Note
Go to the end to download the full example code
Object Detection Tutorial#
In this tutorial, you will learn how to validate your object detection model using deepchecks test suites. You can read more about the different checks and suites for computer vision use cases at the examples section.
If you just want to see the output of this tutorial, jump to the observing the results section.
An object detection tasks usually consist of two parts:
Object Localization, where the model predicts the location of an object in the image,
Object Classification, where the model predicts the class of the detected object.
The common output of an object detection model is a list of bounding boxes around the objects, and their classes.
# Before we start, if you don't have deepchecks vision package installed yet, run:
import sys
!{sys.executable} -m pip install "deepchecks[vision]" --quiet --upgrade # --user
# or install using pip from your python environment
Defining the data and model#
Note
In this tutorial, we use the pytorch to create the dataset and model. To see how this can be done using tensorflow or other frameworks, please visit the The Vision Data Class guide.
Load Data#
The model in this tutorial is used to detect tomatoes in images. The model is trained on a dataset consisted of 895 images of tomatoes, with bounding box annotations provided in PASCAL VOC format. All annotations belong to a single class: tomato.
Note
The dataset is available at the following link: https://www.kaggle.com/andrewmvd/tomato-detection
We thank the authors of the dataset for providing the dataset.
import os
import numpy as np
import torch
from torch.utils.data import DataLoader, Dataset
import albumentations as A
from albumentations.pytorch import ToTensorV2
from PIL import Image
import xml.etree.ElementTree as ET
import urllib.request
import zipfile
url = 'https://figshare.com/ndownloader/files/34488599'
urllib.request.urlretrieve(url, './tomato-detection.zip')
with zipfile.ZipFile('./tomato-detection.zip', 'r') as zip_ref:
zip_ref.extractall('.')
class TomatoDataset(Dataset):
def __init__(self, root, transforms):
self.root = root
self.transforms = transforms
self.images = list(sorted(os.listdir(os.path.join(root, 'images'))))
self.annotations = list(sorted(os.listdir(os.path.join(root, 'annotations'))))
def __getitem__(self, idx):
img_path = os.path.join(self.root, "images", self.images[idx])
ann_path = os.path.join(self.root, "annotations", self.annotations[idx])
img = Image.open(img_path).convert("RGB")
bboxes, labels = [], []
with open(ann_path, 'r') as f:
root = ET.parse(f).getroot()
for obj in root.iter('object'):
difficult = obj.find('difficult').text
if int(difficult) == 1:
continue
cls_id = 1
xmlbox = obj.find('bndbox')
b = [float(xmlbox.find('xmin').text), float(xmlbox.find('ymin').text), float(xmlbox.find('xmax').text),
float(xmlbox.find('ymax').text)]
bboxes.append(b)
labels.append(cls_id)
bboxes = torch.as_tensor(np.array(bboxes), dtype=torch.float32)
labels = torch.as_tensor(np.array(labels), dtype=torch.int64)
if self.transforms is not None:
res = self.transforms(image=np.array(img), bboxes=bboxes, class_labels=labels)
target = {
'boxes': [torch.Tensor(x) for x in res['bboxes']],
'labels': res['class_labels']
}
img = res['image']
return img, target
def __len__(self):
return len(self.images)
data_transforms = A.Compose([
A.Resize(height=256, width=256),
A.CenterCrop(height=224, width=224),
A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2(),
], bbox_params=A.BboxParams(format='pascal_voc', label_fields=['class_labels']))
dataset = TomatoDataset(root='./tomato-detection/data', transforms=data_transforms)
train_dataset, test_dataset = torch.utils.data.random_split(dataset,
[int(len(dataset)*0.9), len(dataset)-int(len(dataset)*0.9)],
generator=torch.Generator().manual_seed(42))
test_dataset.transforms = A.Compose([ToTensorV2()])
Visualize the dataset#
Let’s see how our data looks like.
print(f'Number of training images: {len(train_dataset)}')
print(f'Number of test images: {len(test_dataset)}')
print(f'Example output of an image shape: {train_dataset[0][0].shape}')
print(f'Example output of a label: {train_dataset[0][1]}')
Number of training images: 805
Number of test images: 90
Example output of an image shape: torch.Size([3, 224, 224])
Example output of a label: {'boxes': [tensor([ 0.00000, 75.13600, 39.68000, 165.75999]), tensor([ 0.00000, 0.00000, 94.08000, 93.56800])], 'labels': [tensor(1), tensor(1)]}
Downloading a Pre-trained Model#
In this tutorial, we will download a pre-trained SSDlite model and a MobileNetV3 Large backbone from the official PyTorch repository. For more details, please refer to the official documentation.
After downloading the model, we will fine-tune it for our particular classes. We will do it by replacing the pre-trained head with a new one that matches our needs.
from functools import partial
from torch import nn
import torchvision
from torchvision.models.detection import _utils as det_utils
from torchvision.models.detection.ssdlite import SSDLiteClassificationHead
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model = torchvision.models.detection.ssdlite320_mobilenet_v3_large(pretrained=True)
in_channels = det_utils.retrieve_out_channels(model.backbone, (320, 320))
num_anchors = model.anchor_generator.num_anchors_per_location()
norm_layer = partial(nn.BatchNorm2d, eps=0.001, momentum=0.03)
model.head.classification_head = SSDLiteClassificationHead(in_channels, num_anchors, 2, norm_layer)
_ = model.to(device)
Downloading: "https://download.pytorch.org/models/ssdlite320_mobilenet_v3_large_coco-a79551df.pth" to /home/runner/.cache/torch/hub/checkpoints/ssdlite320_mobilenet_v3_large_coco-a79551df.pth
0%| | 0.00/13.4M [00:00<?, ?B/s]
100%|██████████| 13.4M/13.4M [00:00<00:00, 180MB/s]
Loading Pre-trained Weights#
For this tutorial we will not include the training code itself, but will download and load pre-trained weights.
model.load_state_dict(torch.load('./tomato-detection/ssd_model.pth'))
_ = model.eval()
Validating the Model With Deepchecks#
Now, after we have the training data, test data and the model, we can validate the model with deepchecks test suites.
Implementing the VisionData class#
The checks in the package validate the model & data by calculating various quantities over the data, labels and
predictions. In order to do that, those must be in a pre-defined format, according to the task type.
In the following example we’re using pytorch. To see an implementation of this in tensorflow, please refer to
The Vision Data Class guide.
For pytorch, we will use our DataLoader, but we’ll create a new collate function for it, that transforms the batch to
the correct format. Then, we’ll create a deepchecks.vision.vision_data.vision_data.VisionData object, that will hold the data loader.
To learn more about the expected format please visit Supported Tasks and Formats.
First, we will create some functions that transform our batch to the correct format of images, labels and predictions:
def get_untransformed_images(original_images):
"""
Convert a batch of data to images in the expected format. The expected format is an iterable of images,
where each image is a numpy array of shape (height, width, channels). The numbers in the array should be in the
range [0, 255] in a uint8 format.
"""
inp = torch.stack(list(original_images)).cpu().detach().numpy().transpose((0, 2, 3, 1))
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
# Un-normalize the images
inp = std * inp + mean
inp = np.clip(inp, 0, 1)
return inp * 255
def transform_labels_to_cxywh(original_labels):
"""
Convert a batch of data to labels in the expected format. The expected format is an iterator of arrays, each array
corresponding to a sample. Each array element is in a shape of [B, 5], where B is the number of bboxes
in the image, and each bounding box is in the structure of [class_id, x, y, w, h].
"""
label = []
for annotation in original_labels:
if len(annotation["boxes"]):
bbox = torch.stack(annotation["boxes"])
# Convert the Pascal VOC xyxy format to xywh format
bbox[:, 2:] = bbox[:, 2:] - bbox[:, :2]
# The label shape is [class_id, x, y, w, h]
label.append(
torch.concat([torch.stack(annotation["labels"]).reshape((-1, 1)), bbox], dim=1)
)
else:
# If it's an empty image, we need to add an empty label
label.append(torch.tensor([]))
return label
def infer_on_images(original_images):
"""
Returns the predictions for a batch of data. The expected format is an iterator of arrays, each array
corresponding to a sample. Each array element is in a shape of [B, 6], where B is the number of bboxes in the
predictions, and each bounding box is in the structure of [x, y, w, h, score, class_id].
Note that model and device here are global variables, and are defined in the previous code block, as the collate
function cannot recieve other arguments than the batch.
"""
nm_thrs = 0.2
score_thrs = 0.7
imgs = list(img.to(device) for img in original_images)
# Getting the predictions of the model on the batch
with torch.no_grad():
preds = model(imgs)
processed_pred = []
for pred in preds:
# Performoing non-maximum suppression on the detections
keep_boxes = torchvision.ops.nms(pred['boxes'], pred['scores'], nm_thrs)
score_filter = pred['scores'][keep_boxes] > score_thrs
# get the filtered result
test_boxes = pred['boxes'][keep_boxes][score_filter].reshape((-1, 4))
test_boxes[:, 2:] = test_boxes[:, 2:] - test_boxes[:, :2] # xyxy to xywh
test_labels = pred['labels'][keep_boxes][score_filter]
test_scores = pred['scores'][keep_boxes][score_filter]
processed_pred.append(
torch.concat([test_boxes, test_scores.reshape((-1, 1)), test_labels.reshape((-1, 1))], dim=1))
return processed_pred
Now we’ll create the collate function that will be used by the DataLoader. In pytorch, the collate function is used to transform the output batch to any custom format, and we’ll use that in order to transform the batch to the correct format for the checks.
from deepchecks.vision.vision_data import BatchOutputFormat
def deepchecks_collate_fn(batch) -> BatchOutputFormat:
"""Return a batch of images, labels and predictions in the deepchecks format."""
# batch received as iterable of tuples of (image, label) and transformed to tuple of iterables of images and labels:
batch = tuple(zip(*batch))
images = get_untransformed_images(batch[0])
labels = transform_labels_to_cxywh(batch[1])
predictions = infer_on_images(batch[0])
return BatchOutputFormat(images=images, labels=labels, predictions=predictions)
We have a single label here, which is the tomato class The label_map is a dictionary that maps the class id to the class name, for display purposes.
LABEL_MAP = {
1: 'Tomato'
}
Now that we have our updated collate function, we can recreate the dataloader in the deepchecks format, and use it to create a VisionData object:
from deepchecks.vision.vision_data import VisionData
train_loader = DataLoader(train_dataset, batch_size=64, collate_fn=deepchecks_collate_fn)
test_loader = DataLoader(test_dataset, batch_size=64, collate_fn=deepchecks_collate_fn)
training_data = VisionData(batch_loader=train_loader, task_type='object_detection', label_map=LABEL_MAP)
test_data = VisionData(batch_loader=test_loader, task_type='object_detection', label_map=LABEL_MAP)
torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at ../aten/src/ATen/native/TensorShape.cpp:2157.)
Making sure our data is in the correct format:#
The VisionData object automatically validates your data format and will alert you if there is a problem.
However, you can also manually view your images and labels to make sure they are in the correct format by using
the head function to conveniently visualize your data:
<deepchecks.core.serialization.html_display.HtmlDisplayableResult object at 0x7f90d5d6db80>
Running Deepchecks’ suite on our data and model!#
Now that we have defined the task class, we can validate the model with the deepchecks’ model evaluation suite. This can be done with this simple few lines of code:
from deepchecks.vision.suites import model_evaluation
suite = model_evaluation()
result = suite.run(training_data, test_data)
Processing Batches:Train:
| | 0/1 [Time: 00:00]
Processing Batches:Train:
|█████| 1/1 [Time: 00:30]
Processing Batches:Train:
|█████| 1/1 [Time: 00:30]
Computing Single Dataset Checks Train:
| | 0/4 [Time: 00:00]
Computing Single Dataset Checks Train:
|█▎ | 1/4 [Time: 00:00, Check=Mean Average Precision Report]
Computing Single Dataset Checks Train:
|██▌ | 2/4 [Time: 00:00, Check=Mean Average Recall Report]
Computing Single Dataset Checks Train:
|█████| 4/4 [Time: 00:00, Check=Weak Segments Performance]
Computing Single Dataset Checks Train:
|█████| 4/4 [Time: 00:00, Check=Weak Segments Performance]
Processing Batches:Test:
| | 0/1 [Time: 00:00]
Processing Batches:Test:
|█████| 1/1 [Time: 00:03]
Processing Batches:Test:
|█████| 1/1 [Time: 00:03]
Computing Single Dataset Checks Test:
| | 0/4 [Time: 00:00]
Computing Single Dataset Checks Test:
|█████| 4/4 [Time: 00:00, Check=Weak Segments Performance]
Computing Single Dataset Checks Test:
|█████| 4/4 [Time: 00:00, Check=Weak Segments Performance]
Computing Train Test Checks:
| | 0/2 [Time: 00:00]
Computing Train Test Checks:
| | 0/2 [Time: 00:00, Check=Class Performance]
Computing Train Test Checks:
|██▌ | 1/2 [Time: 00:00, Check=Class Performance]
Computing Train Test Checks:
|██▌ | 1/2 [Time: 00:00, Check=Prediction Drift]
Computing Train Test Checks:
|█████| 2/2 [Time: 00:00, Check=Prediction Drift]
Computing Train Test Checks:
|█████| 2/2 [Time: 00:00, Check=Prediction Drift]
We also have suites for:
data integrity
- validating a single dataset and
train test validation -
validating the dataset split
Observing the results:#
The results can be saved as a html file with the following code:
result.save_as_html('output.html')
'output (3).html'
Or, if working inside a notebook, the output can be displayed directly by simply printing the result object:
result
We can see that our model does not perform well, as can be seen in the “Class Performance” check under the “Didn’t Pass” section of the suite results. This is because the model was trained on a different dataset, and the model was not trained to detect tomatoes. Moreover, we can see that lowering the IoU threshold could have fixed this a bit (as can be seen in the “Mean Average Precision Report” Check), but would still keep the overall precision low. Moreover, under the “Passed” section, we can see that our drift checks have passed, which means that the distribution of the predictions on the training and test data is similar, and the issue is not there but in the model itself.
Total running time of the script: (0 minutes 51.047 seconds)