if args.dtm_type == 'standard':
dtms = prepare_distance_transform_from_mask(masks, mask_size)
elif args.dtm_type == 'reciprocal':
dtms = prepare_reciprocal_DTM_from_mask(masks, mask_size)
elif args.dtm_type == 'complement':
dtms = prepare_complement_DTM_from_mask(masks, mask_size)
elif args.dtm_type == 'other':
dtms = prepare_other_DTM_from_mask(masks, mask_size, args.bg_constant, args.norm_constant)
else:
raise NotImplementedError
all_masks.append(dtms)
# --> encode --> decode.
if args.if_whiten:
centered_dtms = (dtms - shape_mean) / shape_std
dtms_codes = fast_ista(centered_dtms, learned_dict, lmbda=sparse_alpha, max_iter=80)
dtms_rc = torch.matmul(dtms_codes, learned_dict) * shape_std + shape_mean
else:
centered_dtms = dtms - shape_mean
dtms_codes = fast_ista(centered_dtms, learned_dict, lmbda=sparse_alpha, max_iter=80)
dtms_rc = torch.matmul(dtms_codes, learned_dict) + shape_mean
dtms_rc = dtms_rc.numpy()
# evaluate sparsity
sparsity_counts.append(np.sum(np.abs(dtms_codes.numpy()) > 1e-4))
kurtosis_counts.append(dtms_codes.numpy())
# eva.
dtms_rc = np.where(dtms_rc + args.offset >= 0.5, 1, 0) # 0.6 for cityscapes masks, 0.9 for coco masks
IoUevaluate.add_batch(dtms_rc, masks.numpy())
# evaluation.
IoUevaluate = IOUMetric(2)
print("Start evaluation ...")
for i, masks in enumerate(mask_loader):
print("Eva [{} / {}]".format(i, size_data))
# generate the reconstruction mask.
masks = masks.view(masks.shape[0], -1) # a batch of masks: (N, 784)
masks = masks.to(torch.float32)
dtms = prepare_distance_transform_from_mask(
masks, mask_size) # for learn DTMs minusone
# dtms = prepare_overlay_DTMs_from_mask(masks, mask_size) # for learn DTMs overlay
# dtms = prepare_extended_DTMs_from_mask(masks, mask_size) # for learn DTMs extended, 0, 1 range
# --> encode --> decode.
dtms_codes = fast_ista(dtms,
learned_dict,
lmbda=sparse_alpha,
max_iter=100)
dtms_rc = torch.matmul(dtms_codes, learned_dict).numpy()
# evaluate sparsity
sparsity_counts.append(np.sum(np.abs(dtms_codes.numpy()) > 1e-4))
# eva.
dtms_rc = np.where(dtms_rc + 1 - 0.1 > 0.5, 1,
0) # adjust the thresholding to binary masks
# dtms_rc = np.where(dtms_rc - 0.1 > 0.5, 1, 0)
# dtms_rc = np.where(dtms_rc > 0.4, 1, 0)
IoUevaluate.add_batch(dtms_rc, masks.numpy())
codes.append(dtms_codes.numpy())
# do some plots
all_masks = []
reconstruction_error = []
# evaluation.
IoUevaluate = IOUMetric(2)
print("Start evaluation ...")
for i, masks in enumerate(mask_loader):
print("Eva [{} / {}]".format(i, size_data))
# generate the reconstruction mask.
masks = masks.view(masks.shape[0], -1) # a batch of masks: (N, 784)
masks = masks.to(torch.float32)
all_masks.append(masks)
# --> encode --> decode.
mask_codes = fast_ista(masks,
learned_dict,
lmbda=sparse_alpha,
max_iter=80)
mask_rc = torch.matmul(mask_codes, learned_dict).numpy()
# rec_err = np.sum((mask_rc - masks) ** 2, axis=-1).reshape(1, -1)
# reconstruction_error.append(rec_err)
sparsity_counts.append(
np.mean(np.abs(mask_codes.numpy()) > 1e-2, axis=1))
kurtosis_counts.append(mask_codes.numpy())
# eva.
mask_rc = np.where(mask_rc >= 0.5, 1, 0)
IoUevaluate.add_batch(mask_rc, masks.numpy())
# break
num_workers=4)
size_data = len(mask_loader)
# evaluation.
IoUevaluate = IOUMetric(2)
print("Start evaluation ...")
for i, masks in enumerate(mask_loader):
print("Eva [{} / {}]".format(i, size_data))
# generate the reconstruction mask.
masks = masks.view(masks.shape[0], -1) # a batch of masks: (N, 1600)
polys = prepare_polygon_from_mask(masks, mask_size, n_vertices, pads=5)
# masks = masks.to(torch.float32)
# --> encode --> decode.
polygon_codes = fast_ista(polys,
learned_dict,
lmbda=sparse_alpha,
max_iter=80)
polygon_rc = torch.matmul(polygon_codes, learned_dict)
# eva.
mask_rc = poly_to_mask(polygon_rc, mask_size)
# mask_rc = poly_to_mask(polys, mask_size)
# for j in range(mask_rc.shape[0]):
# show_img = np.concatenate([masks[j].numpy().reshape((mask_size, mask_size)),
# mask_rc[j].numpy().reshape((mask_size, mask_size))],
# axis=1).astype(np.uint8) * 255
# cv2.imshow('cat', show_img)
# if cv2.waitKey() & 0xFF == ord('q'):
# break
IoUevaluate.add_batch(mask_rc.numpy(), masks.numpy())
_, _, _, mean_iu, _ = IoUevaluate.evaluate()