Simplified Inference

[glenn-jocher]

Replacement for #1045

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

This PR introduces adjustments to confidence thresholds and the NMS (Non-Maximum Suppression) process, alongside code clean-up and modularization.

📊 Key Changes

• Changed the default object confidence threshold from 0.4 to 0.25 and the IOU threshold for NMS from 0.5 to 0.45 in detect.py.
• Removed an unused NMS import and unnecessary model evaluation call in hubconf.py.
• Added autoShape class in common.py for a robust model wrapper to support various input forms and removed redundant imports.
• Integrated autoShape functionality within yolo.py and allowed the addition/removal of NMS from YOLO models.
• Removed outdated code and imports in various files such as sotabench.py and datasets.py.

🎯 Purpose & Impact

• Lower confidence and NMS thresholds lead to potentially more detections at the cost of increased false positives.
• The introduction of autoShape enables the model to handle different input types more seamlessly, enhancing user-friendliness and the potential for integration into different pipelines without requiring pre-formatting by the user.
• Code modularization and cleanup improve the maintainability of the code base, making it easier to develop and deploy.
• These changes seek to improve both detection performance and user experience when working with YOLO models from the Ultralytics repository.

[glenn-jocher]

This PR allows for using YOLOv5 independent of the this repo, and for automatic handling of input formats. Now you can pass in PIL objects directly, numpy array images, cv2 images, or torch inputs. You can pass in a batch as a list, or as a torch batched input. Options are:

    def forward(self, x, size=640, augment=False, profile=False):
        # supports inference from various sources. For height=720, width=1280, RGB images example inputs are:
        #   opencv:     x = cv2.imread('image.jpg')[:,:,::-1]  # HWC BGR to RGB x(720,1280,3)
        #   PIL:        x = Image.open('image.jpg')  # HWC x(720,1280,3)
        #   numpy:      x = np.zeros((720,1280,3))  # HWC
        #   torch:      x = torch.zeros(16,3,720,1280)  # BCHW
        #   multiple:   x = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...]  # list of images

Input

import cv2
import numpy as np
from PIL import Image, ImageDraw

from models.experimental import attempt_load

# Model
# model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
model = attempt_load('yolov5s.pt').autoshape()  # add autoshape wrapper

# Images
img1 = Image.open('inference/images/zidane.jpg')  # PIL
img2 = cv2.imread('inference/images/zidane.jpg')[:, :, ::-1]  # opencv (BGR to RGB)
img3 = np.zeros((640, 1280, 3))  # numpy
imgs = [img1, img2, img3]

# Inference
prediction = model(imgs, size=640)  # includes NMS

# Plot
for i, img in enumerate(imgs):
    print('\nImage %g/%g: %s ' % (i + 1, len(imgs), img.shape), end='')
    img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img  # from np
    if prediction[i] is not None:  # is not None
        for *box, conf, cls in prediction[i]:  # [xy1, xy2], confidence, class
            print('%s %.2f, ' % (model.names[int(cls)], conf), end='')  # label
            ImageDraw.Draw(img).rectangle(box, width=3)  # plot
    img.save('results%g.jpg' % i)  # save

Output

Downloading https://github.com/ultralytics/yolov5/releases/download/v3.0/yolov5s.pt to yolov5s.pt...
100%
14.5M/14.5M [00:00<00:00, 64.5MB/s]

Fusing layers... 
Adding autoShape... 

Image 1/3: (720, 1280, 3) person 0.87, person 0.80, tie 0.78, tie 0.28, 
Image 2/3: (720, 1280, 3) person 0.87, person 0.80, tie 0.78, tie 0.28, 
Image 3/3: (640, 1280, 3) 

[glenn-jocher]

Torch Hub Example

This repo NOT required

import cv2
import numpy as np
import torch
from PIL import Image, ImageDraw

# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True, force_reload=True).fuse().eval()
model = model.autoshape()  # add autoshape wrapper IMPORTANT

# Image
torch.hub.download_url_to_file('https://github.com/raw/ultralytics/yolov5/master/inference/images/zidane.jpg', 'zidane.jpg')
img1 = Image.open('zidane.jpg')  # PIL
img2 = cv2.imread('zidane.jpg')[:, :, ::-1]  # opencv (BGR to RGB)
img3 = np.zeros((640, 1280, 3))  # numpy
imgs = [img1, img2, img3]  # batched inference

# Inference
prediction = model(imgs, size=640)  # includes NMS

# Plot
for i, img in enumerate(imgs):
    print('\nImage %g/%g: %s ' % (i + 1, len(imgs), img.shape), end='')
    img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img  # from np
    if prediction[i] is not None:  # is not None
        for *box, conf, cls in prediction[i]:  # [xy1, xy2], confidence, class
            print('class %g %.2f, ' % (cls, conf), end='')  # label
            ImageDraw.Draw(img).rectangle(box, width=3)  # plot
    img.save('results%g.jpg' % i)  # save

[dagap]

@glenn-jocher

When I run this, I get the following error:

if prediction[i] is not None:  # is not None
    TypeError: 'Detections' object is not subscriptable

[glenn-jocher]

@dagap I'll inline the latest here in case anyone arrives here.

Official page: https://pytorch.org/hub/ultralytics_yolov5/

Load From PyTorch Hub (NEW FORMAT)

To load YOLOv5 from PyTorch Hub for inference with PIL, OpenCV, Numpy or PyTorch inputs:

import cv2
import torch
from PIL import Image

# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True).fuse().autoshape()  # for PIL/cv2/np inputs and NMS

# Images
for f in ['zidane.jpg', 'bus.jpg']:  # download 2 images
    print(f'Downloading {f}...')
    torch.hub.download_url_to_file('https://github.com/ultralytics/yolov5/releases/download/v1.0/' + f, f)
img1 = Image.open('zidane.jpg')  # PIL image
img2 = cv2.imread('bus.jpg')[:, :, ::-1]  # OpenCV image (BGR to RGB)
imgs = [img1, img2]  # batched list of images

# Inference
results = model(imgs, size=640)  # includes NMS

# Results
results.print()  # print results to screen
results.show()  # display results
results.save()  # save as results1.jpg, results2.jpg... etc.

# Data
print('\n', results.xyxy[0])  # print img1 predictions
#          x1 (pixels)  y1 (pixels)  x2 (pixels)  y2 (pixels)   confidence        class
# tensor([[7.47613e+02, 4.01168e+01, 1.14978e+03, 7.12016e+02, 8.71210e-01, 0.00000e+00],
#         [1.17464e+02, 1.96875e+02, 1.00145e+03, 7.11802e+02, 8.08795e-01, 0.00000e+00],
#         [4.23969e+02, 4.30401e+02, 5.16833e+02, 7.20000e+02, 7.77376e-01, 2.70000e+01],
#         [9.81310e+02, 3.10712e+02, 1.03111e+03, 4.19273e+02, 2.86850e-01, 2.70000e+01]])

[pk-1196]

@glenn-jocher Hi , i have trained yolov5 model using yolov5s.pt and saved the weights as best.pt
Now how to use above weight i.e. best.pt on the stacked image to get inference speed any changes in above code . so that my input image will be of (2,3,640,640)

[glenn-jocher]

@pk-1196 see PyTorch Hub tutorial for inference directions:

YOLOv5 Tutorials

• Train Custom Data  🚀 RECOMMENDED
• Tips for Best Training Results  ☘️ RECOMMENDED
• Weights & Biases Logging  🌟 NEW
• Supervisely Ecosystem  🌟 NEW
• Multi-GPU Training
• PyTorch Hub  ⭐ NEW
• TorchScript, ONNX, CoreML Export 🚀
• Test-Time Augmentation (TTA)
• Model Ensembling
• Model Pruning/Sparsity
• Hyperparameter Evolution
• Transfer Learning with Frozen Layers  ⭐ NEW
• TensorRT Deployment

[ZhangChen0212]

@dagap I'll inline the latest here in case anyone arrives here.

Official page: https://pytorch.org/hub/ultralytics_yolov5/

Load From PyTorch Hub (NEW FORMAT)

To load YOLOv5 from PyTorch Hub for inference with PIL, OpenCV, Numpy or PyTorch inputs:

import cv2
import torch
from PIL import Image

# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True).fuse().autoshape()  # for PIL/cv2/np inputs and NMS

# Images
for f in ['zidane.jpg', 'bus.jpg']:  # download 2 images
    print(f'Downloading {f}...')
    torch.hub.download_url_to_file('https://github.com/ultralytics/yolov5/releases/download/v1.0/' + f, f)
img1 = Image.open('zidane.jpg')  # PIL image
img2 = cv2.imread('bus.jpg')[:, :, ::-1]  # OpenCV image (BGR to RGB)
imgs = [img1, img2]  # batched list of images

# Inference
results = model(imgs, size=640)  # includes NMS

# Results
results.print()  # print results to screen
results.show()  # display results
results.save()  # save as results1.jpg, results2.jpg... etc.

# Data
print('\n', results.xyxy[0])  # print img1 predictions
#          x1 (pixels)  y1 (pixels)  x2 (pixels)  y2 (pixels)   confidence        class
# tensor([[7.47613e+02, 4.01168e+01, 1.14978e+03, 7.12016e+02, 8.71210e-01, 0.00000e+00],
#         [1.17464e+02, 1.96875e+02, 1.00145e+03, 7.11802e+02, 8.08795e-01, 0.00000e+00],
#         [4.23969e+02, 4.30401e+02, 5.16833e+02, 7.20000e+02, 7.77376e-01, 2.70000e+01],
#         [9.81310e+02, 3.10712e+02, 1.03111e+03, 4.19273e+02, 2.86850e-01, 2.70000e+01]])

Hi there! When I run this code, I get an error:

Traceback (most recent call last):
  File "/home/chen/Desktop/test.py", line 6, in <module>
    model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True).fuse().autoshape()  # for PIL/cv2/np inputs and NMS
  File "/usr/local/lib/python3.9/dist-packages/torch/nn/modules/module.py", line 947, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'AutoShape' object has no attribute 'fuse'