outs = network(training_batch)
adam_optim.zero_grad()
loss = objective(outs, trues)
loss.backward()
adam_optim.step()
print(step_num, loss.item())
if step_num < 200000:
lr = 1e-4
elif step_num < 400000:
lr = 1e-5
else:
lr = 1e-6
for params in adam_optim.param_groups:
params["lr"] = lr
writer.add_scalar("train/loss", loss.item(), step_num)
if step_num % 100 == 0:
network.eval()
outs = network(training_batch)
loss = objective(outs, trues)
writer.add_scalar("val/loss", loss.item(), step_num)
for iter in np.arange(32):
out_drawing = outs[iter].cpu().data.numpy()
true_drawing = trues[iter].cpu().data.numpy()
writer.add_image("train/out_drawing{}.png".format(iter),
out_drawing, step_num)
writer.add_image("train/true_drawing{}.png".format(iter),
true_drawing, step_num)
if step_num % 1000 == 0:
save_model()
step_num += 1
train_batch = train_batch.cuda()
ground_truth = ground_truth.cuda()
gen = net(train_batch)
optimizer.zero_grad()
loss = criterion(gen, ground_truth)
loss.backward()
optimizer.step()
print(step, loss.item())
if step < 200000:
lr = 1e-4
elif step < 400000:
lr = 1e-5
else:
lr = 1e-6
for param_group in optimizer.param_groups:
param_group["lr"] = lr
writer.add_scalar("train/loss", loss.item(), step)
if step % 100 == 0:
net.eval()
gen = net(train_batch)
loss = criterion(gen, ground_truth)
writer.add_scalar("val/loss", loss.item(), step)
for i in range(32):
G = gen[i].cpu().data.numpy()
GT = ground_truth[i].cpu().data.numpy()
writer.add_image("train/gen{}.png".format(i), G, step)
writer.add_image("train/ground_truth{}.png".format(i), GT, step)
if step % 1000 == 0:
save_model()
step += 1
lr = 1e-4
elif step < 400000:
lr = 1e-5
else:
lr = 1e-6
for param_group in optimizer.param_groups:
param_group["lr"] = lr
if args.constrained and (step + 1) % dec_brush_width_int == 0:
curr_brush_width -= dec_brush_width
curr_brush_width = max(curr_brush_width, CONSTRAINT_BRUSH_WIDTH)
writer.add_scalar("train/loss", loss.item(), step)
if step % 500 == 0:
net.eval()
gen = net(train_batch)
loss = criterion(gen, ground_truth)
writer.add_scalar("val/loss", loss.item(), step)
for i in range(32):
G = gen[i].cpu().data.numpy() * 255
GT = ground_truth[i].cpu().data.numpy() * 255
# G = np.array([G,G,G])
# GT = np.array([GT,GT,GT])
# writer.add_image("train/img{}.png".format(i), np.transpose(G, (1,2,0)), step)
# writer.add_image("train/img{}_truth.png".format(i), np.transpose(GT, (1,2,0)), step)
writer.add_image("train/img{}.png".format(i), G, step)
writer.add_image("train/img{}_truth.png".format(i), GT, step)
if step % 1000 == 0:
save_model()
step += 1
