conv3x3 | vary fast 3x3 convolution on cpu | GPU library

U-Net family of models (such as the MultiResUNet model above) follow an encoder-decoder architecture. Encoder is a down-sampling path with feature extraction whereas the decoder an upsampling one. Feature maps from encoder are concatenated at the decoder through skip-connections. These feature maps are concatenated at the last axis, the 'channel' axis (considering the features to be having dimensions [batch_size, height, width, channels]). Now, for the features to be concatenated at any axis ('channel' axis, in our case), the dimensions at all the other axes must match.

In the above model architecture, there are 3 downsampling/max-pooling operations being performed (through MaxPooling2D)in the encoder path. At the decoder path 3 upsampling/transpose-conv operations are performed, aiming to restore the image back to the full dimension. However, for the concatenations (through skip-connections) to happen, the downsampled and upsampled feature dimensions of height, width & batch_size should remain identical at every "level" of the model. I'll illustrate this with the examples you mentioned in the question:

1st case : Input dimensions (128,128,3) : 128 -> 64 -> 32 -> 16 -> 32 -> 64 -> 128

2nd case: Input dimensions (36,36,3) : 36 -> 18 -> 9 -> 4 -> 8 -> 16 -> 32

In the 2nd case, when the height and width of feature map reaches 9 in the encoder path, further downsampling leads to a dimension change (loss) that cannot be regained in the decoder while upsampling. Hence, it throws an error due to inability to concatenate feature maps of dimensions [(None, 8, 8, 128)] & [(None, 9, 9, 128)].

In general, for a simple encoder-decoder model (with skip-connections) having 'n' downsampling (MaxPooling2D) layers, the input dimension must be a multiple of 2^n to be able to concatenate the model's encoder features at the decoder. In this case n=3, hence the input must be a multiple of 8 to not run into these dimension mismatch errors.

Hope this helps! :)

The contributor posted the same question on the PGI User Forums where I've answered. (See: https://www.pgroup.com/userforum/viewtopic.php?f=4&t=7614). The topic question is incorrect in that the inner loops are not getting parallelized nor are the cause of the issue.

The problem here is that the code has a race condition on the shared "output_im" pointer. My suggested solution is to compute a per thread offset into the array rather than trying to manipulate the pointer itself.

I finally found out the cause. The error messages with the form of :

While copying the parameter named "xxx.weight", whose dimensions in the model are torch.Size([yyy]) and whose dimensions in the checkpoint are torch.Size([yyy]).

are actually generic messages, only returned when an exception has occured while copying the parameters in question.

Pytorch developers could easily add the actual exception args into this spurious yet unhelpful message, so it could actually help better debug the issue at hand. Anyway, looking at the exception which was by the way :

There is an implementation of laddernet in Keras available here : https://github.com/divamgupta/ladder_network_keras/blob/master/ladder_net.py. Consider this as a starting point, I have used at a point this repository successfully.