christophreich1996 / maxvit Goto Github PK

Hello, thank you for your excellent work! Do you have the pytorch pretrained weight?

hi!
I have a question about the line 400 in maxvit.py.
there is a skip-connection in block attention and grid attention, so may be we should use 'output = output + self.drop_path(self.mlp(self.norm_2(output)))', rather than '*'
Looking forward to your reply, thanks!

Hi @ChristophReich1996, thanks for implementing MaxViT!

I am just wondering whether the grid partition function has been done right.

In lines, you have implemented the grid partition as: windows = input.view(B, C, H // grid_size[0], grid_size[0], W // grid_size[1], grid_size[1]), which seems to be as same to window partition. I'm thinking that the grid partition should be fixing the number of windows instead of setting the window size. It should look like:

windows = input.view(B, C, grid_size[0], H // grid_size[0],  grid_size[1], W // grid_size[1])

Am I right?

I am checking lucidrains's implementation here which seems to indicate this above meaning. Please let me know if I was wrong on this. Thanks~

Thanks to the authors for sharing Maxvit open source, I really enjoyed this project and studied it for a few days. However, I didn't understand this part of the work on the Grid Partition. In my opinion, it looks almost the same as SWIN V1, so how does it accomplish the grid operation shown below? Looking forward to your advice, thank you

Hi good job! I have a question that How can I convert the TF version pretained weight checkpoint file to pytorch version ?

Hello dear, Thanks for your reproducing this paper, and I want to do this model too.In your code, I find that grid operation seems some problem .I think that your operation refers to block operation rather than global grid operator. How to do global grid operator? And Im waiting for your continuous releasing the code. Thank you.

In this line of code
output = self.main_path(input)：line 69, in forward
Error
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking arugment for argument weight in method wrapper_cudnn_batch_norm)

Do you have a trained pre-trained model? And could you share it please？

Hi Chrstoph, thanks for code skeleton for MaxViT paper.

I checked the number of parameters of your code and paper, and both seems to be difference. MaxViT tiny give 24M parameter in this github repo, whereas paper reports 31M. Can you please help me out?

Also I believe the main_path in MBConv block should be like :-

`

   self.main_path = nn.Sequential(

    norm_layer(in_channels),

nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=(1, 1)),   # not in original code

norm_layer(out_channels),

act_layer(),

     DepthwiseSeparableConv(in_chs=out_channels, out_chs=out_channels, stride=2 if downscale else 1,
                               act_layer=act_layer, norm_layer=norm_layer, drop_path_rate=drop_path),

      SqueezeExcite(in_chs=out_channels, rd_ratio=0.25),

      nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=(1, 1))
    )

`

Here you missed first conv2d of kernel 1x1 in your code.

Thanks,
Saarthak

Thank you for your work.
I import the model, and push it to gpu, but the error happen:
RuntimeError: Tensor for argument #2 'weight' is on CPU, but expected it to be on GPU (while checking arguments for cudnn_batch_norm)
Lookiing forward for your reply!