Skip to content

PSEDiceLoss support any input shape (fix issue#333) #339

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 2, 2023
Merged

PSEDiceLoss support any input shape (fix issue#333) #339

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 5 additions & 19 deletions mindocr/data/transforms/det_transforms.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -447,7 +447,7 @@ def __call__(self, data: dict):
data['polys'][:, :, 1] = data['polys'][:, :, 1] * scale_h
#print('transform GT polys to: ', data['polys'])

if 'shape_list' not in data:
if 'shape_list' not in data:
src_h, src_w = data.get('raw_img_shape', (h, w))
data['shape_list'] = [src_h, src_w, scale_h, scale_w]
else:
Expand Down Expand Up @@ -503,33 +503,21 @@ def __init__(self, kernel_num=7, min_shrink_ratio=0.4, min_shortest_edge=640, **
self.min_shrink_ratio = min_shrink_ratio
self.min_shortest_edge = min_shortest_edge

@staticmethod
def _dist(point_1, point_2):
return np.sqrt(np.sum((point_1 - point_2) ** 2))

def _perimeter(self, bbox):
peri = 0.0
for i in range(bbox.shape[0]):
peri += self._dist(bbox[i], bbox[(i + 1) % bbox.shape[0]])
return peri

def _shrink(self, text_polys, rate, max_shr=20):
rate = rate * rate
shrinked_text_polys = []
for bbox in text_polys:
area = Polygon(bbox).area
peri = self._perimeter(bbox)
poly = Polygon(bbox)
area, peri = poly.area, poly.length

pco = pyclipper.PyclipperOffset()
pco.AddPath(bbox, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
offset = min((int)(area * (1 - rate) / (peri + 0.001) + 0.5), max_shr)

shrinked_bbox = pco.Execute(-offset) # (N, 2) shape, N maybe larger than or smaller than 4.
shrinked_bbox = expand_poly(bbox, -offset, pyclipper.JT_ROUND) # (N, 2) shape, N maybe larger than or smaller than 4.
if not shrinked_bbox:
shrinked_text_polys.append(bbox)
continue

shrinked_bbox = np.array(shrinked_bbox)[0]
shrinked_bbox = np.array(shrinked_bbox)
if shrinked_bbox.shape[0] <= 2:
shrinked_text_polys.append(bbox)
continue
Expand Down Expand Up @@ -618,7 +606,6 @@ def __call__(self, data: dict):
poly = poly.exterior
poly = poly.coords[::-1] if poly.is_ccw else poly.coords # sort in clockwise order
new_polys.append(np.array(poly[:-1]))

else: # the polygon is fully outside the image
continue
new_tags.append(ignore)
Expand All @@ -627,5 +614,4 @@ def __call__(self, data: dict):
data['polys'] = new_polys
data['texts'] = new_texts
data['ignore_tags'] = np.array(new_tags)

return data
53 changes: 28 additions & 25 deletions mindocr/losses/det_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,12 +156,12 @@ def construct(self, pred, gt, mask):


class PSEDiceLoss(nn.Cell):
def __init__(self):
def __init__(self, alpha=0.7, ohem_ratio=3):
super().__init__()

self.threshold0 = Tensor(0.5, mstype.float32)
self.zero_float32 = Tensor(0.0, mstype.float32)
self.k = int(640 * 640)
self.alpha = alpha
self.ohem_ratio = ohem_ratio
self.negative_one_int32 = Tensor(-1, mstype.int32)
self.concat = ops.Concat()
self.less_equal = ops.LessEqual()
Expand Down Expand Up @@ -197,16 +197,17 @@ def ohem_batch(self, scores, gt_texts, training_masks):
:return: [N * H * W]
'''
batch_size = scores.shape[0]
h, w = scores.shape[1:]
selected_masks = ()
for i in range(batch_size):
score = self.slice(scores, (i, 0, 0), (1, 640, 640))
score = self.reshape(score, (640, 640))
score = self.slice(scores, (i, 0, 0), (1, h, w))
score = self.reshape(score, (h, w))

gt_text = self.slice(gt_texts, (i, 0, 0), (1, 640, 640))
gt_text = self.reshape(gt_text, (640, 640))
gt_text = self.slice(gt_texts, (i, 0, 0), (1, h, w))
gt_text = self.reshape(gt_text, (h, w))

training_mask = self.slice(training_masks, (i, 0, 0), (1, 640, 640))
training_mask = self.reshape(training_mask, (640, 640))
training_mask = self.slice(training_masks, (i, 0, 0), (1, h, w))
training_mask = self.reshape(training_mask, (h, w))

selected_mask = self.ohem_single(score, gt_text, training_mask)
selected_masks = selected_masks + (selected_mask,)
Expand All @@ -215,22 +216,24 @@ def ohem_batch(self, scores, gt_texts, training_masks):
return selected_masks

def ohem_single(self, score, gt_text, training_mask):
h, w = score.shape[0:2]
k = int(h * w)
Copy link
Collaborator

@hadipash hadipash Jun 2, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not important in this case because we don't support dynamic shaping, but for the future:
during the graph construction all constants (integers, floats, etc.) will be calculated once and will not be changed during execution anymore. So, need to be careful if you intend k to change.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, noted.

pos_num = self.logical_and(self.greater(gt_text, self.threshold0),
self.greater(training_mask, self.threshold0))
pos_num = self.reduce_sum(self.cast(pos_num, mstype.float32))

neg_num = self.less_equal(gt_text, self.threshold0)
neg_num = self.reduce_sum(self.cast(neg_num, mstype.float32))
neg_num = self.minimum(3 * pos_num, neg_num)
neg_num = self.minimum(self.ohem_ratio * pos_num, neg_num)
neg_num = self.cast(neg_num, mstype.int32)

neg_num = neg_num + self.k - 1
neg_num = neg_num + k - 1
neg_mask = self.less_equal(gt_text, self.threshold0)
ignore_score = self.fill(mstype.float32, (640, 640), -1e3)
ignore_score = self.fill(mstype.float32, (h, w), -1e3)
neg_score = self.select(neg_mask, score, ignore_score)
neg_score = self.reshape(neg_score, (640 * 640,))
neg_score = self.reshape(neg_score, (h * w,))

topk_values, _ = self.topk(neg_score, self.k)
topk_values, _ = self.topk(neg_score, k)
threshold = self.gather(topk_values, neg_num, 0)

selected_mask = self.logical_and(
Expand All @@ -254,9 +257,9 @@ def dice_loss(self, input_params, target, mask):
batch_size = input_params.shape[0]
input_sigmoid = self.sigmoid(input_params)

input_reshape = self.reshape(input_sigmoid, (batch_size, 640 * 640))
target = self.reshape(target, (batch_size, 640 * 640))
mask = self.reshape(mask, (batch_size, 640 * 640))
input_reshape = self.reshape(input_sigmoid, (batch_size, -1))
target = self.reshape(target, (batch_size, -1))
mask = self.reshape(mask, (batch_size, -1))

input_mask = input_reshape * mask
target = target * mask
Expand Down Expand Up @@ -286,16 +289,16 @@ def construct(self, model_predict, gt_texts, gt_kernels, training_masks):
'''
batch_size = model_predict.shape[0]
model_predict = self.upsample(model_predict, scale_factor=4)
texts = self.slice(model_predict, (0, 0, 0, 0), (batch_size, 1, 640, 640))
texts = self.reshape(texts, (batch_size, 640, 640))
h, w = model_predict.shape[2:]
texts = self.slice(model_predict, (0, 0, 0, 0), (batch_size, 1, h, w))
texts = self.reshape(texts, (batch_size, h, w))
selected_masks_text = self.ohem_batch(texts, gt_texts, training_masks)
loss_text = self.dice_loss(texts, gt_texts, selected_masks_text)

kernels = []
loss_kernels = []
for i in range(1, 7):
kernel = self.slice(model_predict, (0, i, 0, 0), (batch_size, 1, 640, 640))
kernel = self.reshape(kernel, (batch_size, 640, 640))
kernel = self.slice(model_predict, (0, i, 0, 0), (batch_size, 1, h, w))
kernel = self.reshape(kernel, (batch_size, h, w))
kernels.append(kernel)

mask0 = self.sigmoid(texts)
Expand All @@ -304,13 +307,13 @@ def construct(self, model_predict, gt_texts, gt_kernels, training_masks):
selected_masks_kernels = self.cast(selected_masks_kernels, mstype.float32)

for i in range(6):
gt_kernel = self.slice(gt_kernels, (0, i, 0, 0), (batch_size, 1, 640, 640))
gt_kernel = self.reshape(gt_kernel, (batch_size, 640, 640))
gt_kernel = self.slice(gt_kernels, (0, i, 0, 0), (batch_size, 1, h, w))
gt_kernel = self.reshape(gt_kernel, (batch_size, h, w))
loss_kernel_i = self.dice_loss(kernels[i], gt_kernel, selected_masks_kernels)
loss_kernels.append(loss_kernel_i)
loss_kernel = self.avg_losses(loss_kernels)

loss = 0.7 * loss_text + 0.3 * loss_kernel
loss = self.alpha * loss_text + (1 - self.alpha) * loss_kernel
return loss


Expand Down