mindspore环境配置及运行minst数据集

1、安装GPU版

安装GPU版需要CUDA和cudnn，虚拟环境下的安装看我另一篇文章

在Ubuntu系统的conda虚拟环境下安装cuda和cudnn-CSDN博客

安装完cuda和cudnn后，进入mindspore官网

MindSpore官网

pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.2.12/MindSpore/unified/x86_64/mindspore-2.2.12-cp39-cp39-linux_x86_64.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple

复制命令直接安装即可，这里不建议使用conda命令安装，我这边经常出现找不到镜像的情况，只要是切换到conda创建的环境下，使用pip安装只会安装到当前环境

2、minist数据集实验

这里有官网教程： AI Gallery 资产详情_华为云

官网教程里的代码是用jupyter写的，我把代码改了一下：

删了一部分，增加了指定gpu的代码

import mindspore
from mindspore import nn, value_and_grad
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset
from mindspore import context

# 设置运行设备为GPU，device_id为0
context.set_context(device_target="GPU", device_id=1)

# Download data from open datasets
from download import download
''' 第一次运行时取消注释，下载数据集后加上注释即可，防止重复下载
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/" \
      "notebook/datasets/MNIST_Data.zip"
path = download(url, "./", kind="zip", replace=True)
'''
train_dataset = MnistDataset('MNIST_Data/train')
test_dataset = MnistDataset('MNIST_Data/test')

# 定义数据处理Pipline
def datapipe(dataset, batch_size):
    image_transforms = [
        vision.Rescale(1.0 / 255.0, 0),
        # 归一化
        vision.Normalize(mean=(0.1307,), std=(0.3081,)),
        # 28, 28, 1 --> channel, height, width 1, 28, 28
        vision.HWC2CHW()
    ]
    # uint8 --> int32
    label_transform = transforms.TypeCast(mindspore.int32)

    dataset = dataset.map(image_transforms, 'image')
    dataset = dataset.map(label_transform, 'label')
    # 1, 28, 28 --> 64, 1, 28, 28
    dataset = dataset.batch(batch_size, drop_remainder=True)
    return dataset

train_dataset = datapipe(train_dataset, 64)
test_dataset = datapipe(test_dataset, 64)

# Define model
class Network(nn.Cell):
    def __init__(self):
        super().__init__()
        self.flatten = nn.Flatten()
        self.dense_relu_sequential = nn.SequentialCell(
            nn.Dense(28*28, 512),
            nn.ReLU(),
            nn.Dense(512, 512),
            nn.ReLU(),
            nn.Dense(512, 10)
        )

    def construct(self, x):
        x = self.flatten(x)
        logits = self.dense_relu_sequential(x)
        return logits

model = Network()

# Instantiate loss function and optimizer
loss_fn = nn.CrossEntropyLoss()
optimizer = nn.SGD(model.trainable_params(), 1e-2)

# 1. Define forward function
def forward_fn(data, label):
    logits = model(data)
    loss = loss_fn(logits, label)
    return loss, logits

# 2. Get gradient function
grad_fn = value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)

def train(model, dataset):
    size = dataset.get_dataset_size()
    model.set_train()
    for batch, (data, label) in enumerate(dataset.create_tuple_iterator()):
        #loss = train_step(data, label)
        #logits = model(data)
        #loss = loss_fn(logits, label)
        #grad_fn = value_and_grad(loss_fn(logits, label), None, optimizer.parameters)
        #optimizer(grad_fn)
        (loss, _), grads = grad_fn(data, label)
        optimizer(grads)

        if batch % 100 == 0:
            loss, current = loss.asnumpy(), batch
            print(f"loss: {loss:>7f}  [{current:>3d}/{size:>3d}]")
            
def test(model, dataset, loss_fn):
    num_batches = dataset.get_dataset_size()
    model.set_train(False)
    total, test_loss, correct = 0, 0, 0
    for data, label in dataset.create_tuple_iterator():
        pred = model(data)
        total += len(data)
        test_loss += loss_fn(pred, label).asnumpy()
        correct += (pred.argmax(1) == label).asnumpy().sum()
    test_loss /= num_batches
    correct /= total
    print(f"Test: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
    
epochs = 100
for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train(model, train_dataset)
    test(model, test_dataset, loss_fn)
print("Done!")

运行：我装的cuda的版本是11.8，稍微高了一点，但似乎不影响正常使用

没运行时显存占用：

运行时显存占用：

使用nvidia-smi

使用gpustat

3、部分代码讲解

optimizer = nn.SGD(model.trainable_params(), 1e-2)

# 1. Define forward function
def forward_fn(data, label):
    logits = model(data)
    loss = loss_fn(logits, label)
    return loss, logits

# 2. Get gradient function
grad_fn = value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)

  for batch, (data, label) in enumerate(dataset.create_tuple_iterator()):
        #loss = train_step(data, label)
        #logits = model(data)
        #loss = loss_fn(logits, label)
        #grad_fn = value_and_grad(loss_fn(logits, label), None, optimizer.parameters)
        #optimizer(grad_fn)
        (loss, _), grads = grad_fn(data, label)
        optimizer(grads)

个人理解，主需要介绍的是mindspore中value_and_grad函数的用法，其中，如以value_and_grad函数中的第一个参数forward_fn，这里代表的应该是一个函数对象，这个函数可以是一个嵌套多次的函数，而需要求梯度的部分是optimizer.parameters中的部分，这也意味着，函数对象forward_fn中的参数是大于等于（包含）optimizer.parameters中的部分

还有就是反向传播的过程是需要知道输出和期望的差别，也就是损失，在pytorch里可以直接使用loss.back()，而在value_and_grad中，默认第一个参数forward_fn的返回值相当于loss，如果forward_fn有多个返回值，则需要在value_and_grad中把has_aux=True，这时，默认第一个返回值作为loss参与反向传播

如有问题请留言