versions_log.txt

# version 0 -> no attention
# version 1 -> self attention added (using nn.MultiheadAttention)
# version 2 -> Changed model architecture and used Silu instead of ReLU
# version 3 -> Used hugging face diffusers.UNet2DModel  UNet2DModel( in_channels= 3 , out_channels = 3, block_out_channels = (32,64 , 128 , 512) , norm_num_groups=8).to(device)
# version 4 -> Normal model in Version 3 .. changed dataset to Lsun bridges
# version 5 -> Added EMA
# Version 6 -> used fashion mnist 32*32 dataset
# version 7 -> used fashion mnist 32*32 dataset with the diffusers UNet2DModel ..lr -> 2e-3
# Version 8 -> fixed the stupid ass bug where the generation size is set to 128
# version 9 -> changed the model to the SmallUnetWithEmb( img_channels=1). Ds is still the same
# Version 10 -> Used the celeb ds with 64*64 images also lr from 2e-3 to 1e-3
# version 11 -> Changed the max pool in the down module to conv2d with 3*3 kernel size and padding 1 and the upsample to convtranspose2d
# version 12 -> attention layer added to the 3rd layer in both the decoder and the encoder
# version 13 -> ciphar10 and removed the 3rd attention layer and added classes
# version 14 -> fasion mnist, same as before