program.py 18.9 KB
Newer Older
M
refine  
MissPenguin 已提交
1
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
L
LDOUBLEV 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
#
# 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.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

W
WenmuZhou 已提交
19
import os
L
LDOUBLEV 已提交
20
import sys
21
import platform
L
LDOUBLEV 已提交
22 23
import yaml
import time
W
WenmuZhou 已提交
24 25 26 27 28 29
import shutil
import paddle
import paddle.distributed as dist
from tqdm import tqdm
from argparse import ArgumentParser, RawDescriptionHelpFormatter

L
LDOUBLEV 已提交
30 31
from ppocr.utils.stats import TrainingStats
from ppocr.utils.save_load import save_model
D
dyning 已提交
32 33
from ppocr.utils.utility import print_dict
from ppocr.utils.logging import get_logger
L
LDOUBLEV 已提交
34
from ppocr.utils import profiler
D
dyning 已提交
35 36
from ppocr.data import build_dataloader
import numpy as np
L
LDOUBLEV 已提交
37

D
dyning 已提交
38

L
LDOUBLEV 已提交
39 40 41 42 43 44 45
class ArgsParser(ArgumentParser):
    def __init__(self):
        super(ArgsParser, self).__init__(
            formatter_class=RawDescriptionHelpFormatter)
        self.add_argument("-c", "--config", help="configuration file to use")
        self.add_argument(
            "-o", "--opt", nargs='+', help="set configuration options")
L
LDOUBLEV 已提交
46 47 48 49 50 51 52
        self.add_argument(
            '-p',
            '--profiler_options',
            type=str,
            default=None,
            help='The option of profiler, which should be in format \"key1=value1;key2=value2;key3=value3\".'
        )
L
LDOUBLEV 已提交
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

    def parse_args(self, argv=None):
        args = super(ArgsParser, self).parse_args(argv)
        assert args.config is not None, \
            "Please specify --config=configure_file_path."
        args.opt = self._parse_opt(args.opt)
        return args

    def _parse_opt(self, opts):
        config = {}
        if not opts:
            return config
        for s in opts:
            s = s.strip()
            k, v = s.split('=')
            config[k] = yaml.load(v, Loader=yaml.Loader)
        return config


def load_config(file_path):
    """
    Load config from yml/yaml file.
    Args:
        file_path (str): Path of the config file to be loaded.
    Returns: global config
    """
    _, ext = os.path.splitext(file_path)
    assert ext in ['.yml', '.yaml'], "only support yaml files for now"
81 82
    config = yaml.load(open(file_path, 'rb'), Loader=yaml.Loader)
    return config
L
LDOUBLEV 已提交
83 84


85
def merge_config(config, opts):
L
LDOUBLEV 已提交
86 87 88 89 90 91
    """
    Merge config into global config.
    Args:
        config (dict): Config to be merged.
    Returns: global config
    """
92
    for key, value in opts.items():
L
LDOUBLEV 已提交
93
        if "." not in key:
94 95
            if isinstance(value, dict) and key in config:
                config[key].update(value)
L
LDOUBLEV 已提交
96
            else:
97
                config[key] = value
L
LDOUBLEV 已提交
98 99
        else:
            sub_keys = key.split('.')
T
tink2123 已提交
100
            assert (
101
                sub_keys[0] in config
T
tink2123 已提交
102
            ), "the sub_keys can only be one of global_config: {}, but get: {}, please check your running command".format(
103 104
                config.keys(), sub_keys[0])
            cur = config[sub_keys[0]]
L
LDOUBLEV 已提交
105 106 107 108 109
            for idx, sub_key in enumerate(sub_keys[1:]):
                if idx == len(sub_keys) - 2:
                    cur[sub_key] = value
                else:
                    cur = cur[sub_key]
110
    return config
L
LDOUBLEV 已提交
111 112 113 114 115 116 117 118 119 120 121 122 123 124


def check_gpu(use_gpu):
    """
    Log error and exit when set use_gpu=true in paddlepaddle
    cpu version.
    """
    err = "Config use_gpu cannot be set as true while you are " \
          "using paddlepaddle cpu version ! \nPlease try: \n" \
          "\t1. Install paddlepaddle-gpu to run model on GPU \n" \
          "\t2. Set use_gpu as false in config file to run " \
          "model on CPU"

    try:
W
WenmuZhou 已提交
125
        if use_gpu and not paddle.is_compiled_with_cuda():
W
WenmuZhou 已提交
126
            print(err)
L
LDOUBLEV 已提交
127 128 129 130 131
            sys.exit(1)
    except Exception as e:
        pass


W
WenmuZhou 已提交
132
def train(config,
D
dyning 已提交
133 134 135
          train_dataloader,
          valid_dataloader,
          device,
W
WenmuZhou 已提交
136 137 138 139 140 141 142 143
          model,
          loss_class,
          optimizer,
          lr_scheduler,
          post_process_class,
          eval_class,
          pre_best_model_dict,
          logger,
S
stephon 已提交
144 145
          vdl_writer=None,
          scaler=None):
W
WenmuZhou 已提交
146 147
    cal_metric_during_train = config['Global'].get('cal_metric_during_train',
                                                   False)
L
LDOUBLEV 已提交
148 149 150 151
    log_smooth_window = config['Global']['log_smooth_window']
    epoch_num = config['Global']['epoch_num']
    print_batch_step = config['Global']['print_batch_step']
    eval_batch_step = config['Global']['eval_batch_step']
L
LDOUBLEV 已提交
152
    profiler_options = config['profiler_options']
W
WenmuZhou 已提交
153

D
dyning 已提交
154
    global_step = 0
155 156
    if 'global_step' in pre_best_model_dict:
        global_step = pre_best_model_dict['global_step']
L
LDOUBLEV 已提交
157 158 159 160
    start_eval_step = 0
    if type(eval_batch_step) == list and len(eval_batch_step) >= 2:
        start_eval_step = eval_batch_step[0]
        eval_batch_step = eval_batch_step[1]
W
WenmuZhou 已提交
161 162 163 164 165
        if len(valid_dataloader) == 0:
            logger.info(
                'No Images in eval dataset, evaluation during training will be disabled'
            )
            start_eval_step = 1e111
L
LDOUBLEV 已提交
166 167 168
        logger.info(
            "During the training process, after the {}th iteration, an evaluation is run every {} iterations".
            format(start_eval_step, eval_batch_step))
L
LDOUBLEV 已提交
169 170
    save_epoch_step = config['Global']['save_epoch_step']
    save_model_dir = config['Global']['save_model_dir']
171 172
    if not os.path.exists(save_model_dir):
        os.makedirs(save_model_dir)
W
WenmuZhou 已提交
173 174 175 176
    main_indicator = eval_class.main_indicator
    best_model_dict = {main_indicator: 0}
    best_model_dict.update(pre_best_model_dict)
    train_stats = TrainingStats(log_smooth_window, ['lr'])
T
tink2123 已提交
177
    model_average = False
W
WenmuZhou 已提交
178 179
    model.train()

T
tink2123 已提交
180
    use_srn = config['Architecture']['algorithm'] == "SRN"
T
tink2123 已提交
181
    extra_input = config['Architecture'][
L
LDOUBLEV 已提交
182
        'algorithm'] in ["SRN", "NRTR", "SAR", "SEED"]
183
    try:
L
fix bug  
LDOUBLEV 已提交
184
        model_type = config['Architecture']['model_type']
185
    except:
L
fix bug  
LDOUBLEV 已提交
186
        model_type = None
T
tink2123 已提交
187
    algorithm = config['Architecture']['algorithm']
T
tink2123 已提交
188

189 190 191 192 193 194 195 196 197 198
    start_epoch = best_model_dict[
        'start_epoch'] if 'start_epoch' in best_model_dict else 1

    train_reader_cost = 0.0
    train_run_cost = 0.0
    total_samples = 0
    reader_start = time.time()

    max_iter = len(train_dataloader) - 1 if platform.system(
    ) == "Windows" else len(train_dataloader)
W
WenmuZhou 已提交
199

T
tink2123 已提交
200
    for epoch in range(start_epoch, epoch_num + 1):
201 202 203 204 205 206
        if train_dataloader.dataset.need_reset:
            train_dataloader = build_dataloader(
                config, 'Train', device, logger, seed=epoch)
            max_iter = len(train_dataloader) - 1 if platform.system(
            ) == "Windows" else len(train_dataloader)

W
WenmuZhou 已提交
207
        for idx, batch in enumerate(train_dataloader):
L
LDOUBLEV 已提交
208
            profiler.add_profiler_step(profiler_options)
文幕地方's avatar
文幕地方 已提交
209
            train_reader_cost += time.time() - reader_start
J
Jane-Ding 已提交
210
            if idx >= max_iter:
W
WenmuZhou 已提交
211 212 213
                break
            lr = optimizer.get_lr()
            images = batch[0]
T
tink2123 已提交
214
            if use_srn:
T
tink2123 已提交
215
                model_average = True
S
stephon 已提交
216

文幕地方's avatar
文幕地方 已提交
217
            train_start = time.time()
S
stephon 已提交
218 219 220 221 222 223 224
            # use amp
            if scaler:
                with paddle.amp.auto_cast():
                    if model_type == 'table' or extra_input:
                        preds = model(images, data=batch[1:])
                    else:
                        preds = model(images)
T
tink2123 已提交
225
            else:
S
stephon 已提交
226 227
                if model_type == 'table' or extra_input:
                    preds = model(images, data=batch[1:])
228
                elif model_type in ["kie", 'vqa']:
L
LDOUBLEV 已提交
229
                    preds = model(batch)
S
stephon 已提交
230 231
                else:
                    preds = model(images)
232

W
WenmuZhou 已提交
233 234
            loss = loss_class(preds, batch)
            avg_loss = loss['loss']
S
stephon 已提交
235 236 237 238 239 240 241 242

            if scaler:
                scaled_avg_loss = scaler.scale(avg_loss)
                scaled_avg_loss.backward()
                scaler.minimize(optimizer, scaled_avg_loss)
            else:
                avg_loss.backward()
                optimizer.step()
W
WenmuZhou 已提交
243
            optimizer.clear_grad()
W
WenmuZhou 已提交
244

文幕地方's avatar
文幕地方 已提交
245
            train_run_cost += time.time() - train_start
246
            global_step += 1
文幕地方's avatar
文幕地方 已提交
247
            total_samples += len(images)
W
WenmuZhou 已提交
248

D
dyning 已提交
249 250
            if not isinstance(lr_scheduler, float):
                lr_scheduler.step()
W
WenmuZhou 已提交
251 252 253 254 255 256

            # logger and visualdl
            stats = {k: v.numpy().mean() for k, v in loss.items()}
            stats['lr'] = lr
            train_stats.update(stats)

L
LDOUBLEV 已提交
257
            if cal_metric_during_train:  # only rec and cls need
W
WenmuZhou 已提交
258
                batch = [item.numpy() for item in batch]
L
LDOUBLEV 已提交
259
                if model_type in ['table', 'kie']:
M
MissPenguin 已提交
260 261 262 263
                    eval_class(preds, batch)
                else:
                    post_result = post_process_class(preds, batch[1])
                    eval_class(post_result, batch)
littletomatodonkey's avatar
fix doc  
littletomatodonkey 已提交
264 265
                metric = eval_class.get_metric()
                train_stats.update(metric)
W
WenmuZhou 已提交
266 267 268 269 270 271

            if vdl_writer is not None and dist.get_rank() == 0:
                for k, v in train_stats.get().items():
                    vdl_writer.add_scalar('TRAIN/{}'.format(k), v, global_step)
                vdl_writer.add_scalar('TRAIN/lr', lr, global_step)

272 273 274
            if dist.get_rank() == 0 and (
                (global_step > 0 and global_step % print_batch_step == 0) or
                (idx >= len(train_dataloader) - 1)):
W
WenmuZhou 已提交
275
                logs = train_stats.log()
276
                strs = 'epoch: [{}/{}], global_step: {}, {}, avg_reader_cost: {:.5f} s, avg_batch_cost: {:.5f} s, avg_samples: {}, ips: {:.5f}'.format(
W
WenmuZhou 已提交
277
                    epoch, epoch_num, global_step, logs, train_reader_cost /
文幕地方's avatar
文幕地方 已提交
278
                    print_batch_step, (train_reader_cost + train_run_cost) /
279
                    print_batch_step, total_samples / print_batch_step,
文幕地方's avatar
文幕地方 已提交
280
                    total_samples / (train_reader_cost + train_run_cost))
W
WenmuZhou 已提交
281
                logger.info(strs)
282

W
WenmuZhou 已提交
283
                train_reader_cost = 0.0
文幕地方's avatar
文幕地方 已提交
284 285
                train_run_cost = 0.0
                total_samples = 0
W
WenmuZhou 已提交
286 287 288
            # eval
            if global_step > start_eval_step and \
                    (global_step - start_eval_step) % eval_batch_step == 0 and dist.get_rank() == 0:
T
tink2123 已提交
289 290 291 292 293 294 295
                if model_average:
                    Model_Average = paddle.incubate.optimizer.ModelAverage(
                        0.15,
                        parameters=model.parameters(),
                        min_average_window=10000,
                        max_average_window=15625)
                    Model_Average.apply()
T
tink2123 已提交
296 297 298 299 300
                cur_metric = eval(
                    model,
                    valid_dataloader,
                    post_process_class,
                    eval_class,
M
refine  
MissPenguin 已提交
301
                    model_type,
T
tink2123 已提交
302
                    extra_input=extra_input)
L
LDOUBLEV 已提交
303 304 305
                cur_metric_str = 'cur metric, {}'.format(', '.join(
                    ['{}: {}'.format(k, v) for k, v in cur_metric.items()]))
                logger.info(cur_metric_str)
W
WenmuZhou 已提交
306 307 308

                # logger metric
                if vdl_writer is not None:
L
LDOUBLEV 已提交
309
                    for k, v in cur_metric.items():
W
WenmuZhou 已提交
310 311
                        if isinstance(v, (float, int)):
                            vdl_writer.add_scalar('EVAL/{}'.format(k),
L
LDOUBLEV 已提交
312 313
                                                  cur_metric[k], global_step)
                if cur_metric[main_indicator] >= best_model_dict[
W
WenmuZhou 已提交
314
                        main_indicator]:
L
LDOUBLEV 已提交
315
                    best_model_dict.update(cur_metric)
W
WenmuZhou 已提交
316 317 318 319 320 321
                    best_model_dict['best_epoch'] = epoch
                    save_model(
                        model,
                        optimizer,
                        save_model_dir,
                        logger,
322
                        config,
W
WenmuZhou 已提交
323 324 325
                        is_best=True,
                        prefix='best_accuracy',
                        best_model_dict=best_model_dict,
326 327
                        epoch=epoch,
                        global_step=global_step)
L
LDOUBLEV 已提交
328
                best_str = 'best metric, {}'.format(', '.join([
W
WenmuZhou 已提交
329 330 331 332 333 334 335 336
                    '{}: {}'.format(k, v) for k, v in best_model_dict.items()
                ]))
                logger.info(best_str)
                # logger best metric
                if vdl_writer is not None:
                    vdl_writer.add_scalar('EVAL/best_{}'.format(main_indicator),
                                          best_model_dict[main_indicator],
                                          global_step)
337

文幕地方's avatar
文幕地方 已提交
338
            reader_start = time.time()
W
WenmuZhou 已提交
339 340 341 342 343 344
        if dist.get_rank() == 0:
            save_model(
                model,
                optimizer,
                save_model_dir,
                logger,
345
                config,
W
WenmuZhou 已提交
346 347 348
                is_best=False,
                prefix='latest',
                best_model_dict=best_model_dict,
349 350
                epoch=epoch,
                global_step=global_step)
W
WenmuZhou 已提交
351 352 353 354 355 356
        if dist.get_rank() == 0 and epoch > 0 and epoch % save_epoch_step == 0:
            save_model(
                model,
                optimizer,
                save_model_dir,
                logger,
357
                config,
W
WenmuZhou 已提交
358 359 360
                is_best=False,
                prefix='iter_epoch_{}'.format(epoch),
                best_model_dict=best_model_dict,
361 362
                epoch=epoch,
                global_step=global_step)
L
LDOUBLEV 已提交
363
    best_str = 'best metric, {}'.format(', '.join(
W
WenmuZhou 已提交
364 365 366 367
        ['{}: {}'.format(k, v) for k, v in best_model_dict.items()]))
    logger.info(best_str)
    if dist.get_rank() == 0 and vdl_writer is not None:
        vdl_writer.close()
L
LDOUBLEV 已提交
368 369 370
    return


M
refine  
MissPenguin 已提交
371 372 373 374
def eval(model,
         valid_dataloader,
         post_process_class,
         eval_class,
L
LDOUBLEV 已提交
375
         model_type=None,
T
tink2123 已提交
376
         extra_input=False):
W
WenmuZhou 已提交
377 378 379 380
    model.eval()
    with paddle.no_grad():
        total_frame = 0.0
        total_time = 0.0
文幕地方's avatar
文幕地方 已提交
381 382 383 384 385
        pbar = tqdm(
            total=len(valid_dataloader),
            desc='eval model:',
            position=0,
            leave=True)
386 387
        max_iter = len(valid_dataloader) - 1 if platform.system(
        ) == "Windows" else len(valid_dataloader)
W
WenmuZhou 已提交
388
        for idx, batch in enumerate(valid_dataloader):
389
            if idx >= max_iter:
W
WenmuZhou 已提交
390
                break
W
fix bug  
WenmuZhou 已提交
391
            images = batch[0]
W
WenmuZhou 已提交
392
            start = time.time()
T
tink2123 已提交
393
            if model_type == 'table' or extra_input:
M
refine  
MissPenguin 已提交
394
                preds = model(images, data=batch[1:])
395
            elif model_type in ["kie", 'vqa']:
L
LDOUBLEV 已提交
396
                preds = model(batch)
X
xiaoting 已提交
397
            else:
L
LDOUBLEV 已提交
398
                preds = model(images)
399 400 401 402 403 404 405

            batch_numpy = []
            for item in batch:
                if isinstance(item, paddle.Tensor):
                    batch_numpy.append(item.numpy())
                else:
                    batch_numpy.append(item)
W
WenmuZhou 已提交
406 407 408
            # Obtain usable results from post-processing methods
            total_time += time.time() - start
            # Evaluate the results of the current batch
L
LDOUBLEV 已提交
409
            if model_type in ['table', 'kie']:
410 411 412 413
                eval_class(preds, batch_numpy)
            elif model_type in ['vqa']:
                post_result = post_process_class(preds, batch_numpy)
                eval_class(post_result, batch_numpy)
M
MissPenguin 已提交
414
            else:
415 416
                post_result = post_process_class(preds, batch_numpy[1])
                eval_class(post_result, batch_numpy)
L
LDOUBLEV 已提交
417

W
fix bug  
WenmuZhou 已提交
418
            pbar.update(1)
W
WenmuZhou 已提交
419
            total_frame += len(images)
L
LDOUBLEV 已提交
420 421
        # Get final metric,eg. acc or hmean
        metric = eval_class.get_metric()
D
dyning 已提交
422

W
fix bug  
WenmuZhou 已提交
423
    pbar.close()
W
WenmuZhou 已提交
424
    model.train()
L
LDOUBLEV 已提交
425 426
    metric['fps'] = total_frame / total_time
    return metric
L
licx 已提交
427

T
tink2123 已提交
428

B
Bin Lu 已提交
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
def update_center(char_center, post_result, preds):
    result, label = post_result
    feats, logits = preds
    logits = paddle.argmax(logits, axis=-1)
    feats = feats.numpy()
    logits = logits.numpy()

    for idx_sample in range(len(label)):
        if result[idx_sample][0] == label[idx_sample][0]:
            feat = feats[idx_sample]
            logit = logits[idx_sample]
            for idx_time in range(len(logit)):
                index = logit[idx_time]
                if index in char_center.keys():
                    char_center[index][0] = (
                        char_center[index][0] * char_center[index][1] +
                        feat[idx_time]) / (char_center[index][1] + 1)
                    char_center[index][1] += 1
                else:
                    char_center[index] = [feat[idx_time], 1]
    return char_center


def get_center(model, eval_dataloader, post_process_class):
    pbar = tqdm(total=len(eval_dataloader), desc='get center:')
    max_iter = len(eval_dataloader) - 1 if platform.system(
    ) == "Windows" else len(eval_dataloader)
    char_center = dict()
    for idx, batch in enumerate(eval_dataloader):
        if idx >= max_iter:
            break
        images = batch[0]
        start = time.time()
        preds = model(images)

        batch = [item.numpy() for item in batch]
        # Obtain usable results from post-processing methods
        post_result = post_process_class(preds, batch[1])

        #update char_center
        char_center = update_center(char_center, post_result, preds)
        pbar.update(1)

    pbar.close()
    for key in char_center.keys():
        char_center[key] = char_center[key][0]
    return char_center


478
def preprocess(is_train=False):
L
licx 已提交
479
    FLAGS = ArgsParser().parse_args()
L
LDOUBLEV 已提交
480
    profiler_options = FLAGS.profiler_options
L
licx 已提交
481
    config = load_config(FLAGS.config)
482
    config = merge_config(config, FLAGS.opt)
L
LDOUBLEV 已提交
483
    profile_dic = {"profiler_options": FLAGS.profiler_options}
484
    config = merge_config(config, profile_dic)
L
licx 已提交
485

W
WenmuZhou 已提交
486 487 488 489 490 491 492 493 494 495 496
    if is_train:
        # save_config
        save_model_dir = config['Global']['save_model_dir']
        os.makedirs(save_model_dir, exist_ok=True)
        with open(os.path.join(save_model_dir, 'config.yml'), 'w') as f:
            yaml.dump(
                dict(config), f, default_flow_style=False, sort_keys=False)
        log_file = '{}/train.log'.format(save_model_dir)
    else:
        log_file = None
    logger = get_logger(name='root', log_file=log_file)
L
licx 已提交
497 498 499 500 501

    # check if set use_gpu=True in paddlepaddle cpu version
    use_gpu = config['Global']['use_gpu']
    check_gpu(use_gpu)

W
WenmuZhou 已提交
502 503
    alg = config['Architecture']['algorithm']
    assert alg in [
J
Jethong 已提交
504
        'EAST', 'DB', 'SAST', 'Rosetta', 'CRNN', 'STARNet', 'RARE', 'SRN',
T
tink2123 已提交
505
        'CLS', 'PGNet', 'Distillation', 'NRTR', 'TableAttn', 'SAR', 'PSE',
506
        'SEED', 'SDMGR', 'LayoutXLM', 'LayoutLM'
W
WenmuZhou 已提交
507
    ]
L
licx 已提交
508

W
WenmuZhou 已提交
509 510
    device = 'gpu:{}'.format(dist.ParallelEnv().dev_id) if use_gpu else 'cpu'
    device = paddle.set_device(device)
D
dyning 已提交
511

D
dyning 已提交
512
    config['Global']['distributed'] = dist.get_world_size() != 1
W
WenmuZhou 已提交
513

D
dyning 已提交
514 515
    if config['Global']['use_visualdl']:
        from visualdl import LogWriter
L
fix bug  
LDOUBLEV 已提交
516
        save_model_dir = config['Global']['save_model_dir']
D
dyning 已提交
517 518 519 520 521 522 523 524 525
        vdl_writer_path = '{}/vdl/'.format(save_model_dir)
        os.makedirs(vdl_writer_path, exist_ok=True)
        vdl_writer = LogWriter(logdir=vdl_writer_path)
    else:
        vdl_writer = None
    print_dict(config, logger)
    logger.info('train with paddle {} and device {}'.format(paddle.__version__,
                                                            device))
    return config, device, logger, vdl_writer