使用华为云Ascend910在MNIST上面训练LeNet网络(ModelArts版)

1 使用华为云Ascend910在MNIST上面训练LeNet网络

实验题（需要申请环境）：使用华为云Ascend910在MNIST上面训练LeNet网络，上传loss截图和推理精度截图

• MindSpore官网：https://www.mindspore.cn/

• 开源地址：https://gitee.com/mindspore

• 课程案例：https://gitee.com/mindspore/course

1.1 LeNet

Paper: Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. “Gradient-based learning applied to document recognition.” Proceedings of the IEEE, 86(11):2278-2324, November 1998 https://ieeexplore.ieee.org/document/726791

Data: http://yann.lecun.com/exdb/mnist/

MindSpore实现：https://gitee.com/mindspore/course/tree/master/lenet5

LeNet 结构

1.2 数据集下载

原数据：http://yann.lecun.com/exdb/mnist/

• 方式一：下载源数据
• 方式二：ModelArts + OBS （推荐）
  1）创建桶：华北-北京四、单AZ存储、同名称、其他默认或自定义
  
  2）将下载的数据集上传OBS （ MNIST）
  

2.3 创建ModelArts 开始训练

（1）创建Notebook

• ModelArts–开发环境–Noteboot–创建（按需计费、名称、公共镜像->Ascend-Powered-Engine 1.0 (Python3)、对象存储服务（OBS）-> MNIST）
  
  

（2）进入Jupyter 或者jpyter Lab

• 新建 Mindspore-python3.7 lnet.ipynb
• ModelArts 与OBS 数据集拷贝

# 拷贝自己账户下OBS桶内的数据集至执行容器
import moxing
moxing.file.copy_parallel(src_url="s3://sunxiaobei/MNIST/", dst_url='MNIST/')
# src_url形如's3://OBS/PATH'，为OBS桶中数据集的路径，dst_url为执行容器中的路径

# 拷贝共享的OBS桶内的数据集至执行容器
import moxing
moxing.file.copy_parallel(src_url="s3://share-course/dataset/MNIST/", dst_url='MNIST/')

• 导入模块

import os
# os.environ['DEVICE_ID'] = '0'

import mindspore as ms
import mindspore.context as context
import mindspore.dataset.transforms.c_transforms as C
import mindspore.dataset.vision.c_transforms as CV

from mindspore import nn
from mindspore.train import Model
from mindspore.train.callback import LossMonitor

context.set_context(mode=context.GRAPH_MODE, device_target='Ascend') # Ascend, CPU, GPU

• 数据处理

def create_dataset(data_dir, training=True, batch_size=32, resize=(32, 32),
                   rescale=1/(255*0.3081), shift=-0.1307/0.3081, buffer_size=64):
    data_train = os.path.join(data_dir, 'train') # train set
    data_test = os.path.join(data_dir, 'test') # test set
    ds = ms.dataset.MnistDataset(data_train if training else data_test)

    ds = ds.map(input_columns=["image"], operations=[CV.Resize(resize), CV.Rescale(rescale, shift), CV.HWC2CHW()])
    ds = ds.map(input_columns=["label"], operations=C.TypeCast(ms.int32))
    # When `dataset_sink_mode=True` on Ascend, append `ds = ds.repeat(num_epochs) to the end
    ds = ds.shuffle(buffer_size=buffer_size).batch(batch_size, drop_remainder=True)

    return ds

# 可视化，图片的大小为32x32
import matplotlib.pyplot as plt
ds = create_dataset('MNIST', training=False)
data = ds.create_dict_iterator(output_numpy=True).get_next()
images = data['image']
labels = data['label']

for i in range(1, 5):
    plt.subplot(2, 2, i)
    plt.imshow(images[i][0])
    plt.title('Number: %s' % labels[i])
    plt.xticks([])
plt.show()

• 定义模型

模型结构如下图所示：

class LeNet5(nn.Cell):
    def __init__(self):
        super(LeNet5, self).__init__()
        self.conv1 = nn.Conv2d(1, 6, 5, stride=1, pad_mode='valid')
        self.conv2 = nn.Conv2d(6, 16, 5, stride=1, pad_mode='valid')
        self.relu = nn.ReLU()
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
        self.flatten = nn.Flatten()
        self.fc1 = nn.Dense(400, 120)
        self.fc2 = nn.Dense(120, 84)
        self.fc3 = nn.Dense(84, 10)

    def construct(self, x):
        x = self.relu(self.conv1(x))
        x = self.pool(x)
        x = self.relu(self.conv2(x))
        x = self.pool(x)
        x = self.flatten(x)
        x = self.fc1(x)
        x = self.fc2(x)
        x = self.fc3(x)

        return x

• 训练

# - batch size :32
# - number of epochs :3
# - learning rate : 0.01
# - optimizer: Momentum 0.9

def train(data_dir, lr=0.01, momentum=0.9, num_epochs=3):
    ds_train = create_dataset(data_dir)
    ds_eval = create_dataset(data_dir, training=False)

    net = LeNet5()
    loss = nn.loss.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
    opt = nn.Momentum(net.trainable_params(), lr, momentum)
    loss_cb = LossMonitor(per_print_times=ds_train.get_dataset_size())

    model = Model(net, loss, opt, metrics={'acc', 'loss'})
    # dataset_sink_mode can be True when using Ascend
    model.train(num_epochs, ds_train, callbacks=[loss_cb], dataset_sink_mode=False)
    metrics = model.eval(ds_eval, dataset_sink_mode=False)
    print('Metrics:', metrics)

train('MNIST/')

• 训练结果