MAE-pytorch
MAE-pytorch copied to clipboard
pengzhiliang
MAE prediction visualization code
Thank you for your contribution. I wonder if you plan to release the mask prediction visualization code?
Unfortunately, our current visualization code also has some bugs, I will try to solve it!
I have implemented the visualization code here, can I submit a merge request?
Of course. Thank you for your contributions.
And can you provide some visualization results here?
Of course, can you provide pre-trained models and test images? Because the current model is mainly based on a custom datset.
I have uploaded the weight to google drive, please see latest readme.txt.
Hi, I was wondering how to perform inference and run the full encoder-decoder network on a complete, unmasked image? In other words, after it is trained, how would I call the model such that it encodes a complete image with no masks, and then reconstructs the original image using the decoder?
Hello, @avitrost , maybe there are some bugs now when the mask is always 0.
But if you really want to observe the performance when MAE is used as a pure auto-encoder, you should have a change:
For the encoder:
you can directly let x_vis=x in this line.
And for the inputs to decoder:
x_vis = self.encoder_to_decoder(x_vis) # [B, N_vis, C_d]
B, N, C = x_vis.shape
expand_pos_embed = self.pos_embed.expand(B, -1, -1).type_as(x).to(x.device).clone().detach()
pos_emd_vis = expand_pos_embed[~mask].reshape(B, -1, C)
pos_emd_mask = expand_pos_embed[mask].reshape(B, -1, C)
x_full = torch.cat([x_vis + pos_emd_vis, self.mask_token + pos_emd_mask], dim=1)
x = self.decoder(x_full, pos_emd_mask.shape[1]) # [B, N_mask, 3 * 16 * 16]
It needs to be changed to:
x_vis = self.encoder_to_decoder(x_vis) # [B, N, C_d]
expand_pos_embed = self.pos_embed.expand(B, -1, -1).type_as(x).to(x.device).clone().detach()
x = self.decoder(x_vis+expand_pos_embed , x_vis.shape[1]) # [B, N, 3 * 16 * 16]
Maybe there will be some other bugs, you can have a debug. Hope this can help you!
expand_pos_embed = self.pos_embed
Thank you for your contribution! I have a question about this change. Why this change does not update in the latest code?
Hello, I would like to ask if the weight file loaded visually is the pre-trained weight file or the fine-tuned weight file? Error when I load the fine-tuned weight file:
