RuntimeError running the model
Morplson opened this issue · 5 comments
First of all, love your video.
Trying to run the unconditional model gives me this funny error:
File "f:\Coding Projects\ai\diffusinon-test\modules.py", line 99, in forward
return x + emb
RuntimeError: The size of tensor a (192) must match the size of tensor b (12) at non-singleton dimension 0
Do you know how to fix it?
Hey,
did you change anything in the code? If so, could you let me know what exactly was changed?
Hello, I have the same error by changing the input size from 64 to 128.
It appears for some reason to change the batch size in the down and up function, I cant figure out why...
RuntimeError: Caught RuntimeError in replica 0 on device 0.
Original Traceback (most recent call last):
File "/local/home/alibaym/anaconda3/envs/mlenv/lib/python3.10/site-packages/torch/nn/parallel/parallel_apply.py", line 64, in _worker
output = module(*input, **kwargs)
File "/local/home/alibaym/anaconda3/envs/mlenv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1357, in _call_impl
return forward_call(*input, **kwargs)
File "/local/home/alibaym/codes/diffusion_models/modules.py", line 170, in forward
x3 = self.down2(x2, t)
File "/local/home/alibaym/anaconda3/envs/mlenv/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1357, in _call_impl
return forward_call(*input, **kwargs)
File "/local/home/alibaym/codes/diffusion_models/modules.py", line 99, in forward
return x + emb
RuntimeError: The size of tensor a (12) must match the size of tensor b (3) at non-singleton dimension 0
FYI I solved the issue:
Main problem is that image size is hardcoded in the Unet definition for the Self attention module. The line
x = x.view(-1, self.channels, self.size * self.size).swapaxes(1, 2)
Makes the batch size change if the size was not the same. These self attention sizes should be dependant on the image size.
FYI I solved the issue: Main problem is that image size is hardcoded in the Unet definition for the Self attention module. The line x = x.view(-1, self.channels, self.size * self.size).swapaxes(1, 2) Makes the batch size change if the size was not the same. These self attention sizes should be dependant on the image size.
Encountering the same problem on a custom dataset...
Can you give more details about what you had to change in order to get it to work ?
In the UNet initialization, you'll need to specify SelfAttention image size as the image size of your dataset instead of hardcoded numbers like 32, 16 etc. The way I've changed it is following.
`class UNet(nn.Module):
def init(self, c_in=3, c_out=3, time_dim=256, image_size = 256, device="cuda"):
super().init()
self.image_size = image_size
self.sa1 = SelfAttention(128, self.image_size // 2)
self.down2 = Down(128, 256)
self.sa2 = SelfAttention(256, self.image_size // 4)
self.down3 = Down(256, 256)
self.sa3 = SelfAttention(256, self.image_size // 8)
........
self.up1 = Up(512, 128)
self.sa4 = SelfAttention(128, self.image_size // 4)
self.up2 = Up(256, 64)
self.sa5 = SelfAttention(64, self.image_size // 2)
self.up3 = Up(128, 64)
self.sa6 = SelfAttention(64, self.image_size)
self.outc = nn.Conv2d(64, c_out, kernel_size=1)
`