board_train.add_scalar('Pix2PixHD_Discriminator/loss_D3',
loss_D3.item(), total_steps)
board_train.add_scalar('Pix2PixHD_Discriminator/loss_D3_real',
loss_D3_real.item(), total_steps)
board_train.add_scalar('Pix2PixHD_Discriminator/loss_D3_fake',
loss_D3_fake.item(), total_steps)
print(
"total_steps={}, epoch={}, iters={}, loss_G={:.5f}, loss_feat={:.5f}, loss_vgg={:.5f}, loss_D={:.5f}"
.format(total_steps, epoch, iterations, loss_G, loss_feat,
loss_vgg, loss_D))
visuals = [
[pre_image, after_image, g_fake_img],
]
board_add_images(board_train, 'fake image', visuals,
total_steps)
#====================================================
# test loss の表示
#====================================================
if (total_steps == 1
or (total_steps % args.n_display_test_step == 0)):
model_G.eval()
model_D.eval()
n_test_loop = 0
test_iterations = 0
loss_D_total, loss_D1_total, loss_D2_total, loss_D3_total = 0, 0, 0, 0
loss_D_real_total, loss_D1_real_total, loss_D2_real_total, loss_D3_real_total = 0, 0, 0, 0
loss_D_fake_total, loss_D1_fake_total, loss_D2_fake_total, loss_D3_fake_total = 0, 0, 0, 0
loss_G_total = 0
# backward() で計算した勾配を元に、設定した optimizer に従って、重みを更新
optimizer.step()
#====================================================
# 学習過程の表示
#====================================================
if (step == 0 or (step % args.n_display_step == 0)):
board_train.add_scalar('Model/loss', loss.item(), iterations)
print("epoch={}, iters={}, loss={:.5f}".format(
epoch, iterations, loss))
visuals = [
[pre_image, after_image, output],
]
board_add_images(board_train, 'images', visuals, iterations)
#====================================================
# test loss の表示
#====================================================
if (step == 0 or (step % args.n_display_test_step == 0)):
model.eval()
n_test_loop = 0
test_iterations = 0
loss_total = 0
for test_inputs in dloader_test:
if test_inputs["aerial_image_tsr"].shape[
0] != args.batch_size_test:
break
board_train.add_scalar('D/loss_D_fake', loss_D_fake.item(),
step)
print(
"step={}, loss_G={:.5f}, loss_l1={:.5f}, loss_vgg={:.5f}, loss_entropy={:.5f}, loss_bce={:.5f}, loss_adv={:.5f}"
.format(step, loss_G, loss_l1, loss_vgg, loss_entropy,
loss_bce, loss_adv))
print(
"step={}, loss_D={:.5f}, loss_D_real={:.5f}, loss_D_fake={:.5f}"
.format(step, loss_D.item(), loss_D_real.item(),
loss_D_fake.item()))
visuals = [
[image, mask, output],
]
board_add_images(board_train, 'train', visuals, step + 1)
#=====================================
# 検証用データの処理
#=====================================
if (step == 0 or (step % args.n_display_valid_step == 0)):
loss_G_total = 0
loss_l1_total, loss_vgg_total = 0, 0
loss_entropy_total, loss_bce_total = 0, 0
loss_adv_total = 0
loss_D_total, loss_D_real_total, loss_D_fake_total = 0, 0, 0
n_valid_loop = 0
for iter, inputs in enumerate(tqdm(dloader_valid)):
model_G.eval()
model_D.eval()
board_train.add_scalar('Generater/loss_l1', loss_l1.item(),
iterations)
board_train.add_scalar('Discriminator/loss_D', loss_D.item(),
iterations)
board_train.add_scalar('Discriminator/loss_D_real',
loss_D_real.item(), iterations)
board_train.add_scalar('Discriminator/loss_D_fake',
loss_D_fake.item(), iterations)
print(
"epoch={}, iters={}, loss_G={:.5f}, loss_D={:.5f}".format(
epoch, iterations, loss_G, loss_D))
visuals = [
[pre_image, after_image, g_fake_img],
]
board_add_images(board_train, 'fake image', visuals,
iterations + 1)
#====================================================
# test loss の表示
#====================================================
if (step == 0 or (step % args.n_display_test_step == 0)):
model_G.eval()
model_D.eval()
n_test_loop = 0
test_iterations = 0
loss_D_total = 0
loss_D_real_total = 0
loss_D_fake_total = 0
loss_G_total = 0
loss_gan_total = 0
.format(step, loss_D.item(), loss_D_real.item(),
loss_D_fake.item()))
if (args.model_type_D == "ganimation"):
print("step={}, loss_G_cond_depth={:.5f}".format(
step,
loss_G_cond_depth,
))
print("step={}, loss_D_cond_depth={:.5f}".format(
step,
loss_D_cond_depth,
))
visuals = [
[image, mask, output],
]
board_add_images(board_train, 'images', visuals, step + 1)
step += 1
n_print -= 1
#====================================================
# モデルの保存
#====================================================
if (epoch % args.n_save_epoches == 0):
save_checkpoint(
model_G, device,
os.path.join(args.save_checkpoints_dir, args.exper_name,
'model_ep%03d.pth' % (epoch)))
save_checkpoint(
model_G, device,
os.path.join(args.save_checkpoints_dir, args.exper_name,