train.py 11.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
# Copyright (c) 2019 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 time
import numpy as np
22 23
import datetime
from collections import deque
24

25

26 27 28 29 30
def set_paddle_flags(**kwargs):
    for key, value in kwargs.items():
        if os.environ.get(key, None) is None:
            os.environ[key] = str(value)

31

32
# NOTE(paddle-dev): All of these flags should be set before
33
# `import paddle`. Otherwise, it would not take any effect.
34 35 36 37 38
set_paddle_flags(
    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
)

from paddle import fluid
39
from paddle.fluid.contrib import mixed_precision
40 41 42
from ppdet.core.workspace import load_config, merge_config, create
from ppdet.data.data_feed import create_reader

43
from ppdet.utils.cli import print_total_cfg
44
from ppdet.utils import dist_utils
45 46 47
from ppdet.utils.eval_utils import parse_fetches, eval_run, eval_results
from ppdet.utils.stats import TrainingStats
from ppdet.utils.cli import ArgsParser
48
from ppdet.utils.check import check_gpu
49 50 51 52 53 54 55 56 57 58
import ppdet.utils.checkpoint as checkpoint
from ppdet.modeling.model_input import create_feed

import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)


def main():
59 60 61 62 63 64 65 66 67
    env = os.environ
    FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env
    if FLAGS.dist:
        trainer_id = int(env['PADDLE_TRAINER_ID'])
        import random
        local_seed = (99 + trainer_id)
        random.seed(local_seed)
        np.random.seed(local_seed)

68 69 70 71 72 73 74
    cfg = load_config(FLAGS.config)
    if 'architecture' in cfg:
        main_arch = cfg.architecture
    else:
        raise ValueError("'architecture' not specified in config file.")

    merge_config(FLAGS.opt)
75 76
    if 'log_iter' not in cfg:
        cfg.log_iter = 20
77

78 79 80
    ignore_params = cfg.finetune_exclude_pretrained_params \
                 if 'finetune_exclude_pretrained_params' in cfg else []

81 82
    # check if set use_gpu=True in paddlepaddle cpu version
    check_gpu(cfg.use_gpu)
W
wangguanzhong 已提交
83
    print_total_cfg(cfg)
84

85 86 87
    if cfg.use_gpu:
        devices_num = fluid.core.get_cuda_device_count()
    else:
88
        devices_num = int(os.environ.get('CPU_NUM', 1))
89 90 91 92 93 94 95 96 97 98 99 100

    if 'train_feed' not in cfg:
        train_feed = create(main_arch + 'TrainFeed')
    else:
        train_feed = create(cfg.train_feed)

    if FLAGS.eval:
        if 'eval_feed' not in cfg:
            eval_feed = create(main_arch + 'EvalFeed')
        else:
            eval_feed = create(cfg.eval_feed)

101 102 103 104
    if 'FLAGS_selected_gpus' in env:
        device_id = int(env['FLAGS_selected_gpus'])
    else:
        device_id = 0
W
wangguanzhong 已提交
105
    place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace()
106 107 108 109 110 111 112 113 114 115
    exe = fluid.Executor(place)

    lr_builder = create('LearningRate')
    optim_builder = create('OptimizerBuilder')

    # build program
    startup_prog = fluid.Program()
    train_prog = fluid.Program()
    with fluid.program_guard(train_prog, startup_prog):
        with fluid.unique_name.guard():
116
            model = create(main_arch)
117 118 119 120 121
            train_pyreader, feed_vars = create_feed(train_feed)
            train_fetches = model.train(feed_vars)
            loss = train_fetches['loss']
            lr = lr_builder()
            optimizer = optim_builder(lr)
122 123 124 125 126
            if FLAGS.fp16:
                optimizer = mixed_precision.decorate(
                    optimizer=optimizer,
                    init_loss_scaling=FLAGS.loss_scale,
                    use_dynamic_loss_scaling=False)
127 128 129 130 131 132 133 134 135 136
            optimizer.minimize(loss)

    # parse train fetches
    train_keys, train_values, _ = parse_fetches(train_fetches)
    train_values.append(lr)

    if FLAGS.eval:
        eval_prog = fluid.Program()
        with fluid.program_guard(eval_prog, startup_prog):
            with fluid.unique_name.guard():
137
                model = create(main_arch)
138 139 140 141
                eval_pyreader, feed_vars = create_feed(eval_feed)
                fetches = model.eval(feed_vars)
        eval_prog = eval_prog.clone(True)

W
wangguanzhong 已提交
142
        eval_reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir)
143 144
        eval_pyreader.decorate_sample_list_generator(eval_reader, place)

145
        # parse eval fetches
146 147 148 149 150
        extra_keys = []
        if cfg.metric == 'COCO':
            extra_keys = ['im_info', 'im_id', 'im_shape']
        if cfg.metric == 'VOC':
            extra_keys = ['gt_box', 'gt_label', 'is_difficult']
151 152 153 154 155
        eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog,
                                                         extra_keys)

    # compile program for multi-devices
    build_strategy = fluid.BuildStrategy()
156
    build_strategy.fuse_all_optimizer_ops = False
157 158
    if FLAGS.fp16:
        build_strategy.fuse_all_reduce_ops = False
K
Kaipeng Deng 已提交
159
    # only enable sync_bn in multi GPU devices
160
    sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn'
161 162
    build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \
        and cfg.use_gpu
163 164 165 166 167 168

    exec_strategy = fluid.ExecutionStrategy()
    # iteration number when CompiledProgram tries to drop local execution scopes.
    # Set it to be 1 to save memory usages, so that unused variables in
    # local execution scopes can be deleted after each iteration.
    exec_strategy.num_iteration_per_drop_scope = 1
169
    if FLAGS.dist:
W
wangguanzhong 已提交
170 171
        dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog,
                                             train_prog)
172
        exec_strategy.num_threads = 1
173 174

    exe.run(startup_prog)
175 176 177 178 179 180 181
    compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel(
        loss_name=loss.name,
        build_strategy=build_strategy,
        exec_strategy=exec_strategy)

    if FLAGS.eval:
        compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog)
182

183
    fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel'
Q
qingqing01 已提交
184
    start_iter = 0
185

186 187
    if FLAGS.resume_checkpoint:
        checkpoint.load_checkpoint(exe, train_prog, FLAGS.resume_checkpoint)
Q
qingqing01 已提交
188
        start_iter = checkpoint.global_step()
189
    elif cfg.pretrain_weights and fuse_bn and not ignore_params:
190 191
        checkpoint.load_and_fusebn(exe, train_prog, cfg.pretrain_weights)
    elif cfg.pretrain_weights:
192 193
        checkpoint.load_params(
            exe, train_prog, cfg.pretrain_weights, ignore_params=ignore_params)
194

W
wangguanzhong 已提交
195 196
    train_reader = create_reader(train_feed, (cfg.max_iters - start_iter) *
                                 devices_num, FLAGS.dataset_dir)
197 198
    train_pyreader.decorate_sample_list_generator(train_reader, place)

199 200 201 202 203 204
    # whether output bbox is normalized in model output layer
    is_bbox_normalized = False
    if hasattr(model, 'is_bbox_normalized') and \
            callable(model.is_bbox_normalized):
        is_bbox_normalized = model.is_bbox_normalized()

K
Kaipeng Deng 已提交
205 206 207
    # if map_type not set, use default 11point, only use in VOC eval
    map_type = cfg.map_type if 'map_type' in cfg else '11point'

208 209 210 211 212 213 214
    train_stats = TrainingStats(cfg.log_smooth_window, train_keys)
    train_pyreader.start()
    start_time = time.time()
    end_time = time.time()

    cfg_name = os.path.basename(FLAGS.config).split('.')[0]
    save_dir = os.path.join(cfg.save_dir, cfg_name)
215
    time_stat = deque(maxlen=cfg.log_iter)
216
    best_box_ap_list = [0.0, 0]  #[map, iter]
217 218 219 220 221 222 223 224

    # use tb-paddle to log data
    if FLAGS.use_tb:
        from tb_paddle import SummaryWriter
        tb_writer = SummaryWriter(FLAGS.tb_log_dir)
        tb_loss_step = 0
        tb_mAP_step = 0

Q
qingqing01 已提交
225
    for it in range(start_iter, cfg.max_iters):
226 227
        start_time = end_time
        end_time = time.time()
228 229 230 231
        time_stat.append(end_time - start_time)
        time_cost = np.mean(time_stat)
        eta_sec = (cfg.max_iters - it) * time_cost
        eta = str(datetime.timedelta(seconds=int(eta_sec)))
232
        outs = exe.run(compiled_train_prog, fetch_list=train_values)
233
        stats = {k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1])}
234 235 236 237 238 239 240 241

        # use tb-paddle to log loss
        if FLAGS.use_tb:
            if it % cfg.log_iter == 0:
                for loss_name, loss_value in stats.items():
                    tb_writer.add_scalar(loss_name, loss_value, tb_loss_step)
                tb_loss_step += 1

242 243
        train_stats.update(stats)
        logs = train_stats.log()
244
        if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0):
245 246 247
            strs = 'iter: {}, lr: {:.6f}, {}, time: {:.3f}, eta: {}'.format(
                it, np.mean(outs[-1]), logs, time_cost, eta)
            logger.info(strs)
248

249 250
        if (it > 0 and it % cfg.snapshot_iter == 0 or it == cfg.max_iters - 1) \
           and (not FLAGS.dist or trainer_id == 0):
251 252
            save_name = str(it) if it != cfg.max_iters - 1 else "model_final"
            checkpoint.save(exe, train_prog, os.path.join(save_dir, save_name))
253 254 255

            if FLAGS.eval:
                # evaluation
256
                results = eval_run(exe, compiled_eval_prog, eval_pyreader,
257 258 259 260
                                   eval_keys, eval_values, eval_cls)
                resolution = None
                if 'mask' in results[0]:
                    resolution = model.mask_head.resolution
261 262 263
                box_ap_stats = eval_results(
                    results, eval_feed, cfg.metric, cfg.num_classes, resolution,
                    is_bbox_normalized, FLAGS.output_eval, map_type)
264

265 266 267 268
                # use tb_paddle to log mAP
                if FLAGS.use_tb:
                    tb_writer.add_scalar("mAP", box_ap_stats[0], tb_mAP_step)
                    tb_mAP_step += 1
269

270 271 272
                if box_ap_stats[0] > best_box_ap_list[0]:
                    best_box_ap_list[0] = box_ap_stats[0]
                    best_box_ap_list[1] = it
273 274
                    checkpoint.save(exe, train_prog,
                                    os.path.join(save_dir, "best_model"))
275
                logger.info("Best test box ap: {}, in iter: {}".format(
276
                    best_box_ap_list[0], best_box_ap_list[1]))
277 278 279 280 281 282

    train_pyreader.reset()


if __name__ == '__main__':
    parser = ArgsParser()
283 284 285 286 287 288 289 290 291 292
    parser.add_argument(
        "--fp16",
        action='store_true',
        default=False,
        help="Enable mixed precision training.")
    parser.add_argument(
        "--loss_scale",
        default=8.,
        type=float,
        help="Mixed precision training loss scale.")
293 294 295 296 297 298 299 300 301 302 303 304
    parser.add_argument(
        "-r",
        "--resume_checkpoint",
        default=None,
        type=str,
        help="Checkpoint path for resuming training.")
    parser.add_argument(
        "--eval",
        action='store_true',
        default=False,
        help="Whether to perform evaluation in train")
    parser.add_argument(
305
        "--output_eval",
306 307
        default=None,
        type=str,
308
        help="Evaluation directory, default is current directory.")
W
wangguanzhong 已提交
309 310 311 312 313 314
    parser.add_argument(
        "-d",
        "--dataset_dir",
        default=None,
        type=str,
        help="Dataset path, same as DataFeed.dataset.dataset_dir")
315 316 317 318 319 320 321 322 323 324
    parser.add_argument(
        "--use_tb",
        type=bool,
        default=False,
        help="whether to record the data to Tensorboard.")
    parser.add_argument(
        '--tb_log_dir',
        type=str,
        default="tb_log_dir/scalar",
        help='Tensorboard logging directory for scalar.')
325 326
    FLAGS = parser.parse_args()
    main()