0.300900</td>
<td>0.038400</td>
<td>0.344600</td>
<td>0.177500</td>
<td>0.364400</td>
</tr>
<tr>
<td>12</td>
<td>1.599700</td>
<td>1.517532</td>
<td>0.198300</td>
<td>0.429800</td>
<td>0.159800</td>
<td>0.066400</td>
<td>0.162900</td>
<td>0.383300</td>
<td>0.220700</td>
<td>0.382100</td>
<td>0.405400</td>
<td>0.214800</td>
<td>0.383200</td>
<td>0.672900</td>
<td>0.469000</td>
<td>0.610400</td>
<td>0.167800</td>
<td>0.379700</td>
<td>0.119700</td>
<td>0.307100</td>
<td>0.038100</td>
<td>0.335400</td>
<td>0.196800</td>
<td>0.394200</td>
</tr>
<tr>
<td>13</td>
<td>1.599700</td>
<td>1.488849</td>
<td>0.209800</td>
<td>0.452300</td>
<td>0.172300</td>
<td>0.094900</td>
<td>0.171100</td>
<td>0.437800</td>
<td>0.222000</td>
<td>0.379800</td>
<td>0.411500</td>
<td>0.203800</td>
<td>0.397300</td>
<td>0.707500</td>
<td>0.470700</td>
<td>0.620700</td>
<td>0.186900</td>
<td>0.407600</td>
<td>0.124200</td>
<td>0.306700</td>
<td>0.059300</td>
<td>0.355400</td>
<td>0.207700</td>
<td>0.367100</td>
</tr>
<tr>
<td>14</td>
<td>1.599700</td>
<td>1.482210</td>
<td>0.228900</td>
<td>0.482600</td>
<td>0.187800</td>
<td>0.083600</td>
<td>0.191800</td>
<td>0.444100</td>
<td>0.225900</td>
<td>0.376900</td>
<td>0.407400</td>
<td>0.182500</td>
<td>0.384800</td>
<td>0.700600</td>
<td>0.512100</td>
<td>0.640100</td>
<td>0.175000</td>
<td>0.363300</td>
<td>0.144300</td>
<td>0.300000</td>
<td>0.083100</td>
<td>0.363100</td>
<td>0.229900</td>
<td>0.370700</td>
</tr>
<tr>
<td>15</td>
<td>1.326800</td>
<td>1.475198</td>
<td>0.216300</td>
<td>0.455600</td>
<td>0.174900</td>
<td>0.088500</td>
<td>0.183500</td>
<td>0.424400</td>
<td>0.226900</td>
<td>0.373400</td>
<td>0.404300</td>
<td>0.199200</td>
<td>0.396400</td>
<td>0.677800</td>
<td>0.496300</td>
<td>0.633800</td>
<td>0.166300</td>
<td>0.392400</td>
<td>0.128900</td>
<td>0.312900</td>
<td>0.085200</td>
<td>0.312300</td>
<td>0.205000</td>
<td>0.370200</td>
</tr>
<tr>
<td>16</td>
<td>1.326800</td>
<td>1.459697</td>
<td>0.233200</td>
<td>0.504200</td>
<td>0.192200</td>
<td>0.096000</td>
<td>0.202000</td>
<td>0.430800</td>
<td>0.239100</td>
<td>0.382400</td>
<td>0.412600</td>
<td>0.219500</td>
<td>0.403100</td>
<td>0.670400</td>
<td>0.485200</td>
<td>0.625200</td>
<td>0.196500</td>
<td>0.410100</td>
<td>0.135700</td>
<td>0.299600</td>
<td>0.123100</td>
<td>0.356900</td>
<td>0.225300</td>
<td>0.371100</td>
</tr>
<tr>
<td>17</td>
<td>1.326800</td>
<td>1.407340</td>
<td>0.243400</td>
<td>0.511900</td>
<td>0.204500</td>
<td>0.121000</td>
<td>0.215700</td>
<td>0.468000</td>
<td>0.246200</td>
<td>0.394600</td>
<td>0.424200</td>
<td>0.225900</td>
<td>0.416100</td>
<td>0.705200</td>
<td>0.494900</td>
<td>0.638300</td>
<td>0.224900</td>
<td>0.430400</td>
<td>0.157200</td>
<td>0.317900</td>
<td>0.115700</td>
<td>0.369200</td>
<td>0.224200</td>
<td>0.365300</td>
</tr>
<tr>
<td>18</td>
<td>1.326800</td>
<td>1.419522</td>
<td>0.245100</td>
<td>0.521500</td>
<td>0.210000</td>
<td>0.116100</td>
<td>0.211500</td>
<td>0.489900</td>
<td>0.255400</td>
<td>0.391600</td>
<td>0.419700</td>
<td>0.198800</td>
<td>0.421200</td>
<td>0.701400</td>
<td>0.501800</td>
<td>0.634200</td>
<td>0.226700</td>
<td>0.410100</td>
<td>0.154400</td>
<td>0.321400</td>
<td>0.105900</td>
<td>0.352300</td>
<td>0.236700</td>
<td>0.380400</td>
</tr>
<tr>
<td>19</td>
<td>1.158600</td>
<td>1.398764</td>
<td>0.253600</td>
<td>0.519200</td>
<td>0.213600</td>
<td>0.135200</td>
<td>0.207700</td>
<td>0.491900</td>
<td>0.257300</td>
<td>0.397300</td>
<td>0.428000</td>
<td>0.241400</td>
<td>0.401800</td>
<td>0.703500</td>
<td>0.509700</td>
<td>0.631100</td>
<td>0.236700</td>
<td>0.441800</td>
<td>0.155900</td>
<td>0.330800</td>
<td>0.128100</td>
<td>0.352300</td>
<td>0.237500</td>
<td>0.384000</td>
</tr>
<tr>
<td>20</td>
<td>1.158600</td>
<td>1.390591</td>
<td>0.248800</td>
<td>0.520200</td>
<td>0.216600</td>
<td>0.127500</td>
<td>0.211400</td>
<td>0.471900</td>
<td>0.258300</td>
<td>0.407000</td>
<td>0.429100</td>
<td>0.240300</td>
<td>0.407600</td>
<td>0.708500</td>
<td>0.505800</td>
<td>0.623400</td>
<td>0.235500</td>
<td>0.431600</td>
<td>0.150000</td>
<td>0.325000</td>
<td>0.125700</td>
<td>0.375400</td>
<td>0.227200</td>
<td>0.390200</td>
</tr>
<tr>
<td>21</td>
<td>1.158600</td>
<td>1.360608</td>
<td>0.262700</td>
<td>0.544800</td>
<td>0.222100</td>
<td>0.134700</td>
<td>0.230000</td>
<td>0.487500</td>
<td>0.269500</td>
<td>0.413300</td>
<td>0.436300</td>
<td>0.236200</td>
<td>0.419100</td>
<td>0.709300</td>
<td>0.514100</td>
<td>0.637400</td>
<td>0.257200</td>
<td>0.450600</td>
<td>0.165100</td>
<td>0.338400</td>
<td>0.139400</td>
<td>0.372300</td>
<td>0.237700</td>
<td>0.382700</td>
</tr>
<tr>
<td>22</td>
<td>1.158600</td>
<td>1.368296</td>
<td>0.262800</td>
<td>0.542400</td>
<td>0.236400</td>
<td>0.137400</td>
<td>0.228100</td>
<td>0.498500</td>
<td>0.266500</td>
<td>0.409000</td>
<td>0.433000</td>
<td>0.239900</td>
<td>0.418500</td>
<td>0.697500</td>
<td>0.520500</td>
<td>0.641000</td>
<td>0.257500</td>
<td>0.455700</td>
<td>0.162600</td>
<td>0.334800</td>
<td>0.140200</td>
<td>0.353800</td>
<td>0.233200</td>
<td>0.379600</td>
</tr>
<tr>
<td>23</td>
<td>1.158600</td>
<td>1.368176</td>
<td>0.264800</td>
<td>0.541100</td>
<td>0.233100</td>
<td>0.138200</td>
<td>0.223900</td>
<td>0.498700</td>
<td>0.272300</td>
<td>0.407400</td>
<td>0.434400</td>
<td>0.233100</td>
<td>0.418300</td>
<td>0.702000</td>
<td>0.524400</td>
<td>0.642300</td>
<td>0.262300</td>
<td>0.444300</td>
<td>0.159700</td>
<td>0.335300</td>
<td>0.140500</td>
<td>0.366200</td>
<td>0.236900</td>
<td>0.384000</td>
</tr>
<tr>
<td>24</td>
<td>1.049700</td>
<td>1.355271</td>
<td>0.269700</td>
<td>0.549200</td>
<td>0.239100</td>
<td>0.134700</td>
<td>0.229900</td>
<td>0.519200</td>
<td>0.274800</td>
<td>0.412700</td>
<td>0.437600</td>
<td>0.245400</td>
<td>0.417200</td>
<td>0.711200</td>
<td>0.523200</td>
<td>0.644100</td>
<td>0.272100</td>
<td>0.440500</td>
<td>0.166700</td>
<td>0.341500</td>
<td>0.137700</td>
<td>0.373800</td>
<td>0.249000</td>
<td>0.388000</td>
</tr>
<tr>
<td>25</td>
<td>1.049700</td>
<td>1.355180</td>
<td>0.272500</td>
<td>0.547900</td>
<td>0.243800</td>
<td>0.149700</td>
<td>0.229900</td>
<td>0.523100</td>
<td>0.272500</td>
<td>0.415700</td>
<td>0.442200</td>
<td>0.256200</td>
<td>0.420200</td>
<td>0.705800</td>
<td>0.523900</td>
<td>0.639600</td>
<td>0.271700</td>
<td>0.451900</td>
<td>0.166300</td>
<td>0.346900</td>
<td>0.153700</td>
<td>0.383100</td>
<td>0.247000</td>
<td>0.389300</td>
</tr>
<tr>
<td>26</td>
<td>1.049700</td>
<td>1.349337</td>
<td>0.275600</td>
<td>0.556300</td>
<td>0.246400</td>
<td>0.146700</td>
<td>0.234800</td>
<td>0.516300</td>
<td>0.274200</td>
<td>0.418300</td>
<td>0.440900</td>
<td>0.248700</td>
<td>0.418900</td>
<td>0.705800</td>
<td>0.523200</td>
<td>0.636500</td>
<td>0.274700</td>
<td>0.440500</td>
<td>0.172400</td>
<td>0.349100</td>
<td>0.155600</td>
<td>0.384600</td>
<td>0.252300</td>
<td>0.393800</td>
</tr>
<tr>
<td>27</td>
<td>1.049700</td>
<td>1.350782</td>
<td>0.275200</td>
<td>0.548700</td>
<td>0.246800</td>
<td>0.147300</td>
<td>0.236400</td>
<td>0.527200</td>
<td>0.280100</td>
<td>0.416200</td>
<td>0.442600</td>
<td>0.253400</td>
<td>0.424000</td>
<td>0.710300</td>
<td>0.526600</td>
<td>0.640100</td>
<td>0.273200</td>
<td>0.445600</td>
<td>0.167000</td>
<td>0.346900</td>
<td>0.160100</td>
<td>0.387700</td>
<td>0.249200</td>
<td>0.392900</td>
</tr>
<tr>
<td>28</td>
<td>1.049700</td>
<td>1.346533</td>
<td>0.277000</td>
<td>0.552800</td>
<td>0.252900</td>
<td>0.147400</td>
<td>0.240000</td>
<td>0.527600</td>
<td>0.280900</td>
<td>0.420900</td>
<td>0.444100</td>
<td>0.255500</td>
<td>0.424500</td>
<td>0.711200</td>
<td>0.530200</td>
<td>0.646800</td>
<td>0.277400</td>
<td>0.441800</td>
<td>0.170900</td>
<td>0.346900</td>
<td>0.156600</td>
<td>0.389200</td>
<td>0.249600</td>
<td>0.396000</td>
</tr>
<tr>
<td>29</td>
<td>0.993700</td>
<td>1.346575</td>
<td>0.277100</td>
<td>0.554800</td>
<td>0.252900</td>
<td>0.148400</td>
<td>0.239700</td>
<td>0.523600</td>
<td>0.278400</td>
<td>0.420000</td>
<td>0.443300</td>
<td>0.256300</td>
<td>0.424000</td>
<td>0.705600</td>
<td>0.529600</td>
<td>0.647300</td>
<td>0.273900</td>
<td>0.439200</td>
<td>0.174300</td>
<td>0.348700</td>
<td>0.157600</td>
<td>0.386200</td>
<td>0.250100</td>
<td>0.395100</td>
</tr>
<tr>
<td>30</td>
<td>0.993700</td>
<td>1.346446</td>
<td>0.277400</td>
<td>0.554700</td>
<td>0.252700</td>
<td>0.147900</td>
<td>0.240800</td>
<td>0.523600</td>
<td>0.278800</td>
<td>0.420400</td>
<td>0.443300</td>
<td>0.256100</td>
<td>0.424200</td>
<td>0.705500</td>
<td>0.530100</td>
<td>0.646800</td>
<td>0.275600</td>
<td>0.440500</td>
<td>0.174500</td>
<td>0.348700</td>
<td>0.157300</td>
<td>0.386200</td>
<td>0.249200</td>
<td>0.394200</td>
</tr>
</tbody>
</table><p>
If you have set `push_to_hub` to `True` in the `training_args`, the training checkpoints are pushed to the If you have set `push_to_hub` to `True` in the `training_args`, the training checkpoints are pushed to the
Hugging Face Hub. Upon training completion, push the final model to the Hub as well by calling the [`~transformers.Trainer.push_to_hub`] method. Hugging Face Hub. Upon training completion, push the final model to the Hub as well by calling the [`~transformers.Trainer.push_to_hub`] method.
...@@ -394,187 +1441,91 @@ Hugging Face Hub. Upon training completion, push the final model to the Hub as w ...@@ -394,187 +1441,91 @@ Hugging Face Hub. Upon training completion, push the final model to the Hub as w
``` ```
## Evaluate ## Evaluate
Object detection models are commonly evaluated with a set of <a href="https://cocodataset.org/#detection-eval">COCO-style metrics</a>.
You can use one of the existing metrics implementations, but here you'll use the one from `torchvision` to evaluate the final
model that you pushed to the Hub.
To use the `torchvision` evaluator, you'll need to prepare a ground truth COCO dataset. The API to build a COCO dataset
requires the data to be stored in a certain format, so you'll need to save images and annotations to disk first. Just like
when you prepared your data for training, the annotations from the `cppe5["test"]` need to be formatted. However, images
should stay as they are.
The evaluation step requires a bit of work, but it can be split in three major steps.
First, prepare the `cppe5["test"]` set: format the annotations and save the data to disk.
```py ```py
>>> import json >>> from pprint import pprint
>>> metrics = trainer.evaluate(eval_dataset=cppe5["test"], metric_key_prefix="test")
>>> # format annotations the same as for training, no need for data augmentation >>> pprint(metrics)
>>> def val_formatted_anns(image_id, objects): {'epoch': 30.0,
... annotations = [] 'test_loss': 1.0877351760864258,
... for i in range(0, len(objects["id"])): 'test_map': 0.4116,
... new_ann = { 'test_map_50': 0.741,
... "id": objects["id"][i], 'test_map_75': 0.3663,
... "category_id": objects["category"][i], 'test_map_Coverall': 0.5937,
... "iscrowd": 0, 'test_map_Face_Shield': 0.5863,
... "image_id": image_id, 'test_map_Gloves': 0.3416,
... "area": objects["area"][i], 'test_map_Goggles': 0.1468,
... "bbox": objects["bbox"][i], 'test_map_Mask': 0.3894,
... } 'test_map_large': 0.5637,
... annotations.append(new_ann) 'test_map_medium': 0.3257,
'test_map_small': 0.3589,
... return annotations 'test_mar_1': 0.323,
'test_mar_10': 0.5237,
'test_mar_100': 0.5587,
>>> # Save images and annotations into the files torchvision.datasets.CocoDetection expects 'test_mar_100_Coverall': 0.6756,
>>> def save_cppe5_annotation_file_images(cppe5): 'test_mar_100_Face_Shield': 0.7294,
... output_json = {} 'test_mar_100_Gloves': 0.4721,
... path_output_cppe5 = f"{os.getcwd()}/cppe5/" 'test_mar_100_Goggles': 0.4125,
'test_mar_100_Mask': 0.5038,
... if not os.path.exists(path_output_cppe5): 'test_mar_large': 0.7283,
... os.makedirs(path_output_cppe5) 'test_mar_medium': 0.4901,
'test_mar_small': 0.4469,
... path_anno = os.path.join(path_output_cppe5, "cppe5_ann.json") 'test_runtime': 1.6526,
... categories_json = [{"supercategory": "none", "id": id, "name": id2label[id]} for id in id2label] 'test_samples_per_second': 17.548,
... output_json["images"] = [] 'test_steps_per_second': 2.42}
... output_json["annotations"] = []
... for example in cppe5:
... ann = val_formatted_anns(example["image_id"], example["objects"])
... output_json["images"].append(
... {
... "id": example["image_id"],
... "width": example["image"].width,
... "height": example["image"].height,
... "file_name": f"{example['image_id']}.png",
... }
... )
... output_json["annotations"].extend(ann)
... output_json["categories"] = categories_json
... with open(path_anno, "w") as file:
... json.dump(output_json, file, ensure_ascii=False, indent=4)
... for im, img_id in zip(cppe5["image"], cppe5["image_id"]):
... path_img = os.path.join(path_output_cppe5, f"{img_id}.png")
... im.save(path_img)
... return path_output_cppe5, path_anno
``` ```
Next, prepare an instance of a `CocoDetection` class that can be used with `cocoevaluator`. These results can be further improved by adjusting the hyperparameters in [`TrainingArguments`]. Give it a go!
```py
>>> import torchvision
>>> class CocoDetection(torchvision.datasets.CocoDetection):
... def __init__(self, img_folder, image_processor, ann_file):
... super().__init__(img_folder, ann_file)
... self.image_processor = image_processor
... def __getitem__(self, idx):
... # read in PIL image and target in COCO format
... img, target = super(CocoDetection, self).__getitem__(idx)
... # preprocess image and target: converting target to DETR format,
... # resizing + normalization of both image and target)
... image_id = self.ids[idx]
... target = {"image_id": image_id, "annotations": target}
... encoding = self.image_processor(images=img, annotations=target, return_tensors="pt")
... pixel_values = encoding["pixel_values"].squeeze() # remove batch dimension
... target = encoding["labels"][0] # remove batch dimension
... return {"pixel_values": pixel_values, "labels": target} ## Inference
>>> image_processor = AutoImageProcessor.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> path_output_cppe5, path_anno = save_cppe5_annotation_file_images(cppe5["test"])
>>> test_ds_coco_format = CocoDetection(path_output_cppe5, image_processor, path_anno)
```
Finally, load the metrics and run the evaluation. Now that you have finetuned a model, evaluated it, and uploaded it to the Hugging Face Hub, you can use it for inference.
```py ```py
>>> import evaluate >>> import torch
>>> from tqdm import tqdm >>> import requests
>>> import numpy as np
>>> import albumentations as A
>>> model = AutoModelForObjectDetection.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5") >>> from PIL import Image
>>> module = evaluate.load("ybelkada/cocoevaluate", coco=test_ds_coco_format.coco) >>> from transformers import AutoImageProcessor, AutoModelForObjectDetection
>>> val_dataloader = torch.utils.data.DataLoader(
... test_ds_coco_format, batch_size=8, shuffle=False, num_workers=4, collate_fn=collate_fn
... )
>>> device = torch.device("cuda") if torch.cuda.is_available() else "cpu" >>> url = "https://images.pexels.com/photos/8413299/pexels-photo-8413299.jpeg?auto=compress&cs=tinysrgb&w=630&h=375&dpr=2"
>>> model.to(device) >>> image = Image.open(requests.get(url, stream=True).raw)
>>> with torch.no_grad(): >>> # Define transformations for inference
... for idx, batch in enumerate(tqdm(val_dataloader)): >>> resize_and_pad = A.Compose([
... pixel_values = batch["pixel_values"].to(device) ... A.LongestMaxSize(max_size=max_size),
... pixel_mask = batch["pixel_mask"].to(device) ... A.PadIfNeeded(max_size, max_size, border_mode=0, value=(128, 128, 128), position="top_left"),
... ])
... labels = [
... {k: v for k, v in t.items()} for t in batch["labels"]
... ] # these are in DETR format, resized + normalized
... # forward pass
... outputs = model(pixel_values=pixel_values, pixel_mask=pixel_mask)
... orig_target_sizes = torch.stack([target["orig_size"] for target in labels], dim=0)
... # convert outputs of model to Pascal VOC format (xmin, ymin, xmax, ymax)
... results = image_processor.post_process_object_detection(outputs, threshold=0, target_sizes=orig_target_sizes)
...
... module.add(prediction=results, reference=labels)
... del batch
>>> results = module.compute()
>>> print(results)
Accumulating evaluation results...
DONE (t=0.08s).
IoU metric: bbox
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.352
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.681
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.292
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.168
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.208
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.429
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.274
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.484
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.501
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.191
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.323
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.590
```
These results can be further improved by adjusting the hyperparameters in [`~transformers.TrainingArguments`]. Give it a go!
## Inference >>> # This one is for visualization with no padding
Now that you have finetuned a DETR model, evaluated it, and uploaded it to the Hugging Face Hub, you can use it for inference. >>> resize_only = A.Compose([
The simplest way to try out your finetuned model for inference is to use it in a [`Pipeline`]. Instantiate a pipeline ... A.LongestMaxSize(max_size=max_size),
for object detection with your model, and pass an image to it: ... ])
```
Load model and image processor from the Hugging Face Hub (skip to use already trained in this session):
```py ```py
>>> from transformers import pipeline >>> device = "cuda"
>>> import requests >>> model_repo = "qubvel-hf/detr_finetuned_cppe5"
>>> url = "https://i.imgur.com/2lnWoly.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> obj_detector = pipeline("object-detection", model="devonho/detr-resnet-50_finetuned_cppe5") >>> image_processor = AutoImageProcessor.from_pretrained(model_repo)
>>> obj_detector(image) >>> model = AutoModelForObjectDetection.from_pretrained(model_repo)
>>> model = model.to(device)
``` ```
You can also manually replicate the results of the pipeline if you'd like: And detect bounding boxes:
```py ```py
>>> image_processor = AutoImageProcessor.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5") >>> np_preprocessed_image = resize_and_pad(image=np.array(image))["image"]
>>> model = AutoModelForObjectDetection.from_pretrained("devonho/detr-resnet-50_finetuned_cppe5")
>>> with torch.no_grad(): >>> with torch.no_grad():
... inputs = image_processor(images=image, return_tensors="pt") ... inputs = image_processor(images=[np_preprocessed_image], return_tensors="pt")
... outputs = model(**inputs) ... outputs = model(inputs["pixel_values"].to(device))
... target_sizes = torch.tensor([image.size[::-1]]) ... target_sizes = torch.tensor([np_preprocessed_image.shape[:2]])
... results = image_processor.post_process_object_detection(outputs, threshold=0.5, target_sizes=target_sizes)[0] ... results = image_processor.post_process_object_detection(outputs, threshold=0.3, target_sizes=target_sizes)[0]
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]): >>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()] ... box = [round(i, 2) for i in box.tolist()]
...@@ -582,13 +1533,19 @@ You can also manually replicate the results of the pipeline if you'd like: ...@@ -582,13 +1533,19 @@ You can also manually replicate the results of the pipeline if you'd like:
... f"Detected {model.config.id2label[label.item()]} with confidence " ... f"Detected {model.config.id2label[label.item()]} with confidence "
... f"{round(score.item(), 3)} at location {box}" ... f"{round(score.item(), 3)} at location {box}"
... ) ... )
Detected Coverall with confidence 0.566 at location [1215.32, 147.38, 4401.81, 3227.08] Detected Gloves with confidence 0.683 at location [244.58, 124.33, 300.35, 185.13]
Detected Mask with confidence 0.584 at location [2449.06, 823.19, 3256.43, 1413.9] Detected Mask with confidence 0.517 at location [143.73, 64.58, 219.57, 125.89]
Detected Gloves with confidence 0.425 at location [179.15, 155.57, 262.4, 226.35]
Detected Coverall with confidence 0.407 at location [307.13, -1.18, 477.82, 318.06]
Detected Coverall with confidence 0.391 at location [68.61, 126.66, 309.03, 318.89]
``` ```
Let's plot the result: Let's plot the result:
```py ```py
>>> draw = ImageDraw.Draw(image) >>> resized_image = resize_only(image=np.array(image))["image"]
>>> resized_image = Image.fromarray(resized_image)
>>> draw = ImageDraw.Draw(resized_image)
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]): >>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()] ... box = [round(i, 2) for i in box.tolist()]
...@@ -596,9 +1553,9 @@ Let's plot the result: ...@@ -596,9 +1553,9 @@ Let's plot the result:
... draw.rectangle((x, y, x2, y2), outline="red", width=1) ... draw.rectangle((x, y, x2, y2), outline="red", width=1)
... draw.text((x, y), model.config.id2label[label.item()], fill="white") ... draw.text((x, y), model.config.id2label[label.item()], fill="white")
>>> image >>> resized_image
``` ```
<div class="flex justify-center"> <div class="flex justify-center">
<img src="https://i.imgur.com/4QZnf9A.png" alt="Object detection result on a new image"/> <img src="https://i.imgur.com/oDUqD0K.png" alt="Object detection result on a new image"/>
</div> </div>
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