def vis_on_batch(self, batch, savedir_image):
image = batch['images']
gt = np.asarray(batch['masks'], np.float32)
gt /= (gt.max() + 1e-8)
res = self.predict_on_batch(batch)
image = F.interpolate(image, size=gt.shape[-2:], mode='bilinear', align_corners=False)
img_res = hu.save_image(savedir_image,
hu.denormalize(image, mode='rgb')[0],
mask=res[0], return_image=True)
img_gt = hu.save_image(savedir_image,
hu.denormalize(image, mode='rgb')[0],
mask=gt[0], return_image=True)
img_gt = models.text_on_image( 'Groundtruth', np.array(img_gt), color=(0,0,0))
img_res = models.text_on_image( 'Prediction', np.array(img_res), color=(0,0,0))
if 'points' in batch:
pts = batch['points'][0].numpy().copy()
pts[pts == 1] = 2
pts[pts == 0] = 1
pts[pts == 255] = 0
img_gt = np.array(hu.save_image(savedir_image, img_gt/255.,
points=pts, radius=2, return_image=True))
img_list = [np.array(img_gt), np.array(img_res)]
hu.save_image(savedir_image, np.hstack(img_list))
def save_tmp(fname, images, logits, points):
from haven import haven_utils as hu
probs = F.softmax(logits, 1)
mask = probs.argmax(dim=1).cpu().numpy().astype('uint8').squeeze() * 255
img_mask = hu.save_image('tmp2.png',
hu.denormalize(images, mode='rgb'),
mask=mask,
return_image=True)
hu.save_image(fname, np.array(img_mask) / 255., points=att_dict['points'])
def vis_on_batch(self, batch, savedir_image, save_preds=False):
self.eval()
images = batch["images"].cuda()
counts = float(batch["counts"][0])
logits = self.model_base(images)
probs = logits.sigmoid()
# get points from attention
att_dict = self.att_model.get_attention_dict(
images_original=torch.FloatTensor(
hu.denormalize(batch['images'], mode='rgb')),
counts=batch['counts'][0],
probs=probs.squeeze(),
return_roi=True)
blobs = lcfcn.get_blobs(probs.squeeze().detach().cpu().numpy())
org_img = hu.denormalize(images.squeeze(), mode='rgb')
rgb_labels = label2rgb(
blobs,
hu.f2l(org_img),
bg_label=0,
bg_color=None,
)
res1 = mark_boundaries(rgb_labels, blobs)
if att_dict['roi_mask'] is not None:
img_mask = hu.save_image('tmp2.png',
org_img,
mask=att_dict['roi_mask'] == 1,
return_image=True)
res2 = hu.save_image('tmp.png',
np.array(img_mask) / 255.,
points=att_dict['points'],
radius=1,
return_image=True)
os.makedirs(os.path.dirname(savedir_image), exist_ok=True)
# plt.savefig(savedir_image.replace('.jpg', '.png')
hu.save_image(savedir_image.replace('.jpg', '.png'),
np.hstack([res1, np.array(res2) / 255.]))
def vis_on_batch(self, batch, savedir_image):
image = batch['images']
res = self.predict_on_batch(batch)
img_gt = hu.save_image('tmp.png',
hu.denormalize(image, mode='rgb')[0],
points=batch['points'][0],
radius=1,
return_image=True)
img_res = hu.save_image(savedir_image,
hu.denormalize(image, mode='rgb')[0],
mask=res.cpu().numpy(),
return_image=True)
# img_gt = semseg.text_on_image( 'Groundtruth', np.array(img_gt), color=(0,0,0))
# img_res = semseg.text_on_image( 'Prediction', np.array(img_res), color=(0,0,0))
hu.save_image(savedir_image,
np.hstack([np.array(img_gt),
np.array(img_res)]))
def vis_on_batch(self, batch, savedir_image):
image = batch['images']
original = hu.denormalize(image, mode='rgb')[0]
img_pred = hu.save_image(savedir_image,
original,
mask=self.predict_on_batch(batch, method='semseg'), return_image=True)
img_gt = hu.save_image(savedir_image,
original,
return_image=True)
gt_counts = float(batch['points'].sum())
pred_counts = self.predict_on_batch(batch, method='counts')
img_gt = models.text_on_image( 'Groundtruth: %d' % gt_counts, np.array(img_gt), color=(0,0,0))
img_pred = models.text_on_image( 'Prediction: %d' % pred_counts, np.array(img_pred), color=(0,0,0))
if 'points' in batch:
pts = (batch['points'][0].numpy().copy() != 0).astype('uint8')
img_gt = np.array(hu.save_image(savedir_image, img_gt/255.,
points=pts.squeeze(), radius=2, return_image=True))
img_list = [np.array(img_gt), np.array(img_pred)]
hu.save_image("%s" %(savedir_image), np.hstack(img_list))
batch = ut.collate_fn([b])
image = batch['images']
gt = np.asarray(batch['masks'], np.float32)
gt /= (gt.max() + 1e-8)
image = F.interpolate(image,
size=gt.shape[-2:],
mode='bilinear',
align_corners=False)
# img_res = hu.save_image('',
# hu.denormalize(image, mode='rgb')[0],
# mask=res[0], return_image=True)
img_gt = hu.save_image('',
hu.denormalize(image, mode='rgb')[0],
mask=gt[0],
return_image=True)
img_gt = models.text_on_image('Groundtruth',
np.array(img_gt),
color=(0, 0, 0))
# img_res = models.text_on_image( 'Prediction', np.array(img_res), color=(0,0,0))
if 'points' in batch:
img_gt = np.array(
hu.save_image('',
img_gt / 255.,
points=batch['points'][0].numpy() != 255,
radius=2,
return_image=True))
img_list = [np.array(img_gt)]
datadir=datadir,
exp_dict=exp_dict,
dataset_size=exp_dict['dataset_size'])
test_loader = DataLoader(
test_set,
# sampler=val_sampler,
batch_size=1,
collate_fn=ut.collate_fn,
num_workers=0)
for i, batch in enumerate(test_loader):
points = (batch['points'].squeeze() == 1).numpy()
if points.sum() == 0:
continue
savedir_image = os.path.join('.tmp/qualitative/%d.png' % (i))
img = hu.denormalize(batch['images'], mode='rgb')
img_org = np.array(
hu.save_image(savedir_image,
img,
mask=batch['masks'].numpy(),
return_image=True))
img_list = [img_org]
with torch.no_grad():
for hash_id in hash_list:
score_path = os.path.join(savedir_base, hash_id,
'score_list_best.pkl')
score_list = hu.load_pkl(score_path)
exp_dict = hu.load_json(
os.path.join(savedir_base, hash_id, 'exp_dict.json'))
exp_dict=exp_dict,
train_set=test_set).cuda()
model_path = '/mnt/public/results/toolkit/weak_supervision/%s/model_best.pth' % hash_dir
# load best model
model.load_state_dict(hu.torch_load(model_path))
# loop over the val_loader and saves image
for i, batch in enumerate(test_loader):
savedir_image = os.path.join("%s" % savedir, "save_preds",
"%s" % hash_dir, "%s" % split,
"%d.png" % i)
image = batch['images']
original = hu.denormalize(image, mode='rgb')[0]
gt = np.asarray(batch['masks'])
image = F.interpolate(image,
size=gt.shape[-2:],
mode='bilinear',
align_corners=False)
img_pred = hu.save_image(savedir_image,
original,
mask=model.predict_on_batch(batch),
return_image=True)
img_gt = hu.save_image(savedir_image,
original,
mask=gt,
return_image=True)
# stop
model = models.get_model(model_dict=exp_dict['model'],
exp_dict=exp_dict,
train_set=train_set).cuda()
exp_id = hu.hash_dict(exp_dict)
fname = os.path.join('/mnt/public/results/toolkit/weak_supervision',
exp_id, 'model.pth')
model.model_base.load_state_dict(torch.load(fname)['model'], strict=False)
for k in range(5):
batch_id = np.where(train_set.labels)[0][k]
batch = ut.collate_fn([train_set[batch_id]])
logits = F.softmax(model.model_base.forward(batch['images'].cuda()),
dim=1)
img = batch['images'].cuda()
logits_new = model_aff.apply_affinity(batch['images'], logits, crf=0)
i1 = hu.save_image('old.png',
img=hu.denormalize(img, mode='rgb'),
mask=logits.argmax(dim=1).cpu().numpy(),
return_image=True)
i2 = hu.save_image('new.png',
img=hu.denormalize(img, mode='rgb'),
mask=logits_new.argmax(dim=1).cpu().numpy(),
return_image=True)
hu.save_image('tmp/tmp%d.png' % k,
np.concatenate([np.array(i1), np.array(i2)], axis=1))
print('saved %d' % k)
def train_on_batch(self, batch, **extras):
self.train()
images = batch["images"].cuda()
counts = float(batch["counts"][0])
logits = self.model_base(images)
if self.exp_dict['model'].get('loss') == 'lcfcn':
loss = lcfcn.compute_lcfcn_loss(logits, batch["points"].cuda(),
None)
probs = F.softmax(logits, 1)
mask = probs.argmax(
dim=1).cpu().numpy().astype('uint8').squeeze() * 255
# img_mask=hu.save_image('tmp2.png',
# hu.denormalize(images, mode='rgb'), mask=mask, return_image=True)
# hu.save_image('tmp2.png',np.array(img_mask)/255. , radius=3,
# points=batch["points"])
elif self.exp_dict['model'].get('loss') == 'glance':
pred_counts = logits[:, 1].mean()
loss = F.mse_loss(pred_counts.squeeze(), counts.float().squeeze())
elif self.exp_dict['model'].get('loss') == 'att_lcfcn':
probs = logits.sigmoid()
# get points from attention
att_dict = self.att_model.get_attention_dict(
images_original=torch.FloatTensor(
hu.denormalize(batch['images'], mode='rgb')),
counts=batch['counts'][0],
probs=probs.squeeze(),
return_roi=True)
if 1:
blobs = lcfcn.get_blobs(probs.squeeze().detach().cpu().numpy())
org_img = hu.denormalize(images.squeeze(), mode='rgb')
rgb_labels = label2rgb(
blobs,
hu.f2l(org_img),
bg_label=0,
bg_color=None,
)
res1 = mark_boundaries(rgb_labels, blobs)
img_mask = hu.save_image('tmp2.png',
org_img,
return_image=True)
res2 = hu.save_image('tmp.png',
np.array(img_mask) / 255.,
points=att_dict['points'],
radius=1,
return_image=True)
hu.save_image('tmp_blobs.png',
np.hstack([res1, np.array(res2) / 255.]))
loss = lcfcn.compute_loss(
probs=probs,
# batch["points"].cuda(),
points=att_dict['points'].cuda(),
roi_mask=att_dict['roi_mask'])
# loss += .5 * F.cross_entropy(logits,
# torch.from_numpy(1 -
# att_dict['mask_bg']).long().cuda()[None],
# ignore_index=1)
self.opt.zero_grad()
loss.backward()
if self.exp_dict['optimizer'] == 'sps':
self.opt.step(loss=loss)
else:
self.opt.step()
return {"train_loss": float(loss)}
def bbox_on_image(batch, bbox_norm, scores, ind_fg):
img_org = hu.denormalize(batch['images'], mode='rgb')
img_org = img_org / img_org.max()
img_box = bbox_on_image_utils(bbox_norm[ind_fg], img_org.squeeze(), text_list=scores)
return np.hstack([img_box])
def vis_on_batch(self, batch, savedir_image):
image = batch['images']
index = batch['meta'][0]['index']
gt = np.asarray(batch['masks'], np.float32)
gt /= (gt.max() + 1e-8)
res = self.predict_on_batch(batch)
image = F.interpolate(image,
size=gt.shape[-2:],
mode='bilinear',
align_corners=False)
original = hu.denormalize(image, mode='rgb')[0]
img_res = hu.save_image(savedir_image,
original,
mask=res[0],
return_image=True)
img_gt = hu.save_image(savedir_image,
original,
mask=gt[0],
return_image=True)
img_gt = models.text_on_image('Groundtruth',
np.array(img_gt),
color=(0, 0, 0))
img_res = models.text_on_image('Prediction',
np.array(img_res),
color=(0, 0, 0))
if 'points' in batch:
pts = batch['points'][0].numpy().copy()
pts[pts == 1] = 2
pts[pts == 0] = 1
pts[pts == 255] = 0
img_gt = np.array(
hu.save_image(savedir_image,
img_gt / 255.,
points=pts,
radius=2,
return_image=True))
# score map
if self.heuristic != 'random':
score_map = self.compute_uncertainty(batch['images'],
replicate=True,
scale_factor=1,
method=self.heuristic)
score_map = F.interpolate(score_map[None],
size=gt.shape[-2:],
mode='bilinear',
align_corners=False).squeeze()
h, w = score_map.shape
bbox_yxyx = get_rect_bbox(h, w, n_regions=self.n_regions)
heatmap = hu.f2l(
hi.gray2cmap(
(score_map / score_map.max()).cpu().numpy().squeeze()))
s_list = np.zeros(len(bbox_yxyx))
for i, (y1, x1, y2, x2) in enumerate(bbox_yxyx):
s_list[i] = score_map[y1:y2, x1:x2].mean()
img_bbox = bbox_yxyx_on_image(bbox_yxyx[[s_list.argmax()]],
original)
img_score_map = img_bbox * 0.5 + heatmap * 0.5
img_list = [
np.array(img_gt),
np.array(img_res), (img_score_map * 255).astype('uint8')
]
else:
img_list = [np.array(img_gt), np.array(img_res)]
hu.save_image(savedir_image, np.hstack(img_list))
def save_images(exp_dict):
dataset_name = exp_dict['dataset']['name']
n_classes = exp_dict['dataset']['n_classes']
model = models.get_model(model_dict=exp_dict['model'],
exp_dict=exp_dict,
train_set=None).cuda()
state_dict = hc.load_checkpoint(exp_dict,
savedir_base,
fname='model_best.pth')
model.load_state_dict(state_dict)
model.eval()
np.random.seed(1)
train_set = datasets.get_dataset(dataset_dict={'name': dataset_name},
datadir=None,
split="test",
exp_dict=exp_dict)
n_images = 0
for _ in range(len(train_set)):
i = np.random.choice(len(train_set))
b = train_set[i]
if n_images > 5:
break
if b['masks'].sum() == 0:
print(i)
continue
n_images += 1
batch = ut.collate_fn([b])
image = batch['images']
gt = np.asarray(batch['masks'], np.float32)
gt /= (gt.max() + 1e-8)
image = F.interpolate(image,
size=gt.shape[-2:],
mode='bilinear',
align_corners=False)
img_rgb = hu.f2l(hu.denormalize(image, mode='rgb')[0])
img_rgb = (np.array(img_rgb) * 255.).astype('uint8')
# save rgb
fname_rgb = '.tmp/covid_qualitative/%s/%s/%d_rgb.png' % (exp_group,
'gt', i)
hu.save_image(fname_rgb, img_rgb)
# save pts
fname_pts = '.tmp/covid_qualitative/%s/%s/%d_pts.png' % (exp_group,
'gt', i)
img_gt = np.array(hu.save_image('', img_rgb, return_image=True))
if 'points' in batch:
pts = batch['points'][0].numpy()
pts[pts == 1] = 2
pts[pts == 0] = 1
pts[pts == 255] = 0
img_gt = np.array(
hu.save_image('',
img_gt / 255.,
points=pts,
radius=2,
return_image=True))
hu.save_image(fname_pts, img_gt)
# save mask
fname_mask = '.tmp/covid_qualitative/%s/%s/%d_mask.png' % (exp_group,
'gt', i)
img_mask = np.array(
hu.save_image('', img_rgb, mask=gt[0], return_image=True))
hu.save_image(fname_mask, img_mask)
# pred
fname_pred = '.tmp/covid_qualitative/%s/%s/%d_%s.png' % (
exp_group, 'preds', i, exp_dict['model']['loss'])
res = model.predict_on_batch(batch)
img_res = hu.save_image('', img_rgb, mask=res[0], return_image=True)
hu.save_image(fname_pred, np.array(img_res))