train_rec.py

# 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.

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

import os
import sys
import time
import multiprocessing
import numpy as np

# from paddle.fluid.contrib.model_stat import summary


def set_paddle_flags(**kwargs):
    for key, value in kwargs.items():
        if os.environ.get(key, None) is None:
            os.environ[key] = str(value)


# NOTE(paddle-dev): All of these flags should be
# set before `import paddle`. Otherwise, it would
# not take any effect. 
set_paddle_flags(
    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
)

from paddle import fluid
from ppocr.utils.utility import create_module
from ppocr.utils.utility import load_config, merge_config
import ppocr.data.rec.reader_main as reader
from ppocr.utils.utility import ArgsParser
from ppocr.utils.character import CharacterOps, cal_predicts_accuracy
from ppocr.utils.check import check_gpu
from ppocr.utils.stats import TrainingStats
from ppocr.utils.checkpoint import load_pretrain, load_checkpoint, save, save_model
from ppocr.utils.eval_utils import eval_run

from ppocr.utils.utility import initial_logger
logger = initial_logger()
from ppocr.utils.utility import create_multi_devices_program


def main():
    config = load_config(FLAGS.config)
    merge_config(FLAGS.opt)
    char_ops = CharacterOps(config['Global'])
    config['Global']['char_num'] = char_ops.get_char_num()
    print(config)

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

    place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
    exe = fluid.Executor(place)

    rec_model = create_module(config['Architecture']['function'])(params=config)

    startup_prog = fluid.Program()
    train_prog = fluid.Program()
    with fluid.program_guard(train_prog, startup_prog):
        with fluid.unique_name.guard():
            train_loader, train_outputs = rec_model(mode="train")
            save_var = train_outputs[1]

            if "gradient_clip" in config['Global']:
                gradient_clip = config['Global']['gradient_clip']
                clip = fluid.clip.GradientClipByGlobalNorm(gradient_clip)
                fluid.clip.set_gradient_clip(clip, program=train_prog)

            train_fetch_list = [v.name for v in train_outputs]
            train_loss = train_outputs[0]
            opt_params = config['Optimizer']
            optimizer = create_module(opt_params['function'])(opt_params)
            optimizer.minimize(train_loss)
            global_lr = optimizer._global_learning_rate()
            global_lr.persistable = True
            train_fetch_list.append(global_lr.name)

    train_reader = reader.train_eval_reader(
        config=config, char_ops=char_ops, mode="train")
    train_loader.set_sample_list_generator(train_reader, places=place)

    eval_prog = fluid.Program()
    with fluid.program_guard(eval_prog, startup_prog):
        with fluid.unique_name.guard():
            eval_loader, eval_outputs = rec_model(mode="eval")
            eval_fetch_list = [v.name for v in eval_outputs]

    eval_prog = eval_prog.clone(for_test=True)
    exe.run(startup_prog)

    eval_reader = reader.train_eval_reader(
        config=config, char_ops=char_ops, mode="eval")
    eval_loader.set_sample_list_generator(eval_reader, places=place)

    # compile program for multi-devices
    train_compile_program = create_multi_devices_program(train_prog,
                                                         train_loss.name)

    pretrain_weights = config['Global']['pretrain_weights']
    if pretrain_weights is not None:
        load_pretrain(exe, train_prog, pretrain_weights)

    train_batch_id = 0
    train_log_keys = ['loss', 'acc']
    log_smooth_window = config['Global']['log_smooth_window']
    epoch_num = config['Global']['epoch_num']
    loss_type = config['Global']['loss_type']
    print_step = config['Global']['print_step']
    eval_step = config['Global']['eval_step']
    save_epoch_step = config['Global']['save_epoch_step']
    save_dir = config['Global']['save_dir']
    train_stats = TrainingStats(log_smooth_window, train_log_keys)
    best_eval_acc = -1
    best_batch_id = 0
    best_epoch = 0
    for epoch in range(epoch_num):
        train_loader.start()
        try:
            while True:
                t1 = time.time()
                train_outs = exe.run(program=train_compile_program,
                                     fetch_list=train_fetch_list,
                                     return_numpy=False)
                loss = np.mean(np.array(train_outs[0]))
                lr = np.mean(np.array(train_outs[-1]))

                preds = np.array(train_outs[1])
                preds_lod = train_outs[1].lod()[0]
                labels = np.array(train_outs[2])
                labels_lod = train_outs[2].lod()[0]

                acc, acc_num, img_num = cal_predicts_accuracy(
                    char_ops, preds, preds_lod, labels, labels_lod)

                t2 = time.time()
                train_batch_elapse = t2 - t1

                stats = {'loss': loss, 'acc': acc}
                train_stats.update(stats)
                if train_batch_id > 0 and train_batch_id % print_step == 0:
                    logs = train_stats.log()
                    strs = 'epoch: {}, iter: {}, lr: {:.6f}, {}, time: {:.3f}'.format(
                        epoch, train_batch_id, lr, logs, train_batch_elapse)
                    logger.info(strs)

                if train_batch_id > 0 and train_batch_id % eval_step == 0:
                    outs = eval_run(exe, eval_prog, eval_loader,
                                    eval_fetch_list, char_ops, train_batch_id,
                                    "eval")
                    eval_acc, acc_num, sample_num = outs
                    if eval_acc > best_eval_acc:
                        best_eval_acc = eval_acc
                        best_batch_id = train_batch_id
                        best_epoch = epoch
                        save_path = save_dir + "/best_accuracy"
                        save_model(train_prog, save_path)

                    strs = 'Test iter: {}, acc:{:.6f}, best_acc:{:.6f}, best_epoch:{}, best_batch_id:{}, sample_num:{}'.format(
                        train_batch_id, eval_acc, best_eval_acc, best_epoch,
                        best_batch_id, sample_num)
                    logger.info(strs)
                train_batch_id += 1

        except fluid.core.EOFException:
            train_loader.reset()

        if epoch > 0 and epoch % save_epoch_step == 0:
            save_path = save_dir + "/iter_epoch_%d" % (epoch)
            save_model(train_prog, save_path)


def test_reader():
    config = load_config(FLAGS.config)
    merge_config(FLAGS.opt)
    char_ops = CharacterOps(config['Global'])
    config['Global']['char_num'] = char_ops.get_char_num()
    print(config)
    #     tmp_reader = reader.train_eval_reader(
    #         config=cfg, char_ops=char_ops, mode="train")
    tmp_reader = reader.train_eval_reader(
        config=config, char_ops=char_ops, mode="eval")
    count = 0
    print_count = 0
    import time
    starttime = time.time()
    for data in tmp_reader():
        count += len(data)
        print_count += 1
        if print_count % 10 == 0:
            batch_time = (time.time() - starttime) / print_count
            print("reader:", count, len(data), batch_time)
    print("finish reader:", count)
    print("success")


if __name__ == '__main__':
    parser = ArgsParser()
    parser.add_argument(
        "-r",
        "--resume_checkpoint",
        default=None,
        type=str,
        help="Checkpoint path for resuming training.")
    FLAGS = parser.parse_args()
    main()
#     test_reader()