blade6570 / photographicimagesynthesiswithcascadedrefinementnetworks-pytorch

Photographic Image Synthesis with Cascaded Refinement Networks - Pytorch Implementation

photographicimagesynthesiswithcascadedrefinementnetworks-pytorch's Issues

pretrained VGG19 and CRN weights are no longer able to download

Hello, I am trying to reproduce your work with the readme instruction. I found I currently do not have access to you pretrained models in google drive, could you release it?

Why use `nn.AvgPool2d` to get downsampled images?

It seems inconsistent with the official implementation.

downsampled=tf.image.resize_area(label,(sp//2,sp),align_corners=False)

Information on training process

Hi there,

I'm just wondering if it is possible to get some information on what hardware you used to train the model, as well as how long you took? I am currently writing my own version of the CRN, however I am encountering extremely long training times and massive memory requirements, and so I thought is worthwhile to find out more about other implementations and their requirments.

Kind regards

Several questions

1. Can we use recursive_generator similar with the official code?
2. LayerNorm is in here http://pytorch.org/docs/master/nn.html#layernorm.(Seems in this week v0.4 will be released)

The way that you get D and D_m seems computation heavy.

def recursive_img(label,res): #Resulution may refers to the final image output i.e. 256x512 or 512x1024
     dim=512 if res>=128 else 1024
#    #M_low will start from 4x8 to resx2*res
     if res == 4:
         downsampled = label #torch.unsqueeze(torch.from_numpy(label).float().permute(2,0,1), dim=0)
     else:
         max1=nn.AvgPool2d(kernel_size=2, padding=0, stride=2)
         downsampled=max1(label)
         img = recursive_img(downsampled, res//2)
         
     global D
     global count
     global D_m

     D.insert(count, downsampled)
     D_m.insert(count, dim)
     count+=1
     return downsampled  

Why not directly assign each D_i and D_m_i with specific values.