img_size=img_size,
z_dim=z_dim,
use_cuda=use_cuda,
conf=conf)
tbar = tqdm(range(num_epochs))
fid = open('losses', 'w')
for epoch in tbar:
train_rec, train_kl = vae.one_epoch(train_loader)
valid_rec, valid_kl = vae.evaluate(validate_loader)
vae.eval()
bx_train, _ = iter(train_loader).next()
bx_valid, _ = iter(validate_loader).next()
bx_train = bx_train[:8]
bx_valid = bx_valid[:8]
rand_samp = vae.random_sample(num_samples=8)
rec_train = vae.reconstruct_img(bx_train)
rec_valid = vae.reconstruct_img(bx_valid)
bx_train = vae.unnormalize(bx_train)
bx_valid = vae.unnormalize(bx_valid)
show_imgs = make_grid(torch.cat(
(rand_samp, bx_train, rec_train, bx_valid, rec_valid), dim=0),
nrow=8)
save_image(show_imgs, 'img_%d.png' % epoch)
vae.train()
tbar.set_description('%8.3f %8.3f %8.3f %8.3f' %
(train_rec, train_kl, valid_rec, valid_kl))
fid.write('%8.3f %8.3f %8.3f %8.3f\n' %
(train_rec, train_kl, valid_rec, valid_kl))
fid.flush()
conf['kl_factor'] = kl_factor
conf['norm_mean'] = mean
conf['norm_std'] = std
vae = VAE(n_channel=3,img_size=img_size,z_dim = z_dim, use_cuda = use_cuda, conf = conf)
tbar = tqdm(range(num_epochs))
fid = open('losses', 'w')
for epoch in tbar:
train_rec, train_kl = vae.one_epoch(train_loader)
valid_rec, valid_kl = vae.evaluate(validate_loader)
vae.eval()
bx_train,_ = iter(train_loader).next()
bx_valid,_ = iter(validate_loader).next()
bx_train = bx_train[:8]
bx_valid = bx_valid[:8]
rand_samp = vae.unnormalize(vae.random_sample(num_samples = 8))
rec_train = vae.unnormalize(vae.reconstruct_img(bx_train))
rec_valid = vae.unnormalize(vae.reconstruct_img(bx_valid))
bx_train = vae.unnormalize(bx_train)
bx_valid = vae.unnormalize(bx_valid)
show_imgs = make_grid(torch.cat((rand_samp, bx_train, rec_train, bx_valid, rec_valid), dim = 0), nrow=8)
save_image(show_imgs,'img_%d.png' % epoch)
vae.train()
tbar.set_description('%8.3f %8.3f %8.3f %8.3f' % (train_rec , train_kl , valid_rec, valid_kl))
fid.write('%8.3f %8.3f %8.3f %8.3f\n' % (train_rec , train_kl , valid_rec, valid_kl))
fid.flush()
fid.close()
vae.eval()
