Closed powermano closed 1 year ago
We greatly appreciate your providing this onnx example and we will include it in the tutorial. We also very welcome your participation in building this open-source project !
We greatly appreciate your providing this onnx example and we will include it in the tutorial. We also very welcome your participation in building this open-source project !
Thanks, i'd like to participate in building this open-source project. I will push some KD methods for DETR families in the future.
We greatly appreciate your providing this onnx example and we will include it in the tutorial. We also very welcome your participation in building this open-source project !
Thanks, i'd like to participate in building this open-source project. I will push some KD methods for DETR families in the future.
Sure, we will really appreciate it! @powermano
BTW, we've pinned this issue, which will help more people when they need onnx output~
As we have already pinned this issue, we're going to close it. Feel free to reopen it if necessary.
Hey @powermano, Thank you for this recipe! I am currently thinking to follow this recipe for DINO. However, I'm wondering, did you achieve lower inference speeds? Did you convert the onnx even to TensorRT (if this is possible?) Thank you in Advance!
Hey @powermano, Thank you for this recipe! I am currently thinking to follow this recipe for DINO. However, I'm wondering, did you achieve lower inference speeds? Did you convert the onnx even to TensorRT (if this is possible?) Thank you in Advance!
I have converted the onnx model to openvino(https://docs.openvino.ai/2023.1/home.html) and used the cpu for inference. The inference speed of onnxruntime and openvino using a single process is almost the same. I didn't test the speed of pytorch because it uses multiple processes.
Does this work for Mask DINO?
Hi! I tried to convert MASK DINO model to onnx using these scripts but it just failed when it came to the forward function. I don't know how I should handle with this. Is this possible to provide scripts for MASK DINO. I would really appriciate!
Taking DN-DETR as example
1 add a dn_detr_onnx.py, you only need to change the
forward
function.2 add new_file dn_detr_r50_onnx.py, change
DNDETR
toDNDETR_ONNX
.3 add new_file
dn_detr_r50_50ep_onnx.py
4 add new_file
torch2onnx.py
in the tools foder.5 run with command
python tools/torch2onnx.py --config-file projects/dn_detr/configs/dn_detr_r50_50ep_onnx.py --eval-only train.init_checkpoint=$<YOUR_MODEL_PATH>
It worked for me. The torch results and onnx results are the same.