predict_rec.py 12.7 KB
Newer Older
LDOUBLEV's avatar
LDOUBLEV committed
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
LDOUBLEV's avatar
LDOUBLEV committed
14
15
import os
import sys
Topdu's avatar
Topdu committed
16
from PIL import Image
17
__dir__ = os.path.dirname(os.path.abspath(__file__))
LDOUBLEV's avatar
LDOUBLEV committed
18
sys.path.append(__dir__)
19
sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
LDOUBLEV's avatar
LDOUBLEV committed
20

21
22
os.environ["FLAGS_allocator_strategy"] = 'auto_growth'

LDOUBLEV's avatar
LDOUBLEV committed
23
24
25
26
import cv2
import numpy as np
import math
import time
WenmuZhou's avatar
WenmuZhou committed
27
import traceback
tink2123's avatar
tink2123 committed
28
import paddle
29
30

import tools.infer.utility as utility
WenmuZhou's avatar
WenmuZhou committed
31
32
from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger
33
from ppocr.utils.utility import get_image_file_list, check_and_read_gif
LDOUBLEV's avatar
LDOUBLEV committed
34

WenmuZhou's avatar
WenmuZhou committed
35
36
logger = get_logger()

LDOUBLEV's avatar
LDOUBLEV committed
37
38
39

class TextRecognizer(object):
    def __init__(self, args):
40
        self.rec_image_shape = [int(v) for v in args.rec_image_shape.split(",")]
dyning's avatar
dyning committed
41
        self.character_type = args.rec_char_type
42
        self.rec_batch_num = args.rec_batch_num
tink2123's avatar
tink2123 committed
43
        self.rec_algorithm = args.rec_algorithm
WenmuZhou's avatar
WenmuZhou committed
44
45
        postprocess_params = {
            'name': 'CTCLabelDecode',
tink2123's avatar
tink2123 committed
46
            "character_type": args.rec_char_type,
47
            "character_dict_path": args.rec_char_dict_path,
WenmuZhou's avatar
WenmuZhou committed
48
            "use_space_char": args.use_space_char
tink2123's avatar
tink2123 committed
49
        }
tink2123's avatar
tink2123 committed
50
51
52
        if self.rec_algorithm == "SRN":
            postprocess_params = {
                'name': 'SRNLabelDecode',
WenmuZhou's avatar
WenmuZhou committed
53
54
55
56
57
58
59
                "character_type": args.rec_char_type,
                "character_dict_path": args.rec_char_dict_path,
                "use_space_char": args.use_space_char
            }
        elif self.rec_algorithm == "RARE":
            postprocess_params = {
                'name': 'AttnLabelDecode',
tink2123's avatar
tink2123 committed
60
61
62
63
                "character_type": args.rec_char_type,
                "character_dict_path": args.rec_char_dict_path,
                "use_space_char": args.use_space_char
            }
Topdu's avatar
Topdu committed
64
65
66
67
68
69
70
        elif self.rec_algorithm == 'NRTR':
            postprocess_params = {
                'name': 'NRTRLabelDecode',
                "character_type": args.rec_char_type,
                "character_dict_path": args.rec_char_dict_path,
                "use_space_char": args.use_space_char
            }
WenmuZhou's avatar
WenmuZhou committed
71
        self.postprocess_op = build_post_process(postprocess_params)
LDOUBLEV's avatar
LDOUBLEV committed
72
        self.predictor, self.input_tensor, self.output_tensors, self.config = \
WenmuZhou's avatar
WenmuZhou committed
73
            utility.create_predictor(args, 'rec', logger)
tink2123's avatar
tink2123 committed
74
75
76
77
        self.benchmark = args.benchmark
        if args.benchmark:
            import auto_log
            pid = os.getpid()
LDOUBLEV's avatar
LDOUBLEV committed
78
            gpu_id = utility.get_infer_gpuid()
tink2123's avatar
tink2123 committed
79
80
81
            self.autolog = auto_log.AutoLogger(
                model_name="rec",
                model_precision=args.precision,
tink2123's avatar
tink2123 committed
82
                batch_size=args.rec_batch_num,
tink2123's avatar
tink2123 committed
83
                data_shape="dynamic",
84
                save_path=None,  #args.save_log_path,
tink2123's avatar
tink2123 committed
85
86
87
                inference_config=self.config,
                pids=pid,
                process_name=None,
LDOUBLEV's avatar
LDOUBLEV committed
88
                gpu_ids=gpu_id if args.use_gpu else None,
tink2123's avatar
tink2123 committed
89
90
91
                time_keys=[
                    'preprocess_time', 'inference_time', 'postprocess_time'
                ],
92
93
                warmup=2,
                logger=logger)
LDOUBLEV's avatar
LDOUBLEV committed
94

95
    def resize_norm_img(self, img, max_wh_ratio):
LDOUBLEV's avatar
LDOUBLEV committed
96
        imgC, imgH, imgW = self.rec_image_shape
Topdu's avatar
Topdu committed
97
        if self.rec_algorithm == 'NRTR':
Topdu's avatar
Topdu committed
98
99
100
101
102
103
104
105
106
            img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
            # return padding_im
            image_pil = Image.fromarray(np.uint8(img))
            img = image_pil.resize([100, 32], Image.ANTIALIAS)
            img = np.array(img)
            norm_img = np.expand_dims(img, -1)
            norm_img = norm_img.transpose((2, 0, 1))
            return norm_img.astype(np.float32) / 128. - 1.

107
        assert imgC == img.shape[2]
tink2123's avatar
tink2123 committed
108
109
        max_wh_ratio = max(max_wh_ratio, imgW / imgH)
        imgW = int((32 * max_wh_ratio))
110
        h, w = img.shape[:2]
111
112
113
114
115
        ratio = w / float(h)
        if math.ceil(imgH * ratio) > imgW:
            resized_w = imgW
        else:
            resized_w = int(math.ceil(imgH * ratio))
tink2123's avatar
tink2123 committed
116
        resized_image = cv2.resize(img, (resized_w, imgH))
LDOUBLEV's avatar
LDOUBLEV committed
117
118
119
120
121
122
123
124
        resized_image = resized_image.astype('float32')
        resized_image = resized_image.transpose((2, 0, 1)) / 255
        resized_image -= 0.5
        resized_image /= 0.5
        padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32)
        padding_im[:, :, 0:resized_w] = resized_image
        return padding_im

tink2123's avatar
tink2123 committed
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
    def resize_norm_img_srn(self, img, image_shape):
        imgC, imgH, imgW = image_shape

        img_black = np.zeros((imgH, imgW))
        im_hei = img.shape[0]
        im_wid = img.shape[1]

        if im_wid <= im_hei * 1:
            img_new = cv2.resize(img, (imgH * 1, imgH))
        elif im_wid <= im_hei * 2:
            img_new = cv2.resize(img, (imgH * 2, imgH))
        elif im_wid <= im_hei * 3:
            img_new = cv2.resize(img, (imgH * 3, imgH))
        else:
            img_new = cv2.resize(img, (imgW, imgH))

        img_np = np.asarray(img_new)
        img_np = cv2.cvtColor(img_np, cv2.COLOR_BGR2GRAY)
        img_black[:, 0:img_np.shape[1]] = img_np
        img_black = img_black[:, :, np.newaxis]

        row, col, c = img_black.shape
        c = 1

        return np.reshape(img_black, (c, row, col)).astype(np.float32)

    def srn_other_inputs(self, image_shape, num_heads, max_text_length):

        imgC, imgH, imgW = image_shape
        feature_dim = int((imgH / 8) * (imgW / 8))

        encoder_word_pos = np.array(range(0, feature_dim)).reshape(
            (feature_dim, 1)).astype('int64')
        gsrm_word_pos = np.array(range(0, max_text_length)).reshape(
            (max_text_length, 1)).astype('int64')

        gsrm_attn_bias_data = np.ones((1, max_text_length, max_text_length))
        gsrm_slf_attn_bias1 = np.triu(gsrm_attn_bias_data, 1).reshape(
            [-1, 1, max_text_length, max_text_length])
        gsrm_slf_attn_bias1 = np.tile(
            gsrm_slf_attn_bias1,
            [1, num_heads, 1, 1]).astype('float32') * [-1e9]

        gsrm_slf_attn_bias2 = np.tril(gsrm_attn_bias_data, -1).reshape(
            [-1, 1, max_text_length, max_text_length])
        gsrm_slf_attn_bias2 = np.tile(
            gsrm_slf_attn_bias2,
            [1, num_heads, 1, 1]).astype('float32') * [-1e9]

        encoder_word_pos = encoder_word_pos[np.newaxis, :]
        gsrm_word_pos = gsrm_word_pos[np.newaxis, :]

        return [
            encoder_word_pos, gsrm_word_pos, gsrm_slf_attn_bias1,
            gsrm_slf_attn_bias2
        ]

    def process_image_srn(self, img, image_shape, num_heads, max_text_length):
        norm_img = self.resize_norm_img_srn(img, image_shape)
        norm_img = norm_img[np.newaxis, :]

        [encoder_word_pos, gsrm_word_pos, gsrm_slf_attn_bias1, gsrm_slf_attn_bias2] = \
            self.srn_other_inputs(image_shape, num_heads, max_text_length)

        gsrm_slf_attn_bias1 = gsrm_slf_attn_bias1.astype(np.float32)
        gsrm_slf_attn_bias2 = gsrm_slf_attn_bias2.astype(np.float32)
        encoder_word_pos = encoder_word_pos.astype(np.int64)
        gsrm_word_pos = gsrm_word_pos.astype(np.int64)

        return (norm_img, encoder_word_pos, gsrm_word_pos, gsrm_slf_attn_bias1,
                gsrm_slf_attn_bias2)

LDOUBLEV's avatar
LDOUBLEV committed
197
198
    def __call__(self, img_list):
        img_num = len(img_list)
199
        # Calculate the aspect ratio of all text bars
200
201
202
        width_list = []
        for img in img_list:
            width_list.append(img.shape[1] / float(img.shape[0]))
zhangxin's avatar
zhangxin committed
203
        # Sorting can speed up the recognition process
204
205
        indices = np.argsort(np.array(width_list))
        rec_res = [['', 0.0]] * img_num
206
        batch_num = self.rec_batch_num
LDOUBLEV's avatar
LDOUBLEV committed
207
        st = time.time()
tink2123's avatar
tink2123 committed
208
209
        if self.benchmark:
            self.autolog.times.start()
LDOUBLEV's avatar
LDOUBLEV committed
210
211
212
        for beg_img_no in range(0, img_num, batch_num):
            end_img_no = min(img_num, beg_img_no + batch_num)
            norm_img_batch = []
213
            max_wh_ratio = 0
LDOUBLEV's avatar
LDOUBLEV committed
214
            for ino in range(beg_img_no, end_img_no):
215
                h, w = img_list[indices[ino]].shape[0:2]
216
217
218
                wh_ratio = w * 1.0 / h
                max_wh_ratio = max(max_wh_ratio, wh_ratio)
            for ino in range(beg_img_no, end_img_no):
tink2123's avatar
tink2123 committed
219
220
221
222
223
224
                if self.rec_algorithm != "SRN":
                    norm_img = self.resize_norm_img(img_list[indices[ino]],
                                                    max_wh_ratio)
                    norm_img = norm_img[np.newaxis, :]
                    norm_img_batch.append(norm_img)
                else:
LDOUBLEV's avatar
LDOUBLEV committed
225
226
                    norm_img = self.process_image_srn(
                        img_list[indices[ino]], self.rec_image_shape, 8, 25)
tink2123's avatar
tink2123 committed
227
228
229
230
231
232
233
234
235
                    encoder_word_pos_list = []
                    gsrm_word_pos_list = []
                    gsrm_slf_attn_bias1_list = []
                    gsrm_slf_attn_bias2_list = []
                    encoder_word_pos_list.append(norm_img[1])
                    gsrm_word_pos_list.append(norm_img[2])
                    gsrm_slf_attn_bias1_list.append(norm_img[3])
                    gsrm_slf_attn_bias2_list.append(norm_img[4])
                    norm_img_batch.append(norm_img[0])
LDOUBLEV's avatar
LDOUBLEV committed
236
237
            norm_img_batch = np.concatenate(norm_img_batch)
            norm_img_batch = norm_img_batch.copy()
tink2123's avatar
tink2123 committed
238
239
            if self.benchmark:
                self.autolog.times.stamp()
tink2123's avatar
tink2123 committed
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265

            if self.rec_algorithm == "SRN":
                encoder_word_pos_list = np.concatenate(encoder_word_pos_list)
                gsrm_word_pos_list = np.concatenate(gsrm_word_pos_list)
                gsrm_slf_attn_bias1_list = np.concatenate(
                    gsrm_slf_attn_bias1_list)
                gsrm_slf_attn_bias2_list = np.concatenate(
                    gsrm_slf_attn_bias2_list)

                inputs = [
                    norm_img_batch,
                    encoder_word_pos_list,
                    gsrm_word_pos_list,
                    gsrm_slf_attn_bias1_list,
                    gsrm_slf_attn_bias2_list,
                ]
                input_names = self.predictor.get_input_names()
                for i in range(len(input_names)):
                    input_tensor = self.predictor.get_input_handle(input_names[
                        i])
                    input_tensor.copy_from_cpu(inputs[i])
                self.predictor.run()
                outputs = []
                for output_tensor in self.output_tensors:
                    output = output_tensor.copy_to_cpu()
                    outputs.append(output)
tink2123's avatar
tink2123 committed
266
267
                if self.benchmark:
                    self.autolog.times.stamp()
tink2123's avatar
tink2123 committed
268
269
270
271
272
273
274
275
                preds = {"predict": outputs[2]}
            else:
                self.input_tensor.copy_from_cpu(norm_img_batch)
                self.predictor.run()
                outputs = []
                for output_tensor in self.output_tensors:
                    output = output_tensor.copy_to_cpu()
                    outputs.append(output)
tink2123's avatar
tink2123 committed
276
277
                if self.benchmark:
                    self.autolog.times.stamp()
Topdu's avatar
Topdu committed
278
279
280
281
                if len(outputs) != 1:
                    preds = outputs
                else:
                    preds = outputs[0]
WenmuZhou's avatar
WenmuZhou committed
282
283
284
            rec_result = self.postprocess_op(preds)
            for rno in range(len(rec_result)):
                rec_res[indices[beg_img_no + rno]] = rec_result[rno]
tink2123's avatar
tink2123 committed
285
286
            if self.benchmark:
                self.autolog.times.end(stamp=True)
LDOUBLEV's avatar
LDOUBLEV committed
287
        return rec_res, time.time() - st
LDOUBLEV's avatar
LDOUBLEV committed
288
289


290
def main(args):
dyning's avatar
dyning committed
291
    image_file_list = get_image_file_list(args.image_dir)
LDOUBLEV's avatar
LDOUBLEV committed
292
293
294
    text_recognizer = TextRecognizer(args)
    valid_image_file_list = []
    img_list = []
LDOUBLEV's avatar
LDOUBLEV committed
295

296
    # warmup 2 times
LDOUBLEV's avatar
LDOUBLEV committed
297
298
    if args.warmup:
        img = np.random.uniform(0, 255, [32, 320, 3]).astype(np.uint8)
299
        for i in range(2):
LDOUBLEV's avatar
LDOUBLEV committed
300
            res = text_recognizer([img] * int(args.rec_batch_num))
LDOUBLEV's avatar
LDOUBLEV committed
301

LDOUBLEV's avatar
LDOUBLEV committed
302
    for image_file in image_file_list:
LDOUBLEV's avatar
LDOUBLEV committed
303
304
305
        img, flag = check_and_read_gif(image_file)
        if not flag:
            img = cv2.imread(image_file)
LDOUBLEV's avatar
LDOUBLEV committed
306
307
308
309
310
        if img is None:
            logger.info("error in loading image:{}".format(image_file))
            continue
        valid_image_file_list.append(image_file)
        img_list.append(img)
LDOUBLEV's avatar
LDOUBLEV committed
311
312
313
314
315
316
317
318
319
320
    try:
        rec_res, _ = text_recognizer(img_list)

    except Exception as E:
        logger.info(traceback.format_exc())
        logger.info(E)
        exit()
    for ino in range(len(img_list)):
        logger.info("Predicts of {}:{}".format(valid_image_file_list[ino],
                                               rec_res[ino]))
tink2123's avatar
tink2123 committed
321
322
    if args.benchmark:
        text_recognizer.autolog.report()
323
324
325
326


if __name__ == "__main__":
    main(utility.parse_args())