|
- from mindspore import context
- import os
- import random
- import argparse
- import ast
- import numpy as np
- from mindspore import Tensor
- from mindspore import dataset as de
- import mindspore.ops as ops
- import moxing as mox
- from mindspore import dtype as mstype
- from mindspore.parallel._auto_parallel_context import auto_parallel_context
- from mindspore.nn.optim.momentum import Momentum
- from mindspore.train.model import Model, ParallelMode
- from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
- from mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
- from mindspore.train.loss_scale_manager import FixedLossScaleManager
- from mindspore.train.serialization import load_checkpoint, load_param_into_net
- from mindspore.communication.management import init, get_rank, get_group_size
- import mindspore.nn as nn
- import mindspore.common.initializer as weight_init
- from easydict import EasyDict
- from src.network.dataset import create_dataset_Cifar10
- from src.network.dataset import create_dataset_ImageNet
- from src.network.lr_generator import get_lr, power_lr
- from src.network.HarDNet import HarDNet68
- from src.network.optimizers import get_param_groups
-
- parser = argparse.ArgumentParser(description='Image classification with HarDNet on Imagenet')
- """
- parser.add_argument('--dataset_path', type=str, default='/opt_data/xidian_wks/mmq/cifar-10-batches-bin',
- help='Dataset path')
- """
- # parser.add_argument('--dataset_path', type=str, default='/opt_data/xidian_wks/imagenet_original/train/',
- # help='Dataset path')
- parser.add_argument('--run_distribute', type=ast.literal_eval, default=True, help='Run distribute')
- parser.add_argument('--device_target', type=str, default='Ascend', help='Device target')
- parser.add_argument('--device_num', type=int, default=8, help='Device num')
- parser.add_argument('--pre_trained', type=str, default=None)
- parser.add_argument('--train_url', type=str)
- parser.add_argument('--data_url', type=str, default='/opt_data/xidian_wks/imagenet_original/train/',
- help='Dataset path')
- args = parser.parse_args()
-
- config = EasyDict({
- "class_num": 10,
- "batch_size": 256,
- "loss_scale": 1024,
- "momentum": 0.9,
- "weight_decay": 6e-5,
- "epoch_size": 50,
- "pretrain_epoch_size": 0,
- "save_checkpoint": True,
- "save_checkpoint_epochs": 10,
- "keep_checkpoint_max": 10,
- "warmup_epochs": 5,
- "lr_decay_mode": "cosine",
- "lr_init": 0.05,
- "lr_end": 0.00001,
- "lr_max": 0.1
- })
-
- if __name__ == '__main__':
- target = args.device_target
- config.save_checkpoint_path = '/cache/train_output'
- ckpt_save_dir = config.save_checkpoint_path
-
- # init context
- device_id = int(os.getenv('DEVICE_ID'))
- context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
- context.set_context(device_id=device_id)
-
- # download dataset from obs to cache
- mox.file.copy_parallel(src_url=args.data_url, dst_url='/cache/data_path')
-
- # create dataset
- dataset_path = '/cache/data_path'
- # train_dataset = ImageNet(args.dataset_path)
- # train_dataset = create_dataset_ImageNet(dataset_path=dataset_path,
- # do_train=True,
- # repeat_num=1,
- # batch_size=config.batch_size,
- # target=target)
- train_dataset = create_dataset_Cifar10(dataset_path=dataset_path,
- do_train=True,
- repeat_num=1,
- batch_size=config.batch_size,
- target=target)
-
- step_size = train_dataset.get_dataset_size()
-
- # init lr
- lr = get_lr(lr_init=config.lr_init,
- lr_end=config.lr_end,
- lr_max=config.lr_max,
- warmup_epochs=config.warmup_epochs,
- total_epochs=config.epoch_size,
- steps_per_epoch=step_size,
- lr_decay_mode=config.lr_decay_mode)
- lr = Tensor(lr)
-
- # define net
- network = HarDNet68(num_classes=config.class_num)
- print("----network----")
-
- # init weight
- if args.pre_trained:
- param_dict = load_checkpoint(args.pre_ckpt_path)
- load_param_into_net(network, param_dict)
- else:
- for _, cell in network.cells_and_names():
- if isinstance(cell, nn.Conv2d):
- # cell.weight.default_input = weight_init.initializer(weight_init.XavierUniform(),
- # cell.weight.shape,
- # cell.weight.dtype)
-
- cell.weight.default_input = weight_init.initializer(weight_init.HeNormal(mode='fan_out', ),
- cell.weight.shape,
- cell.weight.dtype)
-
- if isinstance(cell, nn.Dense):
- cell.weight.default_input = weight_init.initializer(weight_init.TruncatedNormal(),
- cell.weight.shape,
- cell.weight.dtype)
-
- # define opt
- decayed_params = []
- no_decayed_params = []
- for param in network.trainable_params():
- if 'beta' not in param.name and 'gamma' not in param.name and 'bias' not in param.name:
- decayed_params.append(param)
- else:
- no_decayed_params.append(param)
-
- group_params = [{'params': decayed_params, 'weight_decay': config.weight_decay},
- {'params': no_decayed_params},
- {'order_params': network.trainable_params()}]
-
- net_opt = nn.SGD(filter(lambda x: x.requires_grad, network.get_parameters()),
- learning_rate=lr,
- momentum=config.momentum,
- weight_decay=config.weight_decay,
- loss_scale=config.loss_scale)
-
- # net_opt = Momentum(group_params, lr, config.momentum, loss_scale=config.loss_scale)
-
- # define loss
- loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
- loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
-
- model = Model(network, loss_fn=loss, optimizer=net_opt, loss_scale_manager=loss_scale, metrics={'acc'})
-
- # define callbacks
- time_cb = TimeMonitor(data_size=train_dataset.get_dataset_size())
- loss_cb = LossMonitor()
- cb = [time_cb, loss_cb]
- if config.save_checkpoint:
- config_ck = CheckpointConfig(save_checkpoint_steps= \
- config.save_checkpoint_epochs * \
- train_dataset.get_dataset_size(),
- keep_checkpoint_max=config.keep_checkpoint_max)
- ckpt_cb = ModelCheckpoint(prefix="HarDNet",
- directory='/cache/train_output/device_' + os.getenv('DEVICE_ID') + '/',
- config=config_ck)
- cb += [ckpt_cb]
- for m in network.cells_and_names():
- print(m[0])
-
- print("\n\n========================")
- #print("Dataset path: {}".format(args.dataset_path))
- print("Total epoch: {}".format(config.epoch_size))
- print("Batch size: {}".format(config.batch_size))
- print("Class num: {}".format(config.class_num))
- print("======= Multiple Training begin========")
- model.train(config.epoch_size, train_dataset,
- callbacks=cb, dataset_sink_mode=True)
- mox.file.copy_parallel(src_url='/cache/train_output', dst_url=args.train_url)
|