Open wukurua opened 4 months ago
the same question here...
I tried to get the FLOPs and #param for this RGBPose model. Here it is Input shape: {'rgb': [1, 3, 8, 224, 224], 'pose': [1, 17, 32, 56, 56]} Flops: 56.98 GFLOPs Params: 36.15 M
I used the code repository https://github.com/kennymckormick/pyskl and use the get_flops script and made modification in flops_counter script file as follows:
rgb_batch = torch.ones(()).new_empty( (1, *[1,3,8,224, 224]), dtype=next(flops_model.parameters()).dtype, device=next(flops_model.parameters()).device) pose_batch = torch.ones(()).new_empty( (1, *[1,17,32,56, 56]), dtype=next(flops_model.parameters()).dtype, device=next(flops_model.parameters()).device) .... _ = flops_model(rgb_batch,pose_batch)
What is the problem this feature will solve?
I want to test GFlops for RGBPoseConv3D (
configs\skeleton\posec3d\rgbpose_conv3d\rgbpose_conv3d.py
), but the current code doesn't seem to support MMRecognizer3D . Hope to have a section for testing the Multi-modal 3D recognizer model framework intools/analysis_tools/get_flops.py
. By the way, If you know the params and GFlops of RGBPoseConv3D, or its input_shape, can anyone tell me about it?What is the feature?
get flops of the MMRecognizer3D
What alternatives have you considered?
No response