search

sahilkhose / StackGAN-BERT

PyTorch implementation of StackGAN paper using BERT embeddings
MIT License
12 stars 4 forks source link

Generator 1 conv2d (64, 64, 64) -> (3, 64, 64) #3

Closed sahilkhose closed 3 years ago

sahilkhose commented 3 years ago

https://github.com/sahilkhose/ganctober/blob/37aa750ab95d381bb44a6654eb20174d7c98dacc/src/layers.py#L68 There is no mention of conv2d in the paper They mention using upsampling blocks and the last upsampling block does not have a batch norm and relu activation We can just set output channels of self.up4 to 3, to solve this problem? I think this conv2d is used because of the keras implementation

abhirajtiwari commented 3 years ago

Would upsample be learning any feature? because it has no weights, so running a convolution to reduce all the channels seems like a better option I suppose. Similarly for #4

sahilkhose commented 3 years ago

This argument is incorrect because the _upsample block that we are using has Conv2d, so the learning is obviously present.

sahilkhose commented 3 years ago

Resolving this issue due to lack of clarity in the paper. The implementation details are not clearly stated in the paper. Or are considered obvious. We will adapt the code in accordance to what practices are used in vision and GAN architectures.