forked from SURFZJY/EAST-caffe
-
Notifications
You must be signed in to change notification settings - Fork 0
/
pylayerUtils.py
218 lines (178 loc) · 9.25 KB
/
pylayerUtils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
import caffe
import numpy as np
import cv2
import random
import os
import math
import csv
from icdar import get_whole_data
class DataLayer(caffe.Layer):
def setup(self, bottom, top):
# data layer config
params = eval(self.param_str)
self.data_dir = params['data_dir']
self.dataset = params['dataset']
self.patch_size = int(params['patch_size'])
self.seed = params['seed']
self.batch_size = int(params['batch_size'])
self.mean = np.array(params['mean'])
self.random = True
# three tops: data, label and weight
# if len(top) != 2:
# raise Exception("Need to define three tops: data, label and weight.")
if len(bottom) != 0:
raise Exception("Do not define a bottom.")
# set directory for each dataset here
if self.dataset == 'synth':
self.fnLst = open(self.data_dir+'SynthText/list_train.txt').readlines()
elif self.dataset == 'ic15':
self.fnLst = os.listdir(self.data_dir+'ic15/train_images/')
elif self.dataset == 'invoice':
self.fnLst = os.listdir(self.data_dir+'taxi1200_segall/train_images/')
else:
raise Exception("Invalid dataset.")
# randomization: seed and pick
self.idx = 0
if self.random:
random.seed(self.seed)
self.idx = random.randint(0, len(self.fnLst)-1)
def reshape(self, bottom, top):
# load image, label and weight
if self.dataset == 'synth':
self.data, self.label, self.weight = self.loadsynth(self.fnLst[self.idx])
elif self.dataset == 'ic15':
self.data, self.score_map, self.geo_map = self.loadic15(self.fnLst)
elif self.dataset == 'invoice':
self.data, self.score_map, self.geo_map = self.loadinvoice(self.fnLst)
else:
raise Exception("Invalid dataset.")
# reshape tops to fit (leading 1 is for batch dimension)
top[0].reshape(*self.data.shape)
top[1].reshape(*self.score_map.shape)
top[2].reshape(*self.geo_map.shape)
# top[3].reshape(1, *self.weight.shape)
def forward(self, bottom, top):
# assign output
top[0].data[...] = self.data
top[1].data[...] = self.score_map
top[2].data[...] = self.geo_map
#top[3].data[...] = self.weight
# pick next input
if self.random:
self.idx = random.randint(0, len(self.fnLst)-1)
else:
self.idx += 1
if self.idx == len(self.fnLst):
self.idx = 0
def backward(self, top, propagate_down, bottom):
pass
def loadinvoice(self, fnLst):
basedir = '{}/taxi1200_segall/train_images'.format(self.data_dir)
whole_data = get_whole_data(input_size=self.patch_size,
batch_size=self.batch_size,
basedir=basedir,
image_list=fnLst)
input_images = (np.array(whole_data[0])).transpose(0,3,1,2)
input_score_maps = (np.array(whole_data[2])).transpose(0,3,1,2)
input_geo_maps = (np.array(whole_data[3])).transpose(0,3,1,2)
input_training_masks = (np.array(whole_data[4])).transpose(0,3,1,2)
return input_images,input_score_maps,input_geo_maps
def loadic15(self, fnLst):
basedir = '{}/ic15/train_images'.format(self.data_dir)
whole_data = get_whole_data(input_size=self.patch_size,
batch_size=self.batch_size,
basedir=basedir,
image_list=fnLst)
input_images = (np.array(whole_data[0])).transpose(0,3,1,2)
input_score_maps = (np.array(whole_data[2])).transpose(0,3,1,2)
input_geo_maps = (np.array(whole_data[3])).transpose(0,3,1,2)
input_training_masks = (np.array(whole_data[4])).transpose(0,3,1,2)
return input_images,input_score_maps,input_geo_maps
class DiceCoefLossLayer(caffe.Layer):
"""
self designed loss layer for segmentation. Class weighted, per pixel loss
"""
def setup(self, bottom, top):
# check input pair
if len(bottom) != 2:
raise Exception("Need two inputs to compute distance.")
self.batch_size = bottom[1].data.shape[0]
# print('bottom[1].data.shape=', bottom[1].data.shape) # output = (4, 1, 128, 128)
def reshape(self, bottom, top):
# check input dimensions match
# print(bottom[0].count) # output=65536 = (1*4*128*128 )
# print(bottom[1].count) # output=65536 = (1*4*128*128 )
if bottom[0].count!=bottom[1].count:
raise Exception("Inputs must have the same dimension.")
self.diff=np.zeros_like(bottom[0].data,dtype=np.float32)
# loss output is scalar
top[0].reshape(1)
def forward(self, bottom, top):
self.diff[...]=bottom[1].data
self.sum=bottom[0].data.sum()+bottom[1].data.sum()+1.
self.dice=(2.* (bottom[0].data * bottom [1].data).sum()+1.)/self.sum
top[0].data[...] = 1.- self.dice
def backward(self, top, propagate_down, bottom):
if propagate_down[1]:
raise Exception("label not diff")
elif propagate_down[0]:
bottom[0].diff[...] = (-2.*self.diff + 2.*bottom[0].data*self.dice) / self.sum
else:
raise Exception("no diff")
class RBoxLossLayer(caffe.Layer):
"""
self designed loss layer for segmentation. Class weighted, per pixel loss
bottom: "concat4" # 1*5*128*128 (4+1)
bottom: "geo_map" # 1*5*128*128 (4+1)
bottom: "score_map" # scoremap的ture_gt 1*1*128*128
"""
def setup(self, bottom, top):
# check input pair
if len(bottom) != 3:
raise Exception("Need three inputs to compute total Loss.")
self.batch_size = bottom[1].data.shape[0]
self.pixel_num = bottom[1].data.shape[2] * bottom[1].data.shape[3]
self.ratio = 20.
def reshape(self, bottom, top):
# check input dimensions match
# print(bottom[0].data.shape) # N*5*128*128 pred
# print(bottom[1].data.shape) # N*5*128*128 geo_gt
# print(bottom[2].data.shape) # N*1*128*128 score_gt
if bottom[0].count!=bottom[1].count:
raise Exception("First Two Inputs must have the same dimension.")
self.score_gt = np.zeros_like(bottom[2].data,dtype=np.float32)
self.L_g = np.zeros_like(bottom[0].data[:,0,:,:],dtype=np.float32)
self.top_grad1 = np.zeros_like(bottom[2].data,dtype=np.float32)
self.top_grad2 = np.zeros_like(bottom[2].data,dtype=np.float32)
self.L_theta_grad = np.zeros_like(bottom[2].data,dtype=np.float32)
# loss output is scalar
top[0].reshape(1) # 1,
def forward(self, bottom, top):
self.score_gt[...] = bottom[2].data
self.d1_pred, self.d2_pred, self.d3_pred, self.d4_pred, self.theta_pred = np.array_split(bottom[0].data, indices_or_sections=5, axis=1)
self.d1_gt, self.d2_gt, self.d3_gt, self.d4_gt, self.theta_gt = np.array_split(bottom[1].data, indices_or_sections=5, axis=1)
area_gt = (self.d1_gt + self.d3_gt) * (self.d2_gt + self.d4_gt)
area_pred = (self.d1_pred + self.d3_pred) * (self.d2_pred + self.d4_pred)
self.w_union = np.minimum(self.d2_gt, self.d2_pred) + np.minimum(self.d4_gt, self.d4_pred)
self.h_union = np.minimum(self.d1_gt, self.d1_pred) + np.minimum(self.d3_gt, self.d3_pred)
self.area_intersect = self.w_union * self.h_union
self.area_union = area_gt + area_pred - self.area_intersect
L_theta = 1. - np.cos(self.theta_pred - self.theta_gt)
L_AABB = -np.log((self.area_intersect + 1.)/(self.area_union + 1.))
self.L_g = L_AABB + self.ratio * L_theta
top[0].data[...] = np.mean(self.L_g * self.score_gt)
def backward(self, top, propagate_down, bottom):
ai_grad1 = self.w_union * (1.*(self.d1_pred<=self.d1_gt))
self.top_grad1 = self.score_gt / self.pixel_num / self.batch_size * ((self.d2_pred+self.d4_pred-ai_grad1)/(self.area_union+1.) - ai_grad1/(self.area_intersect+1.))
ai_grad2 = self.h_union * (1.*(self.d2_pred<=self.d2_gt))
self.top_grad2 = self.score_gt / self.pixel_num / self.batch_size * ((self.d1_pred+self.d3_pred-ai_grad2)/(self.area_union+1.) - ai_grad2/(self.area_intersect+1.))
ai_grad3 = self.w_union * (1.*(self.d3_pred<=self.d3_gt))
self.top_grad3 = self.score_gt / self.pixel_num / self.batch_size * ((self.d2_pred+self.d4_pred-ai_grad3)/(self.area_union+1.) - ai_grad3/(self.area_intersect+1.))
ai_grad4 = self.h_union * (1.*(self.d4_pred<=self.d4_gt))
self.top_grad4 = self.score_gt / self.pixel_num / self.batch_size * ((self.d1_pred+self.d3_pred-ai_grad4)/(self.area_union+1.) - ai_grad4/(self.area_intersect+1.))
self.L_theta_grad = self.ratio * self.score_gt / self.pixel_num / self.batch_size * np.sin(self.theta_pred - self.theta_gt)
bottom[0].diff[...] = np.concatenate((self.top_grad1, self.top_grad2, self.top_grad3, self.top_grad4, self.L_theta_grad), axis=1)
bottom[1].diff[...] = 0
bottom[2].diff[...] = 0