Refactor Detect() anchors for ONNX <> OpenCV DNN compatibility
#4833
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
👋 Hello @SamFC10, thank you for submitting a 🚀 PR! To allow your work to be integrated as seamlessly as possible, we advise you to:
- ✅ Verify your PR is up-to-date with origin/master. If your PR is behind origin/master an automatic GitHub actions rebase may be attempted by including the /rebase command in a comment body, or by running the following code, replacing 'feature' with the name of your local branch:
git remote add upstream https://github.com/ultralytics/yolov5.git
git fetch upstream
git checkout feature # <----- replace 'feature' with local branch name
git rebase upstream/master
git push -u origin -f- ✅ Verify all Continuous Integration (CI) checks are passing.
- ✅ Reduce changes to the absolute minimum required for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." -Bruce Lee
|
@SamFC10 thanks for the PR! CI checks are failing, can you take a look at this please? |
|
@glenn-jocher Fixed all CI build failures. |
|
@SamFC10 thanks for the CI fixes! I ran CI on GPU in Colab and saw the error below. To reproduce, run Setup cell and then CI Checks cell in Appendix: 1. Setup!git clone https://github.com/ultralytics/yolov5 # clone repo
%cd yolov5
%pip install -qr requirements.txt # install dependencies
import torch
from IPython.display import Image, clear_output # to display images
clear_output()
print(f"Setup complete. Using torch {torch.__version__} ({torch.cuda.get_device_properties(0).name if torch.cuda.is_available() else 'CPU'})")2. CI Checks%%shell
export PYTHONPATH="$PWD" # to run *.py. files in subdirectories
rm -rf runs # remove runs/
for m in yolov5s; do # models
python train.py --weights $m.pt --epochs 3 --img 320 --device 0 # train pretrained
python train.py --weights '' --cfg $m.yaml --epochs 3 --img 320 --device 0 # train scratch
for d in 0 cpu; do # devices
python detect.py --weights $m.pt --device $d # detect official
python detect.py --weights runs/train/exp/weights/best.pt --device $d # detect custom
python val.py --weights $m.pt --device $d # val official
python val.py --weights runs/train/exp/weights/best.pt --device $d # val custom
done
python hubconf.py # hub
python models/yolo.py --cfg $m.yaml # build PyTorch model
python models/tf.py --weights $m.pt # build TensorFlow model
python export.py --img 128 --batch 1 --weights $m.pt --include torchscript onnx # export
doneErrorTraceback (most recent call last):
File "train.py", line 611, in <module>
main(opt)
File "train.py", line 509, in main
train(opt.hyp, opt, device, callbacks)
File "train.py", line 261, in train
compute_loss = ComputeLoss(model) # init loss class
File "/content/yolov5/utils/loss.py", line 114, in __init__
self.anchors = det.anchors / det.stride.view(-1, 1, 1)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! |
I was able to reproduce the errors. Latest commit fixes them and CI checks are now passed on Colab GPU runtime |
|
@SamFC10 awesome thank you! Unfortunately we don't have a seamless CPU-GPU CI in place, so our current workflow involves manually running Colab CI before merges. |
glenn-jocher
changed the title
Detect() anchors for ONNX <> OpenCV DNN compatibility
|
@SamFC10 a problem came up in testing. Detect, Val, Export operate correctly. Train does not throw an error, but produces near zero results in comparison to master in the Colab notebook Train section. I think when --weights are being loaded for training the anchors are being overwritten or reordered incorrectly. # Train YOLOv5s on COCO128 for 3 epochs
!python train.py --img 640 --batch 16 --epochs 3 --data coco128.yaml --weights yolov5s.pt --cacheThis PR Epoch gpu_mem box obj cls labels img_size
0/2 3.55G 0.09968 0.06903 0.02277 166 640: 100% 8/8 [00:04<00:00, 1.62it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:01<00:00, 3.97it/s]
all 128 929 0.24 0.00101 9.33e-05 1.87e-05
Epoch gpu_mem box obj cls labels img_size
1/2 4.4G 0.09012 0.06828 0.02285 205 640: 100% 8/8 [00:02<00:00, 3.90it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:00<00:00, 4.48it/s]
all 128 929 0.24 0.00101 9.67e-05 1.19e-05
Epoch gpu_mem box obj cls labels img_size
2/2 4.4G 0.07513 0.06287 0.02503 163 640: 100% 8/8 [00:02<00:00, 3.78it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:05<00:00, 1.25s/it]
all 128 929 0.127 0.00101 7.51e-05 3.07e-05Master Epoch gpu_mem box obj cls labels img_size
0/2 3.55G 0.04438 0.07064 0.01989 166 640: 100% 8/8 [00:04<00:00, 1.72it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:01<00:00, 3.79it/s]
all 128 929 0.682 0.578 0.65 0.426
Epoch gpu_mem box obj cls labels img_size
1/2 4.4G 0.04641 0.06915 0.01959 205 640: 100% 8/8 [00:02<00:00, 3.83it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:00<00:00, 4.26it/s]
all 128 929 0.683 0.583 0.657 0.426
Epoch gpu_mem box obj cls labels img_size
2/2 4.4G 0.04365 0.06366 0.02201 163 640: 100% 8/8 [00:02<00:00, 3.85it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 4/4 [00:03<00:00, 1.15it/s]
all 128 929 0.694 0.575 0.662 0.432 |
|
@glenn-jocher Was able to reproduce this behaviour
Quite likely. In train.py when pre-trained weights are used, |
|
Latest commit solves the issue of poor training accuracy with `--weights option This PR |
|
@SamFC10 I looked at this again a couple days ago. It seems like the PR works well as a standalone branch (i.e. to train and detect), but for master branch we need to ensure that any updates are backwards compatible with all of the existing YOLOv5 models out in the wild from earlier releases. It seemed like this might cause problems in those cases, but I wasn't sure. It's hard to know without a more comprehensive set of tests. If I have some more time this week I will try to revisit. |
Backwards compatibility was needed because this PR changed the definition of Looking at the changes made, I think the current solution is not elegant/ideal. I have another branch locally which fixes the problem and does not modify Marking this PR as draft for now. |
|
@SamFC10 I did a suite of tests yesterday and everything seemed to work well: train new from --weights I have not tested export yet. But based on the above tests backwards compatibility seemed to work correctly so far. |
OpenCV DNN Inference Instructions# Export to ONNX
python export.py --weights yolov5s.pt --include onnx --simplify
# Inference
python detect.py --weights yolov5s.onnx # ONNX Runtime inference
# -- or --
python detect.py --weights yolov5s.onnx --dnn # OpenCV DNN inference |
|
Hello @glenn-jocher @SamFC10 and anyone who may have the same issue ! After pulling the repo today I could not use an old weight file (trained about a year ago) to perform inference. I was getting this error : Using I do not have enough data to post a new issue. Nevertheless, I wanted to report it and to propose a temporary solution for whoever ecounters the same error. So, I solved the issue by loading the weights in training script (as if I were continuing training) and saving them with new ckpt before the epochs loop. After this, no more error during detection. Thank you very much for your work ! |
|
@adrienanton thanks for the bug report! I think there may be built-in functionality to do what you need. If you run python detect.py --weights /tmp/model/last_95.pt --update |
|
@glenn-jocher Thanks for the response ! I do not think |
|
@adrienanton ok got it. The workaround then is just to retrain your model. You'll get better results and the error will be resolved. |
|
@glenn-jocher @SamFC10 I'm reading the YOLOV5 code, just curious why do we need to do |
|
@Tyler-D loss is computed in grid space (i.e. 0-19) rather than pixel space (0-639) |
Got it. Thanks ! |
…ralytics#4833) * refactor anchors and anchor_grid in Detect Layer * fix CI failures by adding compatibility * fix tf failure * fix different devices errors * Cleanup * fix anchors overwriting issue * better refactoring * Remove self.anchor_grid shape check (redundant with self.grid check) Also PEP8 / 120 line width * Convert _make_grid() from static to dynamic method * Remove anchor_grid.to(device) clone() should already clone to same device as self.anchors * fix different devices error Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
|
Dear, I try to remove P3 and P5 detection and still tp P4, and I do require a change in the Neck, and everything becomes well and works. When I try to delete C5 from the feature Extraction Backbone and do modifications in the neck, the error happened nc: 80 # number of classes [30,61, 62,45, 59,119] # P4/16 [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 head: [-1, 1, Conv, [256, 1, 1]], [-1, 1, Conv, [256, 3, 2]], [[20], 1, Detect, [nc, anchors]], # Detect( P4) |
|
@alkhalisy it seems like the error is occurring because you are deleting C5 from the feature extraction backbone but not updating the head accordingly. Since the head is referencing the deleted C5, it causes an index out of range error when trying to access it. Make sure to update the head according to the changes made in the backbone to resolve this error. |
Continuation of #4811.
self.anchor_gridwas causing a problem as it is a pytorch tensor, so the memory is already allocated and shape defined. Due to this, expandinganchor_gridto perform element-wise multiplication resulted in added nodes in exported onnx model. So, I have changedself.anchor_gridto a list of tensors instead (similar toself.grid) and refactored some other code related toanchor_gridsto make this change workable.New onnx models (notice the new shapes in Add and Mul node):
The exported onnx model now works with opencv dnn module.
(with --simplify). For onnx model without --simplify option, the opset version must be set to 11 to import in opencv (ONNX changed behaviour of unsqueeze node in opset 13, need a small fix in opencv to import with opset 13 as well).Fixed by opencv/opencv#20713Note:
This change breaks inference for models trained before this PRAdded compatibility with previously trained models🛠️ PR Summary
Made with ❤️ by Ultralytics Actions
🌟 Summary
Enhanced anchor handling in YOLOv5's detection layers for better compatibility and maintainability.
📊 Key Changes
🎯 Purpose & Impact