Accessing Torch Model

[parthnatekar]

Hi,

How do I access the base torch model when using the high-level APIs

pose_model = init_pose_model(pose_config, pose_checkpoint)

pose_model is not an instance of torch.nn.Module, is there any way to get this?

Edit: I see that pose_model is indeed an instance of torch.nn.Module. Is there an easy way to get the model summary? I want to check the layer output shapes of the model, but when I try to use torchsummary with an input shape I get TypeError: forward() missing 1 required positional argument: 'img_metas' which I assume is because the forward pass takes some additional metadata.

[ly015]

You can use the pose_model.forward_dummy instead of pose_model.forward to perform a simple network forward without the need for extra metadata.

[parthnatekar]

Hi,

Where exactly is model.forward_dummy defined? Does it return the output of each module as an iterator?

I can see that it returns a tuple of tensors, but I'm not sure what these are.

[parthnatekar]

Just to clarify, I'm using mmdet for bounding box prediction and then feeding that to a pose model.

When I run forward_dummy on the mmdet model I get an output of shape ((B, N, 2), (B, N, 4)), where B is the batch size. I'm assuming N is the number of object proposals.

However when I run inference_detector(pose_model, image_frame) I get an output of shape ((B, 2), (B, 4)).

I want to access the top 'k' predicted bounding boxes for each image - how can I get these?

[ly015]

Sorry for the late reply. forward_dummy simply performs network forward while inference_detector performs the whole inference pipeline including model forward and post-processings, and that is why they give different outputs. An example of forward_dummy in mmpose can be found here. If you are using mmdet models, you can also consider raising an issue in mmdet for expert feedback.