t = adain(content_feat, style_feat4)
t = alpha * t + (1 - alpha) * content_feat
g_t = decoder(t)
g_t_feats1, g_t_feats2, g_t_feats3, g_t_feats4 = vgg(g_t)
loss_content = F.mean_squared_error(g_t_feats4, t)
loss_style = style_loss(style_feat1, g_t_feats1)
loss_style += style_loss(style_feat2, g_t_feats2)
loss_style += style_loss(style_feat3, g_t_feats3)
loss_style += style_loss(style_feat4, g_t_feats4)
loss = content_weight * loss_content + style_weight * loss_style
decoder.cleargrads()
vgg.cleargrads()
loss.backward()
loss.unchain_backward()
dec_opt.update()
vgg_opt.update()
sum_loss += loss.data.get()
if epoch % interval == 0 and batch == 0:
serializers.save_npz("decoder.model", decoder)
with chainer.using_config("train", False):
style_feat1, style_feat2, style_feat3, style_feat4 = vgg(
style_test)
