-
Notifications
You must be signed in to change notification settings - Fork 2
/
TMEndToEndTest.py
224 lines (223 loc) · 11.6 KB
/
TMEndToEndTest.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
219
220
221
222
223
224
import torch
from torch.utils.data import DataLoader
import torch.nn as nn
import pdb
import os
import json
from typing import List
from torch.autograd import Variable
from IPython.display import clear_output
from datetime import datetime
from torchvision.transforms import transforms
from TMDetecter.TMDetectDataSet import FBDataSet,img_transform,load_img
from TMTextLine.TMTextLineDataSet import TMTextLineDataSet
from Logging import *
from TMDetecter.TMDetectUtils import fovea2boxes,box_nms
from TMUtils import TMcrop_img
from TMDetecter.TMRPN import TMRPN
from TMTextLine.TMTextLineNN import ResNetLSTM,VGGLSTM,VGGFC,DenseFC,DenseLSTM
from TMTextLine.TMTextLineTest import condense,distill_condense
from TMEndToEndConfigure import *
CFG=TMETEcfg()
@torch.no_grad()
def end_to_end_test(detect_model:nn.Module=TMRPN(CFG.DETECT_NUM_CLASS),
detect_dataset=FBDataSet,
recognize_model:nn.Module=VGGLSTM(CFG.RECOGNIZE_NUM_CLASS),
cfg=CFG):
if cfg.DEVICE=='cuda':
if torch.cuda.is_available()==False:
logging.error("can't find a GPU device")
pdb.set_trace()
if os.path.exists(cfg.PATH+cfg.DETECT_MODEL_NAME)==False:
logging.error("can't find a pretrained model:{}".format(cfg.PATH+cfg.DETECT_MODEL_NAME))
pdb.set_trace()
if os.path.exists(cfg.PATH+cfg.RECOGNIZE_MODEL_NAME)==False:
logging.error("can't find a pretrained model:{}".format(cfg.PATH+cfg.RECOGNIZE_MODEL_NAME))
pdb.set_trace()
if os.path.exists(cfg.PATH+cfg.DICTIONARY_NAME)==False:
logging.error("can't find the dictionary{}".format(cfg.PATH+cfg.DICTIONARY_NAME))
pdb.set_trace()
with open(cfg.PATH+cfg.DICTIONARY_NAME,'r') as f:
dictionary_inv=json.load(f)
cfg.__setattr__('dictionary_inv',dictionary_inv)
detect_model.load_state_dict(torch.load(cfg.PATH+cfg.DETECT_MODEL_NAME,map_location=cfg.DEVICE))
recognize_model.load_state_dict(torch.load(cfg.PATH+cfg.RECOGNIZE_MODEL_NAME,map_location=cfg.DEVICE))
detect_model.to(cfg.DEVICE).eval()
recognize_model.to(cfg.DEVICE).eval()
dataset = detect_dataset(cfg.TEST_IMAGE_PATH, expected_img_size=cfg.DETECT_EXPECTED_IMG_SIZE,
img_transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))]),
cfg=cfg, train=False)
dataloader=DataLoader(dataset,batch_size=cfg.BATCH_SIZE,num_workers=4,collate_fn=dataset.collate)
length = len(dataloader)
start_time = datetime.now()
f = open(cfg.OUTPUT_PATH+'cannot_detect.txt', 'w')
r_e_w,r_e_h=cfg.RECOGNIZE_EXPECTED_IMG_SIZE
multiscale_imgs_name=[]
img2tensor = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])
for step, data in enumerate(dataloader):
step_time = datetime.now()
imgs, imgs_name, min_ratioes, images = data
imgs = Variable(imgs, requires_grad=False).to(cfg.DEVICE)
batch_size = imgs.size(0)
h, w = imgs.size(2), imgs.size(3)
loc_preds, cls_preds = detect_model(imgs)
batch_cls, batch_score, batch_coordinate = fovea2boxes(loc_preds.cpu(), cls_preds.cpu(), torch.Tensor([w, h]),
tuple_l=cfg.TUPLE_L)
logging.debug("length:{}|step:{}|imgs_name:{}|memory:{:.3f}MB".format(length, step, imgs_name,torch.cuda.max_memory_allocated()/1024/1024))
crop_imgs,crop_imgs_name=[],[]
for b in range(batch_size):
cls, score, coordinate = batch_cls[b], batch_score[b], batch_coordinate[b]
cls_list, score_list, coordinate_list = [], [], []
for num_class in range(1, cfg.DETECT_NUM_CLASS + 1):
num_class_index = (cls == num_class)
cls_class, score_class, coordinate_class = cls[num_class_index], score[num_class_index], coordinate[
num_class_index]
keep = box_nms(coordinate_class, score_class, threshold=0.3)
cls_class, score_class, coordinate_class = cls_class[keep], score_class[keep], coordinate_class[keep]
coordinate_class /= min_ratioes[b]
coordinate_class[:, [0, 1]].floor_()
coordinate_class[:, [2, 3]].ceil_()
if len(score_class) == 0:
multiscale_imgs_name.append(imgs_name[b])
f.write(imgs_name[b] + '\n')
continue
_, index = score_class.max(0)
"""start cropping"""
crop_img=TMcrop_img(images[b], coordinate_class[index].tolist(), img_name=imgs_name[b],save=False, path=cfg.CROP_PATH)
#assert cfg.LOSS in ('CTC', 'ECP'), 'cfg.LOSS must be \'CTC\' or \'ECP\',but got :{}'.format(cfg.LOSS)
if cfg.LOSS=='CTC':
crop_img=TMTextLineDataSet.resize_img(crop_img,r_e_w,r_e_h)
crop_img=TMTextLineDataSet.pad_img(crop_img,(r_e_w,r_e_h))
else:
crop_img=crop_img.resize((r_e_w,r_e_h))
crop_img=img2tensor(crop_img)
crop_imgs.append(crop_img)
crop_imgs_name.append(imgs_name[b])
"""start recognizing"""
recognizer(recognize_model,crop_imgs,crop_imgs_name,dictionary_inv,cfg)
if step % 200 == 0:
clear_output(wait=True)
logging.debug("step_time cost :{}".format(datetime.now() - step_time))
"""multiscale test to detecte the image that cannot be detected by single scale"""
logging.info("finshed and single detect cost of time is :{}|".format(datetime.now() - start_time))
if len(multiscale_imgs_name)>1:
multiscale_test(multiscale_imgs_name, detect_model,recognize_model=recognize_model, cfg=cfg)
f.close()
logging.info("finshed and total cost of time is :{}|".format(datetime.now() - start_time))
def recognizer(model:nn.Module,crop_imgs:List,names:List,dictionary_inv:dict,cfg):
f=open(cfg.OUTPUT_PATH+'submission.txt','a+')
imgs=torch.stack(crop_imgs,dim=0)
imgs = Variable(imgs).to(cfg.DEVICE)
preds = model(imgs)
#assert cfg.LOSS in ('CTC','ECP'),'cfg.LOSS must be \'CTC\' or \'ECP\',but got :{}'.format(cfg.LOSS)
if cfg.LOSS=='CTC':
preds = preds.permute(1, 0, 2)
batch_size = preds.size(0)
preds = preds.cpu()
_, preds = preds.max(2)
for i in range(batch_size):
pred, _ = condense(preds[i])
if len(pred) > 10:
distill_condense(pred)
pred_str = []
for p in pred:
s = dictionary_inv.get(str(p))
pred_str.append(s)
pred_str = ''.join(pred_str)
if len(pred_str) == 0:
pred_str = '1'
logging.info("image's name:{}|predicting character:{}".format(names[i], pred_str))
name = names[i]
f.write(name + ',' + pred_str + '\n')
else:
preds = preds.cpu()
_,preds=preds.max(-1)
pred_str=[]
for i,p in enumerate(preds):
s=dictionary_inv.get(str(p.item()+1))
if i!=0 and i%10==0 :
pred_str=''.join(pred_str)
index=i//10
logging.info("image's name:{}|predicting character:{}".format(names[index-1], pred_str))
f.write(names[index-1] + ',' + pred_str + '\n')
pred_str=[]
if i==len(preds)-1:
pred_str.append(s)
pred_str=''.join(pred_str)
index=(i+1)//10
logging.info("image's name:{}|predicting character:{}".format(names[index - 1], pred_str))
f.write(names[index - 1] + ',' + pred_str + '\n')
break
pred_str.append(s)
logging.info('recognizer ended')
f.close()
@torch.no_grad()
def multiscale_test(imgs_name: List, model: nn.Module,recognize_model, cfg=TMETEcfg()): # 148KLEW0:fliped
img2tensor = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])
f = open(cfg.OUTPUT_PATH+'multi_cannot_detect.txt', 'w')
logging.info("starting multiscale test")
r_e_w,r_e_h=cfg.RECOGNIZE_EXPECTED_IMG_SIZE
for i, img_name in enumerate(imgs_name):
try:
image = load_img(cfg.TEST_IMAGE_PATH + img_name)
except FileNotFoundError:
break
flag = False
break_time = 0
for j, scale in enumerate(cfg.MULTISCALE_SIZE):
if flag == True:
break
# image = load_img(cfg.IMAGE_PATH + img_name)
logging.debug("length:{}|step:{}|img_name:{}|scale:{}".format(len(imgs_name), i, img_name, scale))
img, min_ratio = img_transform(image, scale)
# w,h=img.size
# pad=transforms.Pad(padding=(0,0,scale[0]-w,scale[1]-h),fill=0)
# img=pad(img_)
img_tensor = img2tensor(img)
h, w = img_tensor.size(1), img_tensor.size(2)
loc_preds, cls_preds = model(img_tensor.unsqueeze(0).to(cfg.DEVICE))
batch_cls, batch_score, batch_coordinate = fovea2boxes(loc_preds.cpu(), cls_preds.cpu(),
torch.Tensor([w, h]), tuple_l=cfg.TUPLE_L)
cls, score, coordinate = batch_cls[0], batch_score[0], batch_coordinate[0]
for num_class in range(1, cfg.DETECT_NUM_CLASS + 1):
num_class_index = (cls == num_class)
cls_class, score_class, coordinate_class = cls[num_class_index], score[num_class_index], coordinate[
num_class_index]
if len(cls_class) == 0:
break_time += 1
continue
print('cls_class:{}'.format(cls_class))
keep = box_nms(coordinate_class, score_class, threshold=0.3)
cls_class, score_class, coordinate_class = cls_class[keep], score_class[keep], coordinate_class[keep]
print('coordinate:{}'.format(coordinate_class))
coordinate_class /= min_ratio
coordinate_class[:, [0, 1]] = coordinate_class[:, [0, 1]].floor()
coordinate_class[:, [2, 3]] = coordinate_class[:, [2, 3]].ceil()
print('coordinate:{}'.format(coordinate_class))
if len(score_class) == 0:
break_time += 1
break
_, index = score_class.max(0)
print('index:{}|coordinate_classs:{}'.format(index, coordinate_class[index]))
# TMcrop_img(image,temp,img_name=img_name,path=cfg.CROP_PATH)
crop_img=TMcrop_img(image, coordinate_class[index], img_name=img_name, save=False,path=cfg.CROP_PATH_COMPLEMENT)
if cfg.LOSS=='CTC':
crop_img = TMTextLineDataSet.resize_img(crop_img, r_e_w, r_e_h)
crop_img = TMTextLineDataSet.pad_img(crop_img, (r_e_w, r_e_h))
else:
crop_img = crop_img.resize((r_e_w,r_e_h))
crop_img=img2tensor(crop_img)
recognizer(recognize_model, [crop_img], [img_name], cfg.dictionary_inv, cfg)
flag = True
if break_time == len(cfg.MULTISCALE_SIZE):
f.write(img_name + '\n')
# image.show(title='test')
# break
logging.info("flag:{}".format(break_time))
f.close()
if __name__=='__main__':
end_to_end_test(recognize_model=VGGLSTM(CFG.RECOGNIZE_NUM_CLASS,NonLocal=False))