模型
neck部分：
- 错误：修改Eval 配置 DetResizeForTest:image_shape: [1000, 1000]报错
- 原因：图片resize后，in4 = self.in4_conv(c4)卷积后，[1, 256, 63, 63] 3,4维不一定是64，out4 = in4 + F.upsample(in5, scale_factor=2, mode="nearest", align_mode=1)维度不一样，无法相加。
File "/data/root/PaddleOCR2.5/ppocr/modeling/architectures/base_model.py", line 82, in forward
x = self.neck(x)
File "/opt/conda/lib/python3.8/site-packages/paddle/fluid/dygraph/layers.py", line 917, in __call__
return self._dygraph_call_func(*inputs, **kwargs)
File "/opt/conda/lib/python3.8/site-packages/paddle/fluid/dygraph/layers.py", line 907, in _dygraph_call_func
outputs = self.forward(*inputs, **kwargs)
File "/data/root/PaddleOCR2.5/ppocr/modeling/necks/db_fpn.py", line 174, in forward
out4 = in4 + F.upsample(
File "/opt/conda/lib/python3.8/site-packages/paddle/fluid/dygraph/math_op_patch.py", line 264, in __impl__
return math_op(self, other_var, 'axis', axis)
ValueError: (InvalidArgument) Broadcast dimension mismatch. Operands could not be broadcast together with the shape of X = [1, 256, 63, 63] and the shape of Y = [1, 256, 64, 64]. Received [63] in X i s not equal to [64] in Y at i:2.
[Hint: Expected x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i] <= 1 == true, but received x_dims_array[i] == y_dims_array[i] || x_dims_array[i] <= 1 || y_dims_array[i ] <= 1:0 != true:1.] (at /paddle/paddle/fluid/operators/elementwise/elementwise_op_function.h:240)
[operator < elementwise_add > error]
对应的代码：
位置：PaddleOCR2.5/ppocr/modeling/necks/db_fpn.py
- 修正1：为保证卷积后int4维度和上采样int5后维度一致，int4需补差值
def forward(self, x):
#print('run FPN')
c2, c3, c4, c5 = x #backbone的输出x2，x3，x4，x5 c2:[1, 64, 160, 160], c3:[1, 128, 80, 80], c4:[1, 256, 40, 40], c5:[1, 512, 20, 20]
# Top-down
p5 = self.reduce_conv_c5(c5) #[1, 512, 20, 20]->[1, 64, 20, 20]
p4 = self._upsample_add(p5, self.reduce_conv_c4(c4)) # p4 & reduce_conv_c4(c4):[1, 64, 40, 40]
p4 = self.smooth_p4(p4) #[1, 64, 40, 40]
p3 = self._upsample_add(p4, self.reduce_conv_c3(c3))#reduce_conv_c3(c3):[1, 64, 80, 80]
p3 = self.smooth_p3(p3) #[1, 64, 80, 80]
p2 = self._upsample_add(p3, self.reduce_conv_c2(c2))
p2 = self.smooth_p2(p2) #[1, 64, 160, 160]
x = self._upsample_cat(p2, p3, p4, p5) #[1, 256, 160, 160]
x = self.conv(x)
return x
# 把x差值变大到y的大小，再相加
def _upsample_add(self, x, y):
return F.interpolate(x, size=y.size()[2:],mode='bilinear',align_corners=True) + y
- 修正2：将图像大小调整为网络所需的32倍大小，原代码无需修改
前处理
图片resize：
- Eval 配置 DetResizeForTest，不同方式resize图片
class DetResizeForTest(object):
def __init__(self, **kwargs):
super(DetResizeForTest, self).__init__()
self.resize_type = 0
if 'image_shape' in kwargs:
self.image_shape = kwargs['image_shape']
self.resize_type = 1
elif 'limit_side_len' in kwargs:
self.limit_side_len = kwargs['limit_side_len']
self.limit_type = kwargs.get('limit_type', 'min')
elif 'resize_long' in kwargs:
self.resize_type = 2
self.resize_long = kwargs.get('resize_long', 960)
else:
self.limit_side_len = 736
self.limit_type = 'min'
对应的代码：
位置：PaddleOCR2.5/ppocr/data/imaug/operators.py
可视化MakeBorderMap：
class MakeBorderMap(object):
def __init__(self,
shrink_ratio=0.4,
thresh_min=0.3,
thresh_max=0.7,
**kwargs):
self.shrink_ratio = shrink_ratio
self.thresh_min = thresh_min
self.thresh_max = thresh_max
def __call__(self, data):
img = data['image']
text_polys = data['polys']
ignore_tags = data['ignore_tags']
canvas = np.zeros(img.shape[:2], dtype=np.float32)
mask = np.zeros(img.shape[:2], dtype=np.float32)
for i in range(len(text_polys)):
if ignore_tags[i]:
continue
self.draw_border_map(text_polys[i], canvas, mask=mask)
canvas = canvas * (self.thresh_max - self.thresh_min) + self.thresh_min
# 可视化代码
import os
canvas = cv2.cvtColor(np.array(canvas)*255, cv2.COLOR_GRAY2BGR)
cv2.imwrite('/data/PaddleOCR2.5/temp_border_map/' + os.path.basename(data['img_path']), canvas)
import shutil
shutil.copyfile(data['img_path'],'/data/PaddleOCR2.5/temp_border_map/' + os.path.splitext(os.path.basename(data['img_path']))[0]+'_'+'.png')
data['threshold_map'] = canvas
data['threshold_mask'] = mask
return data
对应的代码：
位置：PaddleOCR2.5ppocrdataimaugmake_border_map.py
the other
