sensitivity_anal.py 5.8 KB
Newer Older
L
LDOUBLEV 已提交
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
# Copyright (c) 2021 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

__dir__ = os.path.dirname(__file__)
sys.path.append(__dir__)
sys.path.append(os.path.join(__dir__, '..', '..', '..'))
sys.path.append(os.path.join(__dir__, '..', '..', '..', 'tools'))

import paddle
import paddle.distributed as dist
from ppocr.data import build_dataloader
from ppocr.modeling.architectures import build_model
from ppocr.losses import build_loss
from ppocr.optimizer import build_optimizer
from ppocr.postprocess import build_post_process
from ppocr.metrics import build_metric
from ppocr.utils.save_load import init_model
import tools.program as program

dist.get_world_size()


def get_pruned_params(parameters):
    params = []

    for param in parameters:
        if len(
                param.shape
        ) == 4 and 'depthwise' not in param.name and 'transpose' not in param.name and "conv2d_57" not in param.name and "conv2d_56" not in param.name:
            params.append(param.name)
    return params


def main(config, device, logger, vdl_writer):
    # init dist environment
    if config['Global']['distributed']:
        dist.init_parallel_env()

    global_config = config['Global']

    # build dataloader
    train_dataloader = build_dataloader(config, 'Train', device, logger)
    if config['Eval']:
        valid_dataloader = build_dataloader(config, 'Eval', device, logger)
    else:
        valid_dataloader = None

    # build post process
    post_process_class = build_post_process(config['PostProcess'],
                                            global_config)

    # build model
    # for rec algorithm
    if hasattr(post_process_class, 'character'):
        char_num = len(getattr(post_process_class, 'character'))
        config['Architecture']["Head"]['out_channels'] = char_num
    model = build_model(config['Architecture'])

    flops = paddle.flops(model, [1, 3, 640, 640])
L
LDOUBLEV 已提交
78
    logger.info("FLOPs before pruning: {}".format(flops))
L
LDOUBLEV 已提交
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

    from paddleslim.dygraph import FPGMFilterPruner
    model.train()
    pruner = FPGMFilterPruner(model, [1, 3, 640, 640])

    # build loss
    loss_class = build_loss(config['Loss'])

    # build optim
    optimizer, lr_scheduler = build_optimizer(
        config['Optimizer'],
        epochs=config['Global']['epoch_num'],
        step_each_epoch=len(train_dataloader),
        parameters=model.parameters())

    # build metric
    eval_class = build_metric(config['Metric'])
    # load pretrain model
    pre_best_model_dict = init_model(config, model, logger, optimizer)

    logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
                format(len(train_dataloader), len(valid_dataloader)))
    # build metric
    eval_class = build_metric(config['Metric'])

    logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
                format(len(train_dataloader), len(valid_dataloader)))

    def eval_fn():
        metric = program.eval(model, valid_dataloader, post_process_class,
L
LDOUBLEV 已提交
109 110
                              eval_class, False)
        logger.info("metric['hmean']: {}".format(metric['hmean']))
L
LDOUBLEV 已提交
111 112
        return metric['hmean']

L
LDOUBLEV 已提交
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
    run_sensitive_analysis = False
    """
    run_sensitive_analysis=True: 
        Automatically compute the sensitivities of convolutions in a model. 
        The sensitivity of a convolution is the losses of accuracy on test dataset in 
        differenct pruned ratios. The sensitivities can be used to get a group of best 
        ratios with some condition.
    
    run_sensitive_analysis=False: 
        Set prune trim ratio to a fixed value, such as 10%. The larger the value, 
        the more convolution weights will be cropped.

    """

    if run_sensitive_analysis:
        params_sensitive = pruner.sensitive(
            eval_func=eval_fn,
            sen_file="./deploy/slim/prune/sen.pickle",
            skip_vars=[
                "conv2d_57.w_0", "conv2d_transpose_2.w_0",
                "conv2d_transpose_3.w_0"
            ])
        logger.info(
            "The sensitivity analysis results of model parameters saved in sen.pickle"
        )
        # calculate pruned params's ratio
        params_sensitive = pruner._get_ratios_by_loss(
            params_sensitive, loss=0.02)
        for key in params_sensitive.keys():
            logger.info("{}, {}".format(key, params_sensitive[key]))
    else:
        params_sensitive = {}
        for param in model.parameters():
            if 'transpose' not in param.name and 'linear' not in param.name:
                # set prune ratio as 10%. The larger the value, the more convolution weights will be cropped
                params_sensitive[param.name] = 0.1
L
LDOUBLEV 已提交
149 150 151 152

    plan = pruner.prune_vars(params_sensitive, [0])

    flops = paddle.flops(model, [1, 3, 640, 640])
L
LDOUBLEV 已提交
153
    logger.info("FLOPs after pruning: {}".format(flops))
L
LDOUBLEV 已提交
154 155 156 157 158 159 160 161 162 163 164

    # start train

    program.train(config, train_dataloader, valid_dataloader, device, model,
                  loss_class, optimizer, lr_scheduler, post_process_class,
                  eval_class, pre_best_model_dict, logger, vdl_writer)


if __name__ == '__main__':
    config, device, logger, vdl_writer = program.preprocess(is_train=True)
    main(config, device, logger, vdl_writer)