ABCNet deployment

[shuangyichen]

Is there anyway to export model to caffe model like using detectron2 tools caffe2_converter.py?

[shuangyichen]

@blueardour Hi,

Can onnx/export_model_to_onnx.py work for ABCNet? Looking forward to your reply. Thank you!

[blueardour]

Thanks for reporting the requirement. I will pay some effort on the converting.

[shuangyichen]

@blueardour HI, I am tring to convert by myself because the assignment is very urgent. And I am following your work. So could plz give me some hints or explain your pipeline about exporting onnx? Thank you!

[blueardour]

Hi, @shuangyichen

For the onnx converting script, one might consider the following aspects:

• What version of onnx op set is going to support? As far as I know some operations in pytorch are not well supported for the onnx, such as the NMS, RIO-pooling/align. Thus, we should first print the whole model and then decide which sub-modules we would export to the onnx.

Example:

OneStageDetector(
  (backbone): FPN(
    (fpn_lateral3): Conv2d(
      512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
      (norm): NaiveSyncBatchNorm(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (quant_activation): quantization()
      (quant_weight): quantization()
    )
    (fpn_output3): Conv2d(
      256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
      (norm): NaiveSyncBatchNorm(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (quant_activation): quantization()
      (quant_weight): quantization()
    )
    (fpn_lateral4): Conv2d(
      1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False
      (norm): NaiveSyncBatchNorm(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (quant_activation): quantization()
      (quant_weight): quantization()
    )
    (fpn_output4): Conv2d(
      256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False
      (norm): NaiveSyncBatchNorm(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)

• For the modules which cannot be exported to onnx, such as the FCOS module(the FCOS head module inside the FCOS module can be exported), we might hack the wrapper as I did in the current script.

[blueardour]

just update some progress.

• As Group Norm is not well supported in onnx, I'am training the Full-BN version of ABC-net.

• In the meanwhile, I am developing the onnx converting code based on early saved model file from above training.

[shuangyichen]

Thank you for your contribution!

[shuangyichen]

Hello @blueardour , Do you know how to deal with BezierAlign before torch.onnx.export? Looking forward to your reply.

[blueardour]

hi, @shuangyichen

From my perspective, I suggest not exporting the BezierAlign into onnx. Instead, just export other code with normal operators such as conv/fc/bn/relu/interp. When inference on embeded platform or TenorRT, one might re-implement the BezierAlign code on the dedicated platform.

[shuangyichen]

@blueardour So we just use BezierAlign as the boundary to divide the two parts, simply skip BezierAlign. And convert the two parts into onnx respectively?

[blueardour]

Yes. Though it involves with some extra effort, it somehow provides a feasible way.

[jdavidd]

@shuangyichen Hello, did you manage to write the script to conver ABCnet model to onnx?

[jdavidd]

@blueardour Hello, did you worked on the script to convert Abcnet model to onnx?

[evanchien]

@blueardour Were you able to get the conversion work? Would really love to see how good it will work after deployment. Thanks! :)

[ucsky]

Hi @blueardour I'm in the same situation. Do you have some feedback about converting the model to onnx?

[Single430]

@shuangyichen @blueardour blue Could you tell me what progress abcnet to onnx now?

[pustar]

@shuangyichen @blueardour blue Could you tell me what progress abcnet to onnx now?