HYLcool
commented
1 year ago Hi, @ahmedtalbi
Thanks for your suggestion! That's really a good one. We will add this feature in the following updates.
If you need to convert DAMO-YOLO to ONNX models in a hurry, you can add a new parameter dynamic_axes to the export function here:
https://github.com/tinyvision/DAMO-YOLO/blob/b12c2647e614ec3446108ffd3690dba519f7b5db/tools/converter.py#L231-L239
For example, for non-end2end conversion of DAMO-YOLO-S, the new parameter is like the below:
dynamic_axes = {
args.input: {0: 'batch_size'},
args.output: {0: 'batch_size'},
'848': {0: 'batch_size'},
},| Here in this dictionary, the 3rd item (which is also the 2nd output) is actually the bbox prediction of DAMO-YOLO, which doesn't have a specific tensor name and is numbered by onnx for now. Therefore, for different models, we list the corresponding number names here: | Model | name of the 2nd output |
|---|---|---|
| DAMO-YOLO-N | '1452' | |
| DAMO-YOLO-T | '806' | |
| DAMO-YOLO-S | '848' | |
| DAMO-YOLO-M | '1171' |
You can replace the key of the 3rd item with the corresponding name when converting different models.
On the other hand, if you need to convert end2end models to ONNX models, you also need to modify the output keys of dynamic axes to ['num_dets', 'det_boxes', 'det_scores', 'det_classes'].
BTW, we would be very grateful if you would like to submit a PR. And you can contact us anytime if you have further issues.
ahmedtalbi
Search before asking
Description
The idea is to make DAMO-YOLO work with dynamic axes, like other models, i.e. YOLOv5. This will allow to have a different input shapes and batch size.
I tried to do it myself but with no success. It seems the converter is not able to convert the latest Mul operation here is a snippet from Netron. the rest of the network seems to be converted correctly.
I added the dynamic_axes as argument to
torch.onnx._export()as followstorch.onnx._export( model, dummy_input, onnx_name, input_names=[args.input], output_names=['num_dets', 'det_boxes', 'det_scores', 'det_classes'] if args.end2end else [args.output], dynamic_axes={ 'images': [0, 2, 3], # shape(1,3,640,640) 'output': [0, 1], 'bbox_preds': [0, 1], } if args.dynamic else None, opset_version=args.opset, )would be thankful for any help.Use case
No response
Additional
No response
Are you willing to submit a PR?