Pyramid Pooling Transformer for Scene Understanding

Hello, thanks for open source!

I use mmseg, and load weight from image classification result, it warns: WARNING - The model and loaded state dict do not match exactly missing keys in source state_dict: backbone.head.weight, backbone.head.bias unexpected key in source state_dict: cls_token, ln1.bias, ln1.weight, layers.0.ln1.bias, layers.0.ln1.weight, layers.0.ln2.bias, layers.0.ln2.weight, layers.0.ffn.layers.0.0.bias, layers.0.ffn.layers.0.0.weight, layers.0.ffn.layers.1.bias, layers.0.ffn.layers.1.weight, layers.0.attn.attn.out_proj.bias, layers.0.attn.attn.out_proj.weight, layers.0.attn.attn.in_proj_bias, layers.0.attn.attn.in_proj_weight, layers.1.ln1.bias, layers.1.ln1.weight, layers.1.ln2.bias, layers.1.ln2.weight, layers.1.ffn.layers.0.0.bias, layers.1.ffn.layers.0.0.weight, layers.1.ffn.layers.1.bias, layers.1.ffn.layers.1.weight, layers.1.attn.attn.out_proj.bias, layers.1.attn.attn.out_proj.weight ...... And the experimental results are terrible as the experiments initialize weight with random.

So I load weight from ADE20K result, it work and warns: WARNING - The model and loaded state dict do not match exactly missing keys in source state_dict: backbone.head.weight, backbone.head.bias And the result is similar to the result you offer.

Which weight should I load? ImageNet1K or ADE20K? Or should I modify the keys of weight in ImageNet1K to adapt the key in segmentation?

Hello,

I have some questions about your ablation studies of pyramid pooling. Could you detail about your baseline version in Table 9? First, you say that you replace P-MHSA with an MHSA with a single pooling operation, what is the detail about single pooling operation? Ex: Pooling Ratios? Second, do you compared your method with original MHSA?

When I replaced the PVT trunk with P2T in my code, I encountered an error ：
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [16, 512, 3, 3]], which is output 0 of AdaptiveAvgPool2DBackward, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

Hi @yuhuan-wu , thank you for share the code of this excellent work! Have you trained P2T on ImageNet-22K dataset or any further plan to do it? If so, could you please share the pretrained model on ImageNet-22k?

Thank you.