ONNX transfer problem

[zeckireck]

Hello, our script does not support the case that convtranpose group is greater than 1. How can we turn it into the case that each group is equal to 1?

[luminxu]

If you do not want to use deconvolutional layers with the group number greater than 1, you may try the backbone searched by ViPNAS (e.g. ViPNAS_ResNet following s_vipnas_res50_coco_256x192.py), and apply the TopdownHeatmapSimpleHead as the keypoint_head following SimpleBaseline. Make sure that the in_channels of keypoint_head should be consistent with the output_channels of backbone (e.g. 608 for S-ViPNAS-Res50).

[luminxu]

Also, if you still want to use ViPNASHeatmapSimpleHead, you can set num_deconv_groups to be (1, 1, 1). No matter which one you choose, please retrain the new model and verify whether the performance satisfies your needs.

[zeckireck]

are these factors (num_deconv_layers, num_deconv_filters, num_deconv_kernels) need to be modified?

[zeckireck]

or only change num_deconv_groups to be (1, 1, 1)

[luminxu]

You may try to change num_deconv_groups only. As it is no longer the keypoint head searched by ViPNAS, you should verify the accuracy under this setting. If necessary, you can also change other arguments.

[zeckireck]

Why are there 2 inferences here? Just to infer for once, can't it? Will there be a big impact on accuracy and confidence?https://github.com/luminxu/ViPNAS/blob/56a0630efee9d36595c0f5d3268553f273d35ad3/mmpose/models/detectors/top_down.py#L184

[luminxu]

Flip Test is a common test-time technique for more robust prediction, which empirically introduces around 1% AP improvement. If you care more about the inference speed, you can turn it off in the configuration.

[zeckireck]

As for resnet50, do you know where this operator is in the code? Our hardware only supports the case where the divisor of the div operator is constant。

[luminxu]

I cannot recognize the exact operators according to the figures. Also, I am not sure when the problem happens, e.g., instantiation (I guess so), initialization, or inference. According to your description, you may check this line.

[zeckireck]

Is there any way to set out_channels to a fixed value?https://github.com/luminxu/ViPNAS/blob/main/mmpose/models/backbones/vipnas_resnet.py#L115

[zeckireck]

What is this attention used for?https://github.com/luminxu/ViPNAS/blob/main/mmpose/models/backbones/vipnas_resnet.py#L115is it necessary？

[luminxu]

Whether to use attention module is included in search spaces and is searched for better pose estimation performance. Please refer to our paper for more details.

[zeckireck]

https://github.com/luminxu/ViPNAS/blob/main/mmpose/models/backbones/vipnas_resnet.py#L115. Why do we need 16/outchannels here? only to use 1/16,cant it?