save_load.py 5.6 KB
Newer Older
L
LDOUBLEV 已提交
1 2
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
W
WenmuZhou 已提交
3 4 5
# 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
L
LDOUBLEV 已提交
6 7 8
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
W
WenmuZhou 已提交
9 10 11 12 13
# 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.
L
LDOUBLEV 已提交
14 15 16 17 18 19 20

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

import errno
import os
W
WenmuZhou 已提交
21 22
import pickle
import six
L
LDOUBLEV 已提交
23

W
WenmuZhou 已提交
24
import paddle
L
LDOUBLEV 已提交
25

W
WenmuZhou 已提交
26
__all__ = ['init_model', 'save_model', 'load_dygraph_pretrain']
L
LDOUBLEV 已提交
27 28


W
WenmuZhou 已提交
29
def _mkdir_if_not_exist(path, logger):
L
LDOUBLEV 已提交
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
    """
    mkdir if not exists, ignore the exception when multiprocess mkdir together
    """
    if not os.path.exists(path):
        try:
            os.makedirs(path)
        except OSError as e:
            if e.errno == errno.EEXIST and os.path.isdir(path):
                logger.warning(
                    'be happy if some process has already created {}'.format(
                        path))
            else:
                raise OSError('Failed to mkdir {}'.format(path))


D
dyning 已提交
45
def load_dygraph_pretrain(model, logger, path=None, load_static_weights=False):
L
LDOUBLEV 已提交
46 47 48
    if not (os.path.isdir(path) or os.path.exists(path + '.pdparams')):
        raise ValueError("Model pretrain path {} does not "
                         "exists.".format(path))
W
WenmuZhou 已提交
49
    if load_static_weights:
D
dyning 已提交
50
        pre_state_dict = paddle.static.load_program_state(path)
W
WenmuZhou 已提交
51 52 53 54 55 56 57
        param_state_dict = {}
        model_dict = model.state_dict()
        for key in model_dict.keys():
            weight_name = model_dict[key].name
            weight_name = weight_name.replace('binarize', '').replace(
                'thresh', '')  # for DB
            if weight_name in pre_state_dict.keys():
L
LDOUBLEV 已提交
58 59
                # logger.info('Load weight: {}, shape: {}'.format(
                #     weight_name, pre_state_dict[weight_name].shape))
W
WenmuZhou 已提交
60 61 62 63 64 65 66 67 68 69 70
                if 'encoder_rnn' in key:
                    # delete axis which is 1
                    pre_state_dict[weight_name] = pre_state_dict[
                        weight_name].squeeze()
                    # change axis
                    if len(pre_state_dict[weight_name].shape) > 1:
                        pre_state_dict[weight_name] = pre_state_dict[
                            weight_name].transpose((1, 0))
                param_state_dict[key] = pre_state_dict[weight_name]
            else:
                param_state_dict[key] = model_dict[key]
W
WenmuZhou 已提交
71
        model.set_state_dict(param_state_dict)
W
WenmuZhou 已提交
72 73
        return

W
WenmuZhou 已提交
74 75
    param_state_dict = paddle.load(path + '.pdparams')
    model.set_state_dict(param_state_dict)
W
WenmuZhou 已提交
76
    return
L
LDOUBLEV 已提交
77

W
WenmuZhou 已提交
78 79

def init_model(config, model, logger, optimizer=None, lr_scheduler=None):
L
LDOUBLEV 已提交
80 81 82
    """
    load model from checkpoint or pretrained_model
    """
W
WenmuZhou 已提交
83 84 85 86
    gloabl_config = config['Global']
    checkpoints = gloabl_config.get('checkpoints')
    pretrained_model = gloabl_config.get('pretrained_model')
    best_model_dict = {}
L
LDOUBLEV 已提交
87
    if checkpoints:
W
WenmuZhou 已提交
88 89 90 91
        assert os.path.exists(checkpoints + ".pdparams"), \
            "Given dir {}.pdparams not exist.".format(checkpoints)
        assert os.path.exists(checkpoints + ".pdopt"), \
            "Given dir {}.pdopt not exist.".format(checkpoints)
W
WenmuZhou 已提交
92 93
        para_dict = paddle.load(checkpoints + '.pdparams')
        opti_dict = paddle.load(checkpoints + '.pdopt')
W
WenmuZhou 已提交
94
        model.set_state_dict(para_dict)
W
WenmuZhou 已提交
95 96 97 98 99 100 101 102 103 104 105 106 107 108
        if optimizer is not None:
            optimizer.set_state_dict(opti_dict)

        if os.path.exists(checkpoints + '.states'):
            with open(checkpoints + '.states', 'rb') as f:
                states_dict = pickle.load(f) if six.PY2 else pickle.load(
                    f, encoding='latin1')
            best_model_dict = states_dict.get('best_model_dict', {})
            if 'epoch' in states_dict:
                best_model_dict['start_epoch'] = states_dict['epoch'] + 1

        logger.info("resume from {}".format(checkpoints))
    elif pretrained_model:
        load_static_weights = gloabl_config.get('load_static_weights', False)
D
dyning 已提交
109 110 111
        if not isinstance(pretrained_model, list):
            pretrained_model = [pretrained_model]
        if not isinstance(load_static_weights, list):
D
dyning 已提交
112
            load_static_weights = [load_static_weights] * len(pretrained_model)
D
dyning 已提交
113 114 115
        for idx, pretrained in enumerate(pretrained_model):
            load_static = load_static_weights[idx]
            load_dygraph_pretrain(
D
dyning 已提交
116
                model, logger, path=pretrained, load_static_weights=load_static)
D
dyning 已提交
117 118
            logger.info("load pretrained model from {}".format(
                pretrained_model))
119
    else:
W
WenmuZhou 已提交
120 121
        logger.info('train from scratch')
    return best_model_dict
L
LDOUBLEV 已提交
122 123


124
def save_model(model,
W
WenmuZhou 已提交
125 126 127 128 129 130
               optimizer,
               model_path,
               logger,
               is_best=False,
               prefix='ppocr',
               **kwargs):
L
LDOUBLEV 已提交
131 132 133
    """
    save model to the target path
    """
W
WenmuZhou 已提交
134 135
    _mkdir_if_not_exist(model_path, logger)
    model_prefix = os.path.join(model_path, prefix)
136
    paddle.save(model.state_dict(), model_prefix + '.pdparams')
W
WenmuZhou 已提交
137
    paddle.save(optimizer.state_dict(), model_prefix + '.pdopt')
W
WenmuZhou 已提交
138 139 140 141 142 143 144 145

    # save metric and config
    with open(model_prefix + '.states', 'wb') as f:
        pickle.dump(kwargs, f, protocol=2)
    if is_best:
        logger.info('save best model is to {}'.format(model_prefix))
    else:
        logger.info("save model in {}".format(model_prefix))