tw/npu/train.py

"""
######################## single-dataset train lenet example ########################
This example is a single-dataset training tutorial. If it is a multi-dataset, please refer to the multi-dataset training 
tutorial train_for_multidataset.py. This example cannot be used for multi-datasets!

######################## Instructions for using the training environment ########################
The image of the debugging environment and the image of the training environment are two different images, 
and the working local directories are different. In the training task, you need to pay attention to the following points.
1、(1)The structure of the dataset uploaded for single dataset training in this example
 MNISTData.zip
  ├── test
  │   ├── t10k-images-idx3-ubyte
  │   └── t10k-labels-idx1-ubyte
  └── train
      ├── train-images-idx3-ubyte
      └── train-labels-idx1-ubyte    

  (2)The dataset structure of the single dataset in the training image in this example
  workroot
   ├── data
   |     ├── test
   |     └── train 

2、Single dataset training requires predefined functions
(1)Defines whether the task is a training environment or a debugging environment. 
def WorkEnvironment(environment): 
    if environment == 'train':
        workroot = '/home/work/user-job-dir' #The training task uses this parameter to represent the local path of the training image
    elif environment == 'debug':
        workroot = '/home/ma-user/work'      #The debug task uses this parameter to represent the local path of the debug image
    print('current work mode:' + environment + ', workroot:' + workroot)
    return workroot

(2)Copy single dataset from obs to training image.
def ObsToEnv(obs_data_url, data_dir):
    try:     
        mox.file.copy_parallel(obs_data_url, data_dir)
        print("Successfully Download {} to {}".format(obs_data_url, data_dir))
    except Exception as e:
        print('moxing download {} to {} failed: '.format(obs_data_url, data_dir) + str(e))
    return 

(3)Copy the output model to obs.
def EnvToObs(train_dir, obs_train_url):
    try:
        mox.file.copy_parallel(train_dir, obs_train_url)
        print("Successfully Upload {} to {}".format(train_dir,obs_train_url))
    except Exception as e:
        print('moxing upload {} to {} failed: '.format(train_dir,obs_train_url) + str(e))
    return   

3、3 parameters need to be defined
--data_url is the dataset you selected on the Qizhi platform

--data_url,--train_url,--device_target,These 3 parameters must be defined first in a single dataset task, 
otherwise an error will be reported.    
There is no need to add these parameters to the running parameters of the Qizhi platform, 
because they are predefined in the background, you only need to define them in your code.                 

4、How the dataset is used
A single dataset uses data_url as the input, and data_dir (ie: workroot + '/data') as the calling method
of the dataset in the image.
For details, please refer to the following sample code.

"""

import os
import argparse
import moxing as mox
from config import mnist_cfg as cfg
from dataset import create_dataset
from lenet import LeNet5
import mindspore.nn as nn
from mindspore import context
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
from mindspore.train import Model
from mindspore.nn.metrics import Accuracy
from mindspore.common import set_seed

### Defines whether the task is a training environment or a debugging environment ###
def WorkEnvironment(environment): 
    if environment == 'train':
        workroot = '/home/work/user-job-dir'
    elif environment == 'debug':
        workroot = '/home/work' 
    print('current work mode:' + environment + ', workroot:' + workroot)
    return workroot

### Copy single dataset from obs to training image###
def ObsToEnv(obs_data_url, data_dir):
    try:     
        mox.file.copy_parallel(obs_data_url, data_dir)
        print("Successfully Download {} to {}".format(obs_data_url, data_dir))
    except Exception as e:
        print('moxing download {} to {} failed: '.format(obs_data_url, data_dir) + str(e))
    return 
### Copy the output model to obs###
def EnvToObs(train_dir, obs_train_url):
    try:
        mox.file.copy_parallel(train_dir, obs_train_url)
        print("Successfully Upload {} to {}".format(train_dir,obs_train_url))
    except Exception as e:
        print('moxing upload {} to {} failed: '.format(train_dir,obs_train_url) + str(e))
    return      

### --data_url,--train_url,--device_target,These 3 parameters must be defined first in a single dataset, 
### otherwise an error will be reported.
###There is no need to add these parameters to the running parameters of the Qizhi platform, 
###because they are predefined in the background, you only need to define them in your code.
parser = argparse.ArgumentParser(description='MindSpore Lenet Example')
parser.add_argument('--data_url',
                    help='path to training/inference dataset folder',
                    default= WorkEnvironment('train') + '/data/')

parser.add_argument('--train_url',
                    help='model folder to save/load',
                    default= WorkEnvironment('train') + '/model/')

parser.add_argument(
    '--device_target',
    type=str,
    default="Ascend",
    choices=['Ascend', 'CPU'],
    help='device where the code will be implemented (default: Ascend),if to use the CPU on the Qizhi platform:device_target=CPU')

parser.add_argument('--epoch_size',
                    type=int,
                    default=5,
                    help='Training epochs.')

if __name__ == "__main__":
    args, unknown = parser.parse_known_args()
    ### defining the training environment
    environment = 'train'
    workroot = WorkEnvironment(environment)

    ###Initialize the data and model directories in the training image###
    data_dir = workroot + '/data'  
    train_dir = workroot + '/model'
    if not os.path.exists(data_dir):
        os.makedirs(data_dir)
    if not os.path.exists(train_dir):
        os.makedirs(train_dir)

    ### Copy the dataset from obs to the training image ###   
    ObsToEnv(args.data_url,data_dir)
    
    ###Specifies the device CPU or Ascend NPU used for training###
    context.set_context(mode=context.GRAPH_MODE,
                        device_target=args.device_target)
    ds_train = create_dataset(os.path.join(data_dir, "train"),
                              cfg.batch_size)
    if ds_train.get_dataset_size() == 0:
        raise ValueError(
            "Please check dataset size > 0 and batch_size <= dataset size")
    network = LeNet5(cfg.num_classes)
    net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
    net_opt = nn.Momentum(network.trainable_params(), cfg.lr, cfg.momentum)
    time_cb = TimeMonitor(data_size=ds_train.get_dataset_size())

    if args.device_target != "Ascend":
        model = Model(network,
                      net_loss,
                      net_opt,
                      metrics={"accuracy": Accuracy()})
    else:
        model = Model(network,
                      net_loss,
                      net_opt,
                      metrics={"accuracy": Accuracy()},
                      amp_level="O2")

    config_ck = CheckpointConfig(
        save_checkpoint_steps=cfg.save_checkpoint_steps,
        keep_checkpoint_max=cfg.keep_checkpoint_max)
    ckpoint_cb = ModelCheckpoint(prefix="checkpoint_lenet",
                                 directory=train_dir,
                                 config=config_ck)
    print("============== Starting Training ==============")
    epoch_size = cfg['epoch_size']
    if (args.epoch_size):
        epoch_size = args.epoch_size
        print('epoch_size is: ', epoch_size)

    model.train(epoch_size,
                ds_train,
                callbacks=[time_cb, ckpoint_cb,
                           LossMonitor()])

    ###Copy the trained model data from the local running environment back to obs,
    ###and download it in the training task corresponding to the Qizhi platform
    EnvToObs(train_dir, args.train_url)