|
- """train resnet."""
- import os
- import time
- import argparse
- import ast
- import numpy as np
- from mindspore import context
- from mindspore import Tensor
- from mindspore.nn.optim.momentum import Momentum
- from mindspore.train.model import Model
- from mindspore.context import ParallelMode
- from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
- from mindspore.train.loss_scale_manager import FixedLossScaleManager
- from mindspore.train.serialization import load_checkpoint, load_param_into_net
- from mindspore.common import set_seed
- from mindspore.communication.management import init
- from mindspore.train.callback import Callback
-
- from src.loss import Softmaxloss
- from src.loss import Tripletloss
- from src.loss import mix_loss
- from src.lr_generator import get_lr
- from src.resnet import resnet50 as resnet
- from src.config import config
- from src.dataset import create_dataset_triplet as create_dataset
- set_seed(1)
-
- workroot = '/home/work/user-job-dir'
-
- parser = argparse.ArgumentParser(description='Image classification')
- # modelarts parameter
- parser.add_argument('--train_url', type=str, default=workroot + '/model/', help='Train output path')
- parser.add_argument('--data_url', type=str, default=workroot + '/data/', help='Dataset path')
- parser.add_argument('--ckpt_url', type=str, default=workroot + '/ckpt/', help='Pretrained ckpt path')
- parser.add_argument('--checkpoint_name', type=str, default='resnet-120_625.ckpt', help='Checkpoint file')
- parser.add_argument('--loss_name', type=str, default='softmax',
- help='loss name: softmax(pretrained) triplet quadruplet')
- # Ascend parameter
- parser.add_argument('--dataset_path', type=str, default=None, help='Dataset path')
- parser.add_argument('--ckpt_path', type=str, default=None, help='ckpt path name')
- parser.add_argument('--run_distribute', type=ast.literal_eval, default=False, help='Run distribute')
- parser.add_argument('--device_id', type=int, default=0, help='Device id')
- parser.add_argument('--run_modelarts', type=ast.literal_eval, default=True, help='Run distribute')
- parser.add_argument('--device_target',type=str,default="Ascend",choices=['Ascend', 'CPU'],help='device where the code will be implemented (default: CPU),若要在启智平台上使用NPU,需要在启智平台训练界面上加上运行参数device_target=Ascend')
-
- args_opt = parser.parse_args()
-
- class Monitor(Callback):
- """Monitor"""
- def __init__(self, lr_init=None):
- super(Monitor, self).__init__()
- self.lr_init = lr_init
- self.lr_init_len = len(lr_init)
- def epoch_begin(self, run_context):
- self.losses = []
- self.epoch_time = time.time()
- def epoch_end(self, run_context):
- cb_params = run_context.original_args()
- epoch_mseconds = (time.time() - self.epoch_time) * 1000
- per_step_mseconds = epoch_mseconds / cb_params.batch_num
- print("epoch time: {:5.3f}, per step time: {:5.3f}, avg loss: {:8.5f}"
- .format(epoch_mseconds, per_step_mseconds, np.mean(self.losses)))
- print('batch_size:', config.batch_size, 'epochs_size:', config.epoch_size,
- 'lr_model:', config.lr_decay_mode, 'lr:', config.lr_max, 'step_size:', step_size)
- def step_begin(self, run_context):
- self.step_time = time.time()
- def step_end(self, run_context):
- """step_end"""
- cb_params = run_context.original_args()
- step_mseconds = (time.time() - self.step_time) * 1000
- step_loss = cb_params.net_outputs
- if isinstance(step_loss, (tuple, list)) and isinstance(step_loss[0], Tensor):
- step_loss = step_loss[0]
- if isinstance(step_loss, Tensor):
- step_loss = np.mean(step_loss.asnumpy())
- self.losses.append(step_loss)
- cur_step_in_epoch = (cb_params.cur_step_num - 1) % cb_params.batch_num
- print("epochs: [{:3d}/{:3d}], step:[{:5d}/{:5d}], loss:[{:8.5f}/{:8.5f}], time:[{:5.3f}], lr:[{:8.5f}]".format(
- cb_params.cur_epoch_num, config.epoch_size, cur_step_in_epoch, cb_params.batch_num, step_loss,
- np.mean(self.losses), step_mseconds, self.lr_init[cb_params.cur_step_num - 1]))
-
- def init_group_params(net):
- decayed_params = []
- no_decayed_params = []
- for param in net.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': net.trainable_params()}]
- return group_params
-
- if __name__ == '__main__':
-
-
-
- # init distributed
- if args_opt.run_modelarts:
- import moxing as mox
- device_id = int(os.getenv('DEVICE_ID'))
- device_num = int(os.getenv('RANK_SIZE'))
- context.set_context(device_id=device_id)
- data_dir = workroot + '/data'
- train_dir = workroot + '/model/'
-
- #初始化数据存放目录
- if not os.path.exists(data_dir):
- os.mkdir(data_dir)
-
- #初始化模型存放目录
-
-
- print("________________train_model_URL_______________________")
- train_dir = workroot + '/model/'
- if not os.path.exists(train_dir):
- os.mkdir(train_dir)
- local_train_url = train_dir
- print(os.path.exists(train_dir))
-
- #将数据集从local拷贝到推理镜像中:
- local_data_url = args_opt.data_url
- print(os.path.exists(local_data_url))
- args_opt.data_url = '/home/work/user-job-dir/data/'
- try:
- mox.file.copy_parallel(local_data_url, args_opt.data_url)
- print("Successfully Download {} to {}".format(local_data_url,
- args_opt.data_url))
- except Exception as e:
- print('moxing download {} to {} failed: '.format(
- local_data_url, args_opt.data_url) + str(e))
-
- #将模型文件从local拷贝到推理镜像中:
- local_ckpt_url = args_opt.ckpt_url
- print(os.path.exists(local_ckpt_url))
- args_opt.ckpt_url = '/home/work/user-job-dir/checkpoint.ckpt'
- try:
- mox.file.copy(local_ckpt_url, args_opt.ckpt_url)
- print("Successfully Download {} to {}".format(local_ckpt_url,
- args_opt.ckpt_url))
- except Exception as e:
- print('moxing download {} to {} failed: '.format(
- local_ckpt_url, args_opt.ckpt_url) + str(e))
- DATA_DIR = '/home/work/user-job-dir/data/imagenet/train'
-
-
- # create dataset
- TRAIN_LIST = DATA_DIR
-
- dataset = create_dataset(dataset_path=DATA_DIR, do_train=True,
- batch_size=config.batch_size, train_image_size=config.train_image_size,
- eval_image_size=config.eval_image_size, target='Ascend',
- distribute=config.run_distribute
- )
- step_size = dataset.get_dataset_size()
-
- # define net
- net = resnet(class_num=config.class_num)
-
- # init weight
-
- checkpoint_path = os.path.join(local_ckpt_url, args_opt.checkpoint_name)
-
- # init lr
- lr = Tensor(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))
-
-
- # define opt
- group_params = init_group_params(net)
- opt = Momentum(group_params, lr, config.momentum, loss_scale=config.loss_scale)
-
-
- # define loss, model
- loss = mix_loss()
- loss_scale = FixedLossScaleManager(config.loss_scale, drop_overflow_update=False)
-
-
- model = Model(net, loss_fn=loss, optimizer=opt, loss_scale_manager=loss_scale, metrics=None,
- amp_level="O2", keep_batchnorm_fp32=False)
-
-
- #define callback
- cb = []
- if config.save_checkpoint:
- config_ck = CheckpointConfig(save_checkpoint_steps=config.save_checkpoint_epochs * step_size,
- keep_checkpoint_max=config.keep_checkpoint_max)
-
- check_name = 'ResNet50_'
- save_ckpt_path = os.path.join(local_train_url, str(device_id) +'/')
- ckpt_cb = ModelCheckpoint(prefix=check_name, directory=save_ckpt_path, config=config_ck)
- cb += [ckpt_cb]
- cb += [Monitor(lr_init=lr.asnumpy())]
-
- # train model
- model.train(1, dataset, callbacks=cb, dataset_sink_mode=False)
-
- mox.file.copy_parallel(src_url=local_train_url, dst_url=args_opt.train_url)
|