# increase mini batch counter
mini_batch_count += 1
# determine if GPU training is enabled
if (torch.backends.cudnn.version() != None) and (USE_CUDA == True):
# convert mini batch to torch variable
mini_batch_torch = torch.cuda.FloatTensor(mini_batch_data)
else:
# convert mini batch to torch variable
mini_batch_torch = torch.FloatTensor(mini_batch_data)
# reset the networks gradients
encoder_train.zero_grad()
decoder_train.zero_grad()
discriminator_train.zero_grad()
# =================== reconstruction phase =====================
# run autoencoder encoding - decoding
z_sample = encoder_train(mini_batch_torch)
mini_batch_reconstruction = decoder_train(z_sample)
# split input date to numerical and categorical part
batch_cat = mini_batch_torch[:, :ori_dataset_categ_transformed.
shape[1]]
batch_num = mini_batch_torch[:,
ori_dataset_categ_transformed.shape[1]:]
tic = time.time()
for i in range(num_captions):
image_id = shuffled_images[i]
image = dataloader.get_image(image_id)
image = image.unsqueeze(0)
if torch.cuda.is_available():
image = Variable(image).cuda()
caption = torch.cuda.LongTensor(shuffled_captions[i])
else:
image = Variable(image)
caption = torch.LongTensor(shuffled_captions[i])
caption_train = caption[:-1]
encoder.zero_grad()
decoder.zero_grad()
encod_out = encoder(image)
decoder_output = decoder(encod_out, caption_train)
loss = criterion(decoder_output, caption)
loss.backward()
optimizer.step()
loss_list.append(loss)
toc = time.time()
avg_loss = torch.mean(torch.Tensor(loss_list))
