board_train.add_scalar('D/loss_D', loss_D.item(), step)
print(
"step={}, loss_G={:.5f}, loss_l1={:.5f}, loss_vgg={:.5f}".
format(step, loss_G.item(), loss_l1.item(),
loss_vgg.item(), loss_adv.item()))
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()))
# visual images
visuals = [
[pose_parse_onehot_vis, pose_gt, output],
]
board_add_images(board_train, 'train', visuals, step + 1)
#====================================================
# valid データでの処理
#====================================================
if (step % args.n_display_valid_step == 0):
loss_G_total, loss_l1_total, loss_vgg_total, loss_adv_total = 0, 0, 0, 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,
desc="valid")):
model_G.eval()
model_D.eval()
# 一番最後のミニバッチループで、バッチサイズに満たない場合は無視する（後の計算で、shape の不一致をおこすため）
if inputs["pose_parse_onehot"].shape[
loss_rec.item()))
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()))
# visual images
"""
visuals = [
[ noize_image_z, image_gt, output ],
]
"""
visuals = [
[image_gt, output],
]
board_add_images(board_train, 'train', visuals, step + 1)
for i in range(len(output_list)):
visuals = [
[image_gt_list[i], output_list[i]],
]
board_add_images(board_train,
'train_progress_{}'.format(i),
visuals, step + 1)
#====================================================
# valid データでの処理
#====================================================
"""
if( step != 0 and ( step % args.n_display_valid_step == 0 ) ):
loss_G_total, loss_l1_total, loss_vgg_total, loss_adv_total = 0, 0, 0, 0
image_name = inputs["image_name"]
image = inputs["image"].to(device)
if (args.debug and n_print > 0):
print("image.shape : ", image.shape)
#----------------------------------------------------
# 生成器の推論処理
#----------------------------------------------------
with torch.no_grad():
output = model_G(image)
if (args.debug and n_print > 0):
print("output.shape : ", output.shape)
#====================================================
# 推論結果の保存
#====================================================
save_image_w_norm(
output,
os.path.join(args.results_dir, args.exper_name, "output",
image_name[0]))
# tensorboard
visuals = [
[image, output],
]
board_add_images(board_test, 'test', visuals, step + 1)
n_print -= 1
if (step >= args.n_samplings):
break
materials=materials) * 2.0 - 1.0
save_image(
mesh_img_tsr.transpose(1, 3).transpose(2, 3),
os.path.join(args.results_dir, args.exper_name, "mesh.png"))
save_image(
mesh_ffd_img_tsr.transpose(1, 3).transpose(2, 3),
os.path.join(args.results_dir, args.exper_name, "mesh_ffd.png"))
# visual images
visuals = [
[
mesh_img_tsr.transpose(1, 3).transpose(2, 3),
mesh_ffd_img_tsr.transpose(1, 3).transpose(2, 3)
],
]
board_add_images(board_train, 'render_ffd', visuals, step + 1)
#-----------------------------
# カメラの移動＆再レンダリング
#-----------------------------
camera_dist += 0.0
camera_elev += 0.0
camera_azim += 5.0
rot_matrix, trans_matrix = look_at_view_transform(dist=camera_dist,
elev=camera_elev,
azim=camera_azim)
new_cameras = OpenGLPerspectiveCameras(device=device,
R=rot_matrix,
T=trans_matrix)
# メッシュのレンダリング
mesh_img_tsr[:,1,:,:] = mesh_img_tsr[:,1,:,:] * mesh_img_tsr[:,-1,:,:]
mesh_img_tsr[:,2,:,:] = mesh_img_tsr[:,2,:,:] * mesh_img_tsr[:,-1,:,:]
mesh_img_tsr = mesh_img_tsr[:,0:3,:,:]
mesh_img_tsr = torch.from_numpy( (mesh_img_tsr.detach().cpu().numpy() > 0).astype(np.float32) * 2.0 ).to(device) - 1.0
"""
save_image(
mesh_img_tsr,
os.path.join(args.results_dir, args.exper_name,
"mesh_render_light.png"))
# visual images
visuals = [
[mesh_img_tsr],
]
board_add_images(board_train, 'render_light', visuals, step + 1)
#-----------------------------
# カメラの移動＆再レンダリング
#-----------------------------
camera_dist += 0.1
camera_elev += 5.0
camera_azim += 5.0
rot_matrix, trans_matrix = look_at_view_transform(dist=camera_dist,
elev=camera_elev,
azim=camera_azim)
new_cameras = OpenGLPerspectiveCameras(device=device,
R=rot_matrix,
T=trans_matrix)
# メッシュのレンダリング
output[:, i, :, :].unsqueeze(1)
for i in range(args.n_classes // 4 +
1, args.n_classes // 2)
],
[
output[:, i, :, :].unsqueeze(1) for i in range(
args.n_classes // 2 + 1, args.n_classes // 2 +
args.n_classes // 4)
],
[
output[:, i, :, :].unsqueeze(1) for i in range(
args.n_classes // 2 + args.n_classes // 4 +
1, args.n_classes)
],
]
board_add_images(board_train, 'train', visuals, step + 1)
# visual deeplab v3+ output
visuals = [
[
embedded[:, i, :, :].unsqueeze(1)
for i in range(0, args.n_classes // 4)
],
[
embedded[:, i, :, :].unsqueeze(1)
for i in range(args.n_classes // 4 +
1, args.n_classes // 2)
],
[
embedded[:, i, :, :].unsqueeze(1) for i in range(
args.n_classes // 2 + 1, args.n_classes // 2 +
"source_mesh_step{}.png".format(step)),
"source mesh")
mesh_s_img = get_plot3d_mesh_img(mesh_s, "source mesh")
mesh_t_img = get_plot3d_mesh_img(mesh_t, "target mesh")
mesh_s_new_img = get_plot3d_mesh_img(mesh_s_new, "source mesh")
mesh_s_tsr = torchvision.transforms.functional.to_tensor(
mesh_s_img).unsqueeze(0)
mesh_s_new_tsr = torchvision.transforms.functional.to_tensor(
mesh_s_new_img).unsqueeze(0)
mesh_t_tsr = torchvision.transforms.functional.to_tensor(
mesh_t_img).unsqueeze(0)
visuals = [
[mesh_s_tsr, mesh_t_tsr, mesh_s_new_tsr],
]
board_add_images(board_train, 'train', visuals, step + 1)
step += 1
n_print -= 1
#====================================================
# モデルの保存
#====================================================
if (epoch % args.n_save_epoches == 0):
print("saved mesh")
# Fetch the verts and faces of the final predicted mesh
final_verts, final_faces = mesh_s_new.get_mesh_verts_faces(0)
# Scale normalize back to the original target size
final_verts = final_verts * scale + center
[image, target, output],
]
else:
visuals = [
[image, target_rgb, output_vis_rgb],
[
output[:, i, :, :].unsqueeze(1)
for i in range(0, args.n_classes_target // 2)
],
[
output[:, i, :, :].unsqueeze(1)
for i in range(args.n_classes_target // 2 +
1, args.n_classes_target)
],
]
board_add_images(board_train, 'train', visuals, step + 1)
# visual deeplab v3+ output
visuals = [
[
embedded[:, i, :, :].unsqueeze(1)
for i in range(0, args.n_classes_target // 2)
],
[
embedded[:, i, :, :].unsqueeze(1)
for i in range(args.n_classes_target // 2 +
1, args.n_classes_target)
],
]
board_add_images(board_train, 'train_deeplab_embedded',
visuals, step + 1)
