train.py 18.2 KB
Newer Older
W
wuzewu 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
# coding: utf8
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# 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
# GPU memory garbage collection optimization flags
os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0"

import sys
import argparse
import pprint
import shutil
import functools

import paddle
import numpy as np
import paddle.fluid as fluid

from utils.config import cfg
from utils.timer import Timer, calculate_eta
from metrics import ConfusionMatrix
from reader import SegDataset
from models.model_builder import build_model
from models.model_builder import ModelPhase
from models.model_builder import parse_shape_from_file
from eval import evaluate
from vis import visualize
43
from utils import dist_utils
W
wuzewu 已提交
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 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154


def parse_args():
    parser = argparse.ArgumentParser(description='PaddleSeg training')
    parser.add_argument(
        '--cfg',
        dest='cfg_file',
        help='Config file for training (and optionally testing)',
        default=None,
        type=str)
    parser.add_argument(
        '--use_gpu',
        dest='use_gpu',
        help='Use gpu or cpu',
        action='store_true',
        default=False)
    parser.add_argument(
        '--use_mpio',
        dest='use_mpio',
        help='Use multiprocess I/O or not',
        action='store_true',
        default=False)
    parser.add_argument(
        '--log_steps',
        dest='log_steps',
        help='Display logging information at every log_steps',
        default=10,
        type=int)
    parser.add_argument(
        '--debug',
        dest='debug',
        help='debug mode, display detail information of training',
        action='store_true')
    parser.add_argument(
        '--use_tb',
        dest='use_tb',
        help='whether to record the data during training to Tensorboard',
        action='store_true')
    parser.add_argument(
        '--tb_log_dir',
        dest='tb_log_dir',
        help='Tensorboard logging directory',
        default=None,
        type=str)
    parser.add_argument(
        '--do_eval',
        dest='do_eval',
        help='Evaluation models result on every new checkpoint',
        action='store_true')
    parser.add_argument(
        'opts',
        help='See utils/config.py for all options',
        default=None,
        nargs=argparse.REMAINDER)
    return parser.parse_args()


def save_vars(executor, dirname, program=None, vars=None):
    """
    Temporary resolution for Win save variables compatability.
    Will fix in PaddlePaddle v1.5.2
    """

    save_program = fluid.Program()
    save_block = save_program.global_block()

    for each_var in vars:
        # NOTE: don't save the variable which type is RAW
        if each_var.type == fluid.core.VarDesc.VarType.RAW:
            continue
        new_var = save_block.create_var(
            name=each_var.name,
            shape=each_var.shape,
            dtype=each_var.dtype,
            type=each_var.type,
            lod_level=each_var.lod_level,
            persistable=True)
        file_path = os.path.join(dirname, new_var.name)
        file_path = os.path.normpath(file_path)
        save_block.append_op(
            type='save',
            inputs={'X': [new_var]},
            outputs={},
            attrs={'file_path': file_path})

    executor.run(save_program)


def save_checkpoint(exe, program, ckpt_name):
    """
    Save checkpoint for evaluation or resume training
    """
    ckpt_dir = os.path.join(cfg.TRAIN.MODEL_SAVE_DIR, str(ckpt_name))
    print("Save model checkpoint to {}".format(ckpt_dir))
    if not os.path.isdir(ckpt_dir):
        os.makedirs(ckpt_dir)

    save_vars(
        exe,
        ckpt_dir,
        program,
        vars=list(filter(fluid.io.is_persistable, program.list_vars())))

    return ckpt_dir


def load_checkpoint(exe, program):
    """
    Load checkpoiont from pretrained model directory for resume training
    """

W
wuzewu 已提交
155 156
    print('Resume model training from:', cfg.TRAIN.RESUME_MODEL_DIR)
    if not os.path.exists(cfg.TRAIN.RESUME_MODEL_DIR):
W
wuzewu 已提交
157
        raise ValueError("TRAIN.PRETRAIN_MODEL {} not exist!".format(
W
wuzewu 已提交
158
            cfg.TRAIN.RESUME_MODEL_DIR))
W
wuzewu 已提交
159 160

    fluid.io.load_persistables(
W
wuzewu 已提交
161
        exe, cfg.TRAIN.RESUME_MODEL_DIR, main_program=program)
W
wuzewu 已提交
162

W
wuzewu 已提交
163
    model_path = cfg.TRAIN.RESUME_MODEL_DIR
W
wuzewu 已提交
164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
    # Check is path ended by path spearator
    if model_path[-1] == os.sep:
        model_path = model_path[0:-1]
    epoch_name = os.path.basename(model_path)
    # If resume model is final model
    if epoch_name == 'final':
        begin_epoch = cfg.SOLVER.NUM_EPOCHS
    # If resume model path is end of digit, restore epoch status
    elif epoch_name.isdigit():
        epoch = int(epoch_name)
        begin_epoch = epoch + 1
    else:
        raise ValueError("Resume model path is not valid!")
    print("Model checkpoint loaded successfully!")

    return begin_epoch


182 183 184
def print_info(*msg):
    if cfg.TRAINER_ID == 0:
        print(*msg)
W
wuzewu 已提交
185

W
wuzewu 已提交
186

W
wuzewu 已提交
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
def train(cfg):
    startup_prog = fluid.Program()
    train_prog = fluid.Program()
    drop_last = True

    dataset = SegDataset(
        file_list=cfg.DATASET.TRAIN_FILE_LIST,
        mode=ModelPhase.TRAIN,
        shuffle=True,
        data_dir=cfg.DATASET.DATA_DIR)

    def data_generator():
        if args.use_mpio:
            data_gen = dataset.multiprocess_generator(
                num_processes=cfg.DATALOADER.NUM_WORKERS,
                max_queue_size=cfg.DATALOADER.BUF_SIZE)
        else:
            data_gen = dataset.generator()

        batch_data = []
        for b in data_gen:
            batch_data.append(b)
209
            if len(batch_data) == (cfg.BATCH_SIZE // cfg.NUM_TRAINERS):
W
wuzewu 已提交
210 211 212 213 214 215 216 217 218 219
                for item in batch_data:
                    yield item[0], item[1], item[2]
                batch_data = []
        # If use sync batch norm strategy, drop last batch if number of samples
        # in batch_data is less then cfg.BATCH_SIZE to avoid NCCL hang issues
        if not cfg.TRAIN.SYNC_BATCH_NORM:
            for item in batch_data:
                yield item[0], item[1], item[2]

    # Get device environment
220 221 222 223
    # places = fluid.cuda_places() if args.use_gpu else fluid.cpu_places()
    # place = places[0]
    gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0))
    place = fluid.CUDAPlace(gpu_id) if args.use_gpu else fluid.CPUPlace()
W
wuzewu 已提交
224
    places = fluid.cuda_places() if args.use_gpu else fluid.cpu_places()
225

W
wuzewu 已提交
226
    # Get number of GPU
227 228
    dev_count = cfg.NUM_TRAINERS if cfg.NUM_TRAINERS > 1 else len(places)
    print_info("#Device count: {}".format(dev_count))
W
wuzewu 已提交
229 230 231 232 233 234 235

    # Make sure BATCH_SIZE can divided by GPU cards
    assert cfg.BATCH_SIZE % dev_count == 0, (
        'BATCH_SIZE:{} not divisble by number of GPUs:{}'.format(
            cfg.BATCH_SIZE, dev_count))
    # If use multi-gpu training mode, batch data will allocated to each GPU evenly
    batch_size_per_dev = cfg.BATCH_SIZE // dev_count
236
    print_info("batch_size_per_dev: {}".format(batch_size_per_dev))
W
wuzewu 已提交
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251

    py_reader, avg_loss, lr, pred, grts, masks = build_model(
        train_prog, startup_prog, phase=ModelPhase.TRAIN)
    py_reader.decorate_sample_generator(
        data_generator, batch_size=batch_size_per_dev, drop_last=drop_last)

    exe = fluid.Executor(place)
    exe.run(startup_prog)

    exec_strategy = fluid.ExecutionStrategy()
    # Clear temporary variables every 100 iteration
    if args.use_gpu:
        exec_strategy.num_threads = fluid.core.get_cuda_device_count()
    exec_strategy.num_iteration_per_drop_scope = 100
    build_strategy = fluid.BuildStrategy()
252 253 254 255 256

    if cfg.NUM_TRAINERS > 1 and args.use_gpu:
        dist_utils.prepare_for_multi_process(exe, build_strategy, train_prog)
        exec_strategy.num_threads = 1

W
wuzewu 已提交
257 258 259
    if cfg.TRAIN.SYNC_BATCH_NORM and args.use_gpu:
        if dev_count > 1:
            # Apply sync batch norm strategy
260
            print_info("Sync BatchNorm strategy is effective.")
W
wuzewu 已提交
261 262
            build_strategy.sync_batch_norm = True
        else:
W
wuzewu 已提交
263 264 265
            print_info(
                "Sync BatchNorm strategy will not be effective if GPU device"
                " count <= 1")
W
wuzewu 已提交
266 267 268 269 270 271 272
    compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel(
        loss_name=avg_loss.name,
        exec_strategy=exec_strategy,
        build_strategy=build_strategy)

    # Resume training
    begin_epoch = cfg.SOLVER.BEGIN_EPOCH
W
wuzewu 已提交
273
    if cfg.TRAIN.RESUME_MODEL_DIR:
W
wuzewu 已提交
274 275
        begin_epoch = load_checkpoint(exe, train_prog)
    # Load pretrained model
W
wuzewu 已提交
276
    elif os.path.exists(cfg.TRAIN.PRETRAINED_MODEL_DIR):
277
        print_info('Pretrained model dir: ', cfg.TRAIN.PRETRAINED_MODEL_DIR)
W
wuzewu 已提交
278
        load_vars = []
W
wuzewu 已提交
279
        load_fail_vars = []
W
wuzewu 已提交
280 281 282 283 284 285

        def var_shape_matched(var, shape):
            """
            Check whehter persitable variable shape is match with current network
            """
            var_exist = os.path.exists(
W
wuzewu 已提交
286
                os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
W
wuzewu 已提交
287 288
            if var_exist:
                var_shape = parse_shape_from_file(
W
wuzewu 已提交
289
                    os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
W
wuzewu 已提交
290 291
                return var_shape == shape
            return False
W
wuzewu 已提交
292 293 294 295 296 297 298

        for x in train_prog.list_vars():
            if isinstance(x, fluid.framework.Parameter):
                shape = tuple(fluid.global_scope().find_var(
                    x.name).get_tensor().shape())
                if var_shape_matched(x, shape):
                    load_vars.append(x)
W
wuzewu 已提交
299 300
                else:
                    load_fail_vars.append(x)
301 302 303

        fluid.io.load_vars(
            exe, dirname=cfg.TRAIN.PRETRAINED_MODEL_DIR, vars=load_vars)
W
wuzewu 已提交
304
        for var in load_vars:
305
            print_info("Parameter[{}] loaded sucessfully!".format(var.name))
W
wuzewu 已提交
306
        for var in load_fail_vars:
W
wuzewu 已提交
307 308 309
            print_info(
                "Parameter[{}] don't exist or shape does not match current network, skip"
                " to load it.".format(var.name))
310
        print_info("{}/{} pretrained parameters loaded successfully!".format(
W
wuzewu 已提交
311 312
            len(load_vars),
            len(load_vars) + len(load_fail_vars)))
W
wuzewu 已提交
313
    else:
W
wuzewu 已提交
314 315 316
        print_info(
            'Pretrained model dir {} not exists, training from scratch...'.
            format(cfg.TRAIN.PRETRAINED_MODEL_DIR))
W
wuzewu 已提交
317 318 319 320 321 322 323 324 325 326 327 328

    fetch_list = [avg_loss.name, lr.name]
    if args.debug:
        # Fetch more variable info and use streaming confusion matrix to
        # calculate IoU results if in debug mode
        np.set_printoptions(
            precision=4, suppress=True, linewidth=160, floatmode="fixed")
        fetch_list.extend([pred.name, grts.name, masks.name])
        cm = ConfusionMatrix(cfg.DATASET.NUM_CLASSES, streaming=True)

    if args.use_tb:
        if not args.tb_log_dir:
329
            print_info("Please specify the log directory by --tb_log_dir.")
W
wuzewu 已提交
330 331 332 333 334
            exit(1)

        from tb_paddle import SummaryWriter
        log_writer = SummaryWriter(args.tb_log_dir)

335 336
    # trainer_id = int(os.getenv("PADDLE_TRAINER_ID", 0))
    # num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
W
wuzewu 已提交
337 338 339 340 341 342 343 344 345 346 347 348 349 350
    global_step = 0
    all_step = cfg.DATASET.TRAIN_TOTAL_IMAGES // cfg.BATCH_SIZE
    if cfg.DATASET.TRAIN_TOTAL_IMAGES % cfg.BATCH_SIZE and drop_last != True:
        all_step += 1
    all_step *= (cfg.SOLVER.NUM_EPOCHS - begin_epoch + 1)

    avg_loss = 0.0
    timer = Timer()
    timer.start()
    if begin_epoch > cfg.SOLVER.NUM_EPOCHS:
        raise ValueError(
            ("begin epoch[{}] is larger than cfg.SOLVER.NUM_EPOCHS[{}]").format(
                begin_epoch, cfg.SOLVER.NUM_EPOCHS))

W
wuzewu 已提交
351
    if args.use_mpio:
352
        print_info("Use multiprocess reader")
W
wuzewu 已提交
353
    else:
354
        print_info("Use multi-thread reader")
W
wuzewu 已提交
355

C
chenguowei01 已提交
356 357 358
    # 存储评估时最高mIoU
    best_mIoU = 0

W
wuzewu 已提交
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
    for epoch in range(begin_epoch, cfg.SOLVER.NUM_EPOCHS + 1):
        py_reader.start()
        while True:
            try:
                if args.debug:
                    # Print category IoU and accuracy to check whether the
                    # traning process is corresponed to expectation
                    loss, lr, pred, grts, masks = exe.run(
                        program=compiled_train_prog,
                        fetch_list=fetch_list,
                        return_numpy=True)
                    cm.calculate(pred, grts, masks)
                    avg_loss += np.mean(np.array(loss))
                    global_step += 1

                    if global_step % args.log_steps == 0:
                        speed = args.log_steps / timer.elapsed_time()
                        avg_loss /= args.log_steps
                        category_acc, mean_acc = cm.accuracy()
                        category_iou, mean_iou = cm.mean_iou()

380
                        print_info((
W
wuzewu 已提交
381 382 383 384
                            "epoch={} step={} lr={:.5f} loss={:.4f} acc={:.5f} mIoU={:.5f} step/sec={:.3f} | ETA {}"
                        ).format(epoch, global_step, lr[0], avg_loss, mean_acc,
                                 mean_iou, speed,
                                 calculate_eta(all_step - global_step, speed)))
385 386
                        print_info("Category IoU: ", category_iou)
                        print_info("Category Acc: ", category_acc)
W
wuzewu 已提交
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
                        if args.use_tb:
                            log_writer.add_scalar('Train/mean_iou', mean_iou,
                                                  global_step)
                            log_writer.add_scalar('Train/mean_acc', mean_acc,
                                                  global_step)
                            log_writer.add_scalar('Train/loss', avg_loss,
                                                  global_step)
                            log_writer.add_scalar('Train/lr', lr[0],
                                                  global_step)
                            log_writer.add_scalar('Train/step/sec', speed,
                                                  global_step)
                        sys.stdout.flush()
                        avg_loss = 0.0
                        cm.zero_matrix()
                        timer.restart()
                else:
                    # If not in debug mode, avoid unnessary log and calculate
                    loss, lr = exe.run(
                        program=compiled_train_prog,
                        fetch_list=fetch_list,
                        return_numpy=True)
                    avg_loss += np.mean(np.array(loss))
                    global_step += 1

411
                    if global_step % args.log_steps == 0 and cfg.TRAINER_ID == 0:
W
wuzewu 已提交
412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
                        avg_loss /= args.log_steps
                        speed = args.log_steps / timer.elapsed_time()
                        print((
                            "epoch={} step={} lr={:.5f} loss={:.4f} step/sec={:.3f} | ETA {}"
                        ).format(epoch, global_step, lr[0], avg_loss, speed,
                                 calculate_eta(all_step - global_step, speed)))
                        if args.use_tb:
                            log_writer.add_scalar('Train/loss', avg_loss,
                                                  global_step)
                            log_writer.add_scalar('Train/lr', lr[0],
                                                  global_step)
                            log_writer.add_scalar('Train/speed', speed,
                                                  global_step)
                        sys.stdout.flush()
                        avg_loss = 0.0
                        timer.restart()

            except fluid.core.EOFException:
                py_reader.reset()
                break
            except Exception as e:
                print(e)

435
        if epoch % cfg.TRAIN.SNAPSHOT_EPOCH == 0 and cfg.TRAINER_ID == 0:
W
wuzewu 已提交
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450
            ckpt_dir = save_checkpoint(exe, train_prog, epoch)

            if args.do_eval:
                print("Evaluation start")
                _, mean_iou, _, mean_acc = evaluate(
                    cfg=cfg,
                    ckpt_dir=ckpt_dir,
                    use_gpu=args.use_gpu,
                    use_mpio=args.use_mpio)
                if args.use_tb:
                    log_writer.add_scalar('Evaluate/mean_iou', mean_iou,
                                          global_step)
                    log_writer.add_scalar('Evaluate/mean_acc', mean_acc,
                                          global_step)

C
chenguowei01 已提交
451 452 453 454 455 456 457 458
                # 将最优模型拷贝一份至best_model中
                if mean_iou > best_mIoU:
                    best_mIoU = mean_iou
                    best_model_dir = os.path.join(cfg.TRAIN.MODEL_SAVE_DIR, 'best_model')
                    if os.path.exists(best_model_dir):
                        shutil.rmtree(best_model_dir)
                    shutil.copytree(ckpt_dir, best_model_dir)

W
wuzewu 已提交
459 460 461 462 463 464 465 466 467 468 469
            # Use Tensorboard to visualize results
            if args.use_tb and cfg.DATASET.VIS_FILE_LIST is not None:
                visualize(
                    cfg=cfg,
                    use_gpu=args.use_gpu,
                    vis_file_list=cfg.DATASET.VIS_FILE_LIST,
                    vis_dir="visual",
                    ckpt_dir=ckpt_dir,
                    log_writer=log_writer)

    # save final model
470 471
    if cfg.TRAINER_ID == 0:
        save_checkpoint(exe, train_prog, 'final')
W
wuzewu 已提交
472 473 474 475 476


def main(args):
    if args.cfg_file is not None:
        cfg.update_from_file(args.cfg_file)
477
    if args.opts:
W
wuzewu 已提交
478
        cfg.update_from_list(args.opts)
479 480 481 482

    cfg.TRAINER_ID = int(os.getenv("PADDLE_TRAINER_ID", 0))
    cfg.NUM_TRAINERS = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))

W
wuzewu 已提交
483
    cfg.check_and_infer()
484
    print_info(pprint.pformat(cfg))
W
wuzewu 已提交
485 486 487 488 489 490 491 492 493 494 495 496 497 498
    train(cfg)


if __name__ == '__main__':
    args = parse_args()
    if fluid.core.is_compiled_with_cuda() != True and args.use_gpu == True:
        print(
            "You can not set use_gpu = True in the model because you are using paddlepaddle-cpu."
        )
        print(
            "Please: 1. Install paddlepaddle-gpu to run your models on GPU or 2. Set use_gpu=False to run models on CPU."
        )
        sys.exit(1)
    main(args)