imgs = Variable(imgs)
if torch.cuda.is_available():
imgs = imgs.cuda()
rgbs = imgs[:, :3, :, :]
depths = imgs[:, 3:-1, :, :]
boudarys = imgs[:, -1:, :, :]
feature = model_enc(rgbs)
pred_semseg1, pred_semseg2, pred_depth, pred_boundary = model_dec(feature)
# if args.use_f2:
# outputs += F2(feature)
# outputs /= 2
total_ent += calc_entropy(pred_semseg1).data.cpu().numpy()[0]
if args.saves_prob:
# Save probability tensors
prob_outdir = os.path.join(base_outdir, "prob")
mkdir_if_not_exist(prob_outdir)
prob_outfn = os.path.join(prob_outdir,
path.split('/')[-1].replace('png', 'npy'))
np.save(prob_outfn, pred_semseg1[0].data.cpu().numpy())
# Save predicted pixel labels(pngs)
if train_args.add_bg_loss:
pred = pred_semseg1[0, :args.n_class].data.max(0)[1].cpu()
else:
pred = pred_semseg1[0, :args.n_class - 1].data.max(0)[1].cpu()
F2.cuda()
total_ent = 0.
for index, (imgs, _, paths) in tqdm(enumerate(target_loader)):
path = paths[0]
imgs = Variable(imgs)
if torch.cuda.is_available():
imgs = imgs.cuda()
g_rgb_outdic = G_3ch(imgs[:, :3, :, :])
g_ir_outdic = G_1ch(imgs[:, 3:, :, :])
outputs = F1(g_rgb_outdic, g_ir_outdic)
total_ent += calc_entropy(outputs).data.cpu().numpy()[0]
if args.saves_prob:
# Save probability tensors
prob_outdir = os.path.join(base_outdir, "prob")
mkdir_if_not_exist(prob_outdir)
prob_outfn = os.path.join(prob_outdir, path.split('/')[-1].replace('png', 'npy'))
np.save(prob_outfn, outputs[0].data.cpu().numpy())
# Save predicted pixel labels(pngs)
pred = outputs[0, :args.n_class - 1].data.max(0)[1].cpu().numpy()
img = Image.fromarray(np.uint8(pred))
img = img.resize(test_img_shape, Image.NEAREST)
label_outdir = os.path.join(base_outdir, "label")
mkdir_if_not_exist(label_outdir)
data_list_fn = os.path.join(base_outdir, "data_list.txt")
with open(data_list_fn, "w") as f:
pass
total_ent = 0.
for index, (imgs, labels, paths) in tqdm(enumerate(target_loader)):
path = paths[0]
imgs = Variable(imgs)
if torch.cuda.is_available():
imgs = imgs.cuda()
preds = model(imgs)
total_ent += calc_entropy(preds).data.cpu().numpy()[0]
subdir = path.split('/')[-2]
if train_args.net == "psp":
preds = preds[0]
if args.saves_prob:
# Save probability tensors
prob_outdir = os.path.join(base_outdir, "prob")
prob_outdir = add_subdir_if_necessary(prob_outdir, subdir, args.tgt_dataset)
mkdir_if_not_exist(prob_outdir)
prob_outfn = os.path.join(prob_outdir, path.split('/')[-1].replace('png', 'npy'))
np.save(prob_outfn, preds[0].data.cpu().numpy())
# Save predicted pixel labels(pngs)