|
- """
- 本示例是多数据集训练的教程,如果是单数据集,请参考单数据集训练教程train.py,本示例不能用于单数据集!
- ######################## train lenet example ########################
- train lenet and get network model files(.ckpt)
- """
- """
- ######################## 训练环境使用说明 ########################
- 1、(1)本示例中多数据集训练上传的数据集结构
- MNISTData.zip
- ├── test
- │ ├── t10k-images-idx3-ubyte
- │ └── t10k-labels-idx1-ubyte
- └── train
- ├── train-images-idx3-ubyte
- └── train-labels-idx1-ubyte
-
- checkpoint_lenet-1_1875.zip
- ├── checkpoint_lenet-1_1875.ckpt
-
- (2)本示例中多数据集在训练镜像中的数据集结构
- workroot
- ├── MNISTData
- | ├── test
- | └── train
- └── checkpoint_lenet-1_1875
- ├── checkpoint_lenet-1_1875.ckpt
-
- 2、多数据集训练需要预定义的函数:
- ### (1)定义任务是训练环境还是调试环境(可参考的写法)###
- def WorkEnvironment(environment):
- if environment == 'train':
- workroot = '/home/work/user-job-dir' # 训练任务使用该参数,表示训练镜像本地路径
- elif environment == 'debug':
- workroot = '/home/ma-user/work' #调试任务使用该参数,表示调试镜像本地路径
- print('current work mode:' + environment + ', workroot:' + workroot)
- return workroot
-
- ### (2)将多个数据集从obs拷贝到训练镜像中 (可参考的写法)###
- def MultiObsToEnv(multi_data_url, workroot):
- multi_data_json = json.loads(multi_data_url) #将multi_data_url解析
- for i in range(len(multi_data_json)):
- path = workroot + "/" + multi_data_json[i]["dataset_name"]
- if not os.path.exists(path):
- os.makedirs(path)
- try:
- mox.file.copy_parallel(multi_data_json[i]["dataset_url"], path)
- print("Successfully Download {} to {}".format(multi_data_json[i]["dataset_url"],
- path))
- except Exception as e:
- print('moxing download {} to {} failed: '.format(
- multi_data_json[i]["dataset_url"], path) + str(e))
- return
-
- ***在本示例中MultiObsToEnv函数的输入输出:
- multi_data_url的输入:
- [
- {
- "dataset_url": "s3://test-opendata/attachment/e/a/eae3a316-42d6-4a43-a484-1fa573eab388e
- ae3a316-42d6-4a43-a484-1fa573eab388/", #数据集的obs路径
- "dataset_name": "MNIST_Data" #数据集的名称
- },
- {
- "dataset_url": "s3://test-opendata/attachment/2/c/2c59be66-64ec-41ca-b311-f51a486eabf82c
- 59be66-64ec-41ca-b311-f51a486eabf8/",
- "dataset_name": "checkpoint_lenet-1_1875"
- }
- ]
- 输出:
- MultiObsToEnv函数的目的是为了将多个数据集从obs拷贝到训练镜像中,并构建在训练镜像中数据集路径:
- 如本示例中的MNIST_Data数据集的路径是/home/work/user-job-dir/MNISTData,
- checkpoint_lenet-1_1875数据集的路径是/home/work/user-job-dir/checkpoint_lenet-1_1875
-
- ### (3)将输出的模型拷贝到obs(可参考的写法)###
- def EnvToObs(obs_train_url, train_dir):
- 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、需要定义4个参数
- define 4 parameters for running on modelArts;
- --data_url,--multi_data_url,--train_url,--device_target,这4个参数在多数据集中是必须先定义的,缺一不可,否则会报错
-
- 3、数据集的使用方式
- 多数据集使用multi_data_url作为输入,workroot + 数据集名称 + 数据集内文件或文件夹名称 作为镜像中数据集的调用方式,
- 具体请参考以下示例代码
- """
-
- #具体示例代码如下:
-
- 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 json
- 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
- from mindspore import load_checkpoint, load_param_into_net
-
- ######################## 定义任务是训练环境还是调试环境(固定写法)########################
- def WorkEnvironment(environment):
- if environment == 'train':
- workroot = '/home/work/user-job-dir' # 训练任务使用该参数,表示训练镜像本地路径
- elif environment == 'debug':
- workroot = '/home/ma-user/work' #调试任务使用该参数,表示调试镜像本地路径
- print('current work mode:' + environment + ', workroot:' + workroot)
- return workroot
-
- ######################## 将多个数据集从obs拷贝到训练镜像中 (固定写法)########################
- def MultiObsToEnv(multi_data_url, workroot):
- multi_data_json = json.loads(multi_data_url) #将multi_data_url解析
- for i in range(len(multi_data_json)):
- path = workroot + "/" + multi_data_json[i]["dataset_name"]
- if not os.path.exists(path):
- os.makedirs(path)
- try:
- mox.file.copy_parallel(multi_data_json[i]["dataset_url"], path)
- print("Successfully Download {} to {}".format(multi_data_json[i]["dataset_url"],
- path))
- except Exception as e:
- print('moxing download {} to {} failed: '.format(
- multi_data_json[i]["dataset_url"], path) + str(e))
- return
- ######################## 将输出的模型拷贝到obs(固定写法)########################
- def EnvToObs(obs_train_url, train_dir):
- 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
-
-
- parser = argparse.ArgumentParser(description='MindSpore Lenet Example')
- # define 4 parameters for running on modelArts;--data_url,--multi_data_url,--train_url,--device_target
- # 这4个参数在多数据集中是必须先定义的,缺一不可,否则会报错
- # data_url,train_url,device_target是固定用于在modelarts上训练的参数,表示数据集的路径和输出模型的路径,multi_data_url是多数据集路径和名称的json字符串
- parser.add_argument('--data_url',
- help='path to training/inference dataset folder',
- default= WorkEnvironment('train') + '/data/')
-
- parser.add_argument('--multi_data_url',
- help='path to multi dataset',
- default= WorkEnvironment('train'))
-
- 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: 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')
-
- parser.add_argument('--epoch_size',
- type=int,
- default=5,
- help='Training epochs.')
-
- if __name__ == "__main__":
- args = parser.parse_args()
- #多数据集使用,先执行WorkEnv函数和GetMultiDataPath函数,将多个数据集从obs拷贝到训练镜像中
- environment = 'train'
- workroot = WorkEnvironment(environment)
- MultiObsToEnv(args.multi_data_url, workroot) #多数据集时必须执行此函数将多个数据集从obs拷贝到训练镜像中
-
- train_dir = workroot + '/model' #先在训练镜像中定义输出路径
- if not os.path.exists(train_dir):
- os.makedirs(train_dir)
-
- context.set_context(mode=context.GRAPH_MODE,
- device_target=args.device_target)
- #这里使用了数据集路径workroot + "/MNIST_Data" +/train
- ds_train = create_dataset(os.path.join(workroot + "/MNISTData", "train"),
- cfg.batch_size)
- if ds_train.get_dataset_size() == 0:
- raise ValueError(
- "Please check dataset size > 0 and batch_size <= dataset size")
- #创建网络
- 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())
-
- #加载已经训练好的模型,请根据需求修改此部分模型的载入, 这里使用了数据集路径workroot + "/checkpoint_lenet-1_1875"+"/checkpoint_lenet-1_1875.ckpt"
- load_param_into_net(network, load_checkpoint(os.path.join(workroot + "/checkpoint_lenet-1_1875", "checkpoint_lenet-1_1875.ckpt")))
-
- 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)
- #定义模型输出路径
- ckpoint_cb = ModelCheckpoint(prefix="checkpoint_lenet",
- directory=train_dir,
- config=config_ck)
- #开始训练
- 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()])
- #将输出的模型拷贝到obs,把训练后的模型数据从本地的运行环境拷贝回obs,在启智平台相对应的训练任务中会提供下载
- EnvToObs(train_dir, args.train_url)
-
|