shuffle=True,
drop_last=True))
# initialize summary writer
writer = SummaryWriter()
sigma_p_inv, det_p = setup_pz(NUM_FEA, FEA_DIM, FEA)
# creating copies of encoder-decoder objects for style transfer visualization during training
encoder_test = Encoder()
encoder_test.apply(weights_init)
decoder_test = Decoder()
decoder_test.apply(weights_init)
encoder_test.eval()
decoder_test.eval()
if (CUDA):
encoder_test.cuda()
decoder_test.cuda()
lowest_loss = float('inf')
for epoch in range(START_EPOCH, END_EPOCH):
epoch_loss = 0
for iteration in range(len(dataset) // BATCH_SIZE):
# load a batch of videos
X_in = next(loader).float().cuda()
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True))
encoder = Encoder()
encoder.apply(weights_init)
decoder = Decoder()
decoder.apply(weights_init)
encoder.load_state_dict(
torch.load(os.path.join('checkpoints/', ENCODER_SAVE)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints/', DECODER_SAVE)))
encoder.eval().cuda()
decoder.eval().cuda()
video1 = next(loader).float().cuda()[0].unsqueeze(0)
video2 = next(loader).float().cuda()[0].unsqueeze(0)
X1, KL1, muL1, det_q1 = encoder(video1)
X2, KL2, muL2, det_q2 = encoder(video2)
# save reconstructed images
dec_v1 = decoder(X1)
save_image(dec_v1.squeeze(0).transpose(2, 3),
'./results/style_transfer_results/recon_v1.png',
nrow=NUM_FRAMES,
normalize=True)
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True))
encoder = Encoder()
encoder.apply(weights_init)
decoder = Decoder()
decoder.apply(weights_init)
encoder.load_state_dict(
torch.load(os.path.join('checkpoints', ENCODER_SAVE)))
decoder.load_state_dict(
torch.load(os.path.join('checkpoints', DECODER_SAVE)))
encoder.eval()
decoder.eval()
prediction_model = Prediction_Model()
prediction_model.apply(weights_init)
if (CUDA):
encoder.cuda()
decoder.cuda()
prediction_model.cuda()
optimizer = torch.optim.Adam(list(prediction_model.parameters()),
lr=LR,
betas=(BETA1, BETA2))
mse_loss = nn.MSELoss()
g_loss.item(),
'gf_loss':
gf_loss.item(),
'fm_loss':
fm_loss.item(),
'vgg_loss':
vgg_loss.item()
if type(vgg_loss) is torch.Tensor else vgg_loss,
'kl_loss':
kl_loss.item()
}, it, 'G')
add_scalar_dict(writer, {
'lr_G': decayed_lr_G,
'lr_D': decayed_lr_D
}, it, 'LR')
E.eval()
G.eval()
with torch.no_grad():
mu, logvar = E(fixed_reals)
latents = sample_latent(mu, logvar)
samples = G(latents, fixed_annos_onehot)
vutils.save_image(
samples,
join(sample_path,
'{:03d}_{:07d}_fake.jpg'.format(ep, it)),
nrow=4,
padding=0,
normalize=True,
range=(-1., 1.))
it += 1
