def validate(args):
print('\nvalidating ... ', flush=True, end='')
model = get_model(args)
model.eval()
val_loader = test_data_loader(args)
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
with torch.no_grad():
for idx, dat in tqdm(enumerate(val_loader)):
img_name, img, label_in, sal = dat
label = label_in.cuda(non_blocking=True)
logits, _, _, _ = model(img)
last_featmaps = model.module.get_heatmaps()
cv_im = cv2.imread(img_name[0])
cv_im_gray = cv2.cvtColor(cv_im, cv2.COLOR_BGR2GRAY)
height, width = cv_im.shape[:2]
for l, featmap in enumerate(last_featmaps):
maps = featmap.cpu().data.numpy()
im_name = args.save_dir + img_name[0].split('/')[-1][:-4]
labels = label_in.long().numpy()[0]
for i in range(int(args.num_classes)):
if labels[i] == 1:
att = maps[i]
att[att < 0] = 0
att = att / (np.max(att) + 1e-8)
att = np.array(att * 255, dtype=np.uint8)
out_name = im_name + '_{}.png'.format(i)
att = cv2.resize(att, (width, height),
interpolation=cv2.INTER_CUBIC)
cv2.imwrite(out_name, att)
args = parser.parse_args()
print(args)
output_dir = os.path.join(args.img_dir, "refined_pseudo_segmentation_labels")
if not os.path.exists(output_dir):
os.makedirs(output_dir)
""" model load """
#model = vgg16(pretrained=True, delta=args.delta)
model = vgg16()
model = model.cuda()
model.eval()
ckpt = torch.load(args.checkpoint, map_location='cpu')
model.load_state_dict(ckpt['model'], strict=True)
""" dataloader """
data_loader = test_data_loader(args)
palette = get_palette()
for idx, dat in enumerate(data_loader):
print("[%03d/%03d]" % (idx, len(data_loader)), end="\r")
img, label, sal_map, _, img_name = dat
label = label.cuda()
img = img.cuda()
_, H, W = sal_map.shape
localization_maps = np.zeros((20, H, W), dtype=np.float32)
""" single-scale testing """
for s in [256, 320, 384]:
_img = F.interpolate(img,
def validate(args):
print('\nvalidating ... ', flush=True, end='')
model = get_model(args)
model.eval()
val_loader = test_data_loader(args)
if not os.path.exists(args.save_dir):
os.mkdir(args.save_dir)
with torch.no_grad():
for idx, dat in tqdm(enumerate(val_loader)):
if idx <= 500000 and idx >= 0:
img_name1, img_name2, input1, input2_list, input1_transforms, label1, label2 = dat
posi_index = np.where(label1.squeeze().cpu().numpy() == 1)[0]
assert len(posi_index) == len(input2_list)
img_list = []
for input2_all in input2_list:
img_all = []
for input2 in input2_all:
img = [input1, input2]
img_all.append(img)
img_list.append(img_all)
assert len(posi_index) == len(img_list)
img2 = [input1, input1_transforms[0]]
img3 = [input1, input1_transforms[1]]
img4 = [input1, input1_transforms[2]]
label_new = label1 + label2
label_new[label_new != 2] = 0
label_new[label_new == 2] = 1
label1_comple = label1 - label_new
label2_comple = label2 - label_new
assert (label1_comple >= 0).all() and (label2_comple >=
0).all()
co_feature1_list = []
for j in range(len(posi_index)):
co_feature1_all = None
label_one = posi_index[j]
for img in img_list[j]:
_, _ = model(img)
_, _, co_feature1, _, _, _ = model.module.get_heatmaps(
)
if co_feature1_all is None:
co_feature1_all = co_feature1
else:
co_feature1_all = co_feature1_all + co_feature1
# co_feature1_all.append(co_feature1)
co_feature1_all = co_feature1_all / len(img_list[j])
co_feature1_list.append([co_feature1_all])
co_feature1_list = feature_map_merge(co_feature1_list, label1)
logits2, co_logits2 = model(img2)
featmaps2_1, featmaps2_2, co_feature2_1, co_feature2_2, _, _ = model.module.get_heatmaps(
)
co_feature2_2 = co_feature2_2.flip(3)
logits3, co_logits3 = model(img3)
featmaps3_1, featmaps3_2, co_feature3_1, co_feature3_2, _, _ = model.module.get_heatmaps(
)
co_feature3_2 = F.upsample(co_feature3_2, (32, 32),
mode='bicubic')
logits4, co_logits4 = model(img4)
featmaps4_1, featmaps4_2, co_feature4_1, co_feature4_2, _, _ = model.module.get_heatmaps(
)
co_feature4_2 = F.upsample(co_feature4_2, (32, 32),
mode='bicubic')
cv_im = cv2.imread(img_name1[0])
cv_im_gray = cv2.cvtColor(cv_im, cv2.COLOR_BGR2GRAY)
height, width = cv_im.shape[:2]
cv_im2 = cv2.imread(img_name2[0])
cv_im_gray2 = cv2.cvtColor(cv_im2, cv2.COLOR_BGR2GRAY)
height2, width2 = cv_im2.shape[:2]
save_feature_maps_all([
co_feature2_1, co_feature2_2, co_feature3_2, co_feature4_2
] + co_feature1_list, args, label1, img_name1, height, width,
cv_im, cv_im_gray)
else:
continue