Validation_loader = torch.utils.data.DataLoader(validation_data, batch_size=batch_size, shuffle=True) ###### define constant######## input_channels = 3 hidden_size = 64 max_epochs = 100 lr = 3e-4 beta = 0 #0.000000001#0.00000001#0.0001#0.0000001#0.000001#0.00001 #######network################ #epoch=39 #M='/big_disk/akrami/git_repos_new/lesion-detector/VAE_9.5.2019/Brats_results' ##########load low res net########## G = VAE_Generator(input_channels, hidden_size).cuda() #load_model(epoch,G.encoder, G.decoder,LM) opt_enc = optim.Adam(G.parameters(), lr=lr) fixed_noise = Variable(torch.randn(batch_size, hidden_size)).cuda() data = next(iter(Validation_loader)) fixed_batch = Variable(data).cuda() #######losss################# def MSE_loss(Y, X): msk = torch.tensor(X > 1e-6).float() ret = ((X - Y)**2) * msk ret = torch.sum(ret, 1) return ret
############################################ ########## intilaize parameters########## # define constant input_channels = 3 hidden_size = 64 max_epochs = 200 lr = 3e-4 beta = 0 device = 'cuda' ######################################### epoch = 99 LM = '/big_disk/akrami/git_repos_new/ImagePTE/src/Lesion Detection/models/RVAE_final_1' ##########load low res net########## G = VAE_Generator(input_channels, hidden_size).cuda() load_model(epoch, G.encoder, G.decoder, LM) ##########define beta loss########## def MSE_loss(Y, X): msk = torch.tensor(X > 1e-6).float() ret = ((X - Y)**2) * msk return ret def BMSE_loss(Y, X, beta, sigma, Dim): term1 = -((1 + beta) / beta) K1 = 1 / pow((2 * math.pi * (sigma**2)), (beta * Dim / 2)) term2 = MSE_loss(Y, X)