未验证 提交 f8fb8a7b 编写于 作者: D Double_V 提交者: GitHub

Merge pull request #7406 from LDOUBLEV/dygraph

fix cml
...@@ -191,7 +191,6 @@ Eval: ...@@ -191,7 +191,6 @@ Eval:
channel_first: False channel_first: False
- DetLabelEncode: # Class handling label - DetLabelEncode: # Class handling label
- DetResizeForTest: - DetResizeForTest:
# image_shape: [736, 1280]
- NormalizeImage: - NormalizeImage:
scale: 1./255. scale: 1./255.
mean: [0.485, 0.456, 0.406] mean: [0.485, 0.456, 0.406]
......
...@@ -24,6 +24,7 @@ Architecture: ...@@ -24,6 +24,7 @@ Architecture:
model_type: det model_type: det
Models: Models:
Student: Student:
pretrained:
model_type: det model_type: det
algorithm: DB algorithm: DB
Transform: null Transform: null
...@@ -40,6 +41,7 @@ Architecture: ...@@ -40,6 +41,7 @@ Architecture:
name: DBHead name: DBHead
k: 50 k: 50
Student2: Student2:
pretrained:
model_type: det model_type: det
algorithm: DB algorithm: DB
Transform: null Transform: null
...@@ -91,14 +93,11 @@ Loss: ...@@ -91,14 +93,11 @@ Loss:
- ["Student", "Student2"] - ["Student", "Student2"]
maps_name: "thrink_maps" maps_name: "thrink_maps"
weight: 1.0 weight: 1.0
# act: None
model_name_pairs: ["Student", "Student2"] model_name_pairs: ["Student", "Student2"]
key: maps key: maps
- DistillationDBLoss: - DistillationDBLoss:
weight: 1.0 weight: 1.0
model_name_list: ["Student", "Student2"] model_name_list: ["Student", "Student2"]
# key: maps
# name: DBLoss
balance_loss: true balance_loss: true
main_loss_type: DiceLoss main_loss_type: DiceLoss
alpha: 5 alpha: 5
...@@ -197,6 +196,7 @@ Train: ...@@ -197,6 +196,7 @@ Train:
drop_last: false drop_last: false
batch_size_per_card: 8 batch_size_per_card: 8
num_workers: 4 num_workers: 4
Eval: Eval:
dataset: dataset:
name: SimpleDataSet name: SimpleDataSet
...@@ -204,31 +204,21 @@ Eval: ...@@ -204,31 +204,21 @@ Eval:
label_file_list: label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt - ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms: transforms:
- DecodeImage: - DecodeImage: # load image
img_mode: BGR img_mode: BGR
channel_first: false channel_first: False
- DetLabelEncode: null - DetLabelEncode: # Class handling label
- DetResizeForTest: null - DetResizeForTest:
- NormalizeImage: - NormalizeImage:
scale: 1./255. scale: 1./255.
mean: mean: [0.485, 0.456, 0.406]
- 0.485 std: [0.229, 0.224, 0.225]
- 0.456 order: 'hwc'
- 0.406 - ToCHWImage:
std: - KeepKeys:
- 0.229 keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
- 0.224
- 0.225
order: hwc
- ToCHWImage: null
- KeepKeys:
keep_keys:
- image
- shape
- polys
- ignore_tags
loader: loader:
shuffle: false shuffle: False
drop_last: false drop_last: False
batch_size_per_card: 1 batch_size_per_card: 1 # must be 1
num_workers: 2 num_workers: 2
\ No newline at end of file
...@@ -60,19 +60,19 @@ class KLJSLoss(object): ...@@ -60,19 +60,19 @@ class KLJSLoss(object):
], "mode can only be one of ['kl', 'KL', 'js', 'JS']" ], "mode can only be one of ['kl', 'KL', 'js', 'JS']"
self.mode = mode self.mode = mode
def __call__(self, p1, p2, reduction="mean"): def __call__(self, p1, p2, reduction="mean", eps=1e-5):
if self.mode.lower() == 'kl': if self.mode.lower() == 'kl':
loss = paddle.multiply(p2, loss = paddle.multiply(p2,
paddle.log((p2 + 1e-5) / (p1 + 1e-5) + 1e-5)) paddle.log((p2 + eps) / (p1 + eps) + eps))
loss += paddle.multiply( loss += paddle.multiply(p1,
p1, paddle.log((p1 + 1e-5) / (p2 + 1e-5) + 1e-5)) paddle.log((p1 + eps) / (p2 + eps) + eps))
loss *= 0.5 loss *= 0.5
elif self.mode.lower() == "js": elif self.mode.lower() == "js":
loss = paddle.multiply( loss = paddle.multiply(
p2, paddle.log((2 * p2 + 1e-5) / (p1 + p2 + 1e-5) + 1e-5)) p2, paddle.log((2 * p2 + eps) / (p1 + p2 + eps) + eps))
loss += paddle.multiply( loss += paddle.multiply(
p1, paddle.log((2 * p1 + 1e-5) / (p1 + p2 + 1e-5) + 1e-5)) p1, paddle.log((2 * p1 + eps) / (p1 + p2 + eps) + eps))
loss *= 0.5 loss *= 0.5
else: else:
raise ValueError( raise ValueError(
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册