mindspore训练似乎没有用到Ascend卡

宿主机系统：ubuntu 18.04

npu驱动版本：22.0.4

容器镜像：ubuntu 18.04

容器cann-toolkit版本：Ascend-cann-toolkit_6.0.RC1_linux-aarch64

容器mindspore版本：2.0.0-alpha

我刚刚接触深度学习以及mindspore，我根据安装教程安装完npu驱动以及在容器里安装完对应的环境库，版本都是根据对应列表安装的（MindSpore官网），这是安装后的检查结果

然后我运行mindspore快速入门教程中的代码进行训练，但是似乎没有用到Ascend进行训练，使用默认的数据集MNIST_Data（53M）特别的慢

我watch -n 1 npu info可以看到在数据集加载的过程中 "HBM-Usage(MB)"几乎拉满，但是AICore只是0~2%之间跳动，top命令看到的是两个python进程cpu使用率在400%，三个epoch跑完花了792秒，我怀疑整个训练过程没有用到npu

npu-smi info

这是新手教程中( https://www.mindspore.cn/tutorials/zh-CN/r2.0.0-alpha/beginner/quick_start.html )复制下来并运行的代码，为了确保用的是npu，我加了mindspore.set_context(device_target="Ascend")

run.py

#!/usr/bin/python3
import mindspore
from mindspore.dataset import MnistDataset
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset

# Download data from open datasets
from download import download

import time

mindspore.set_context(device_target="Ascend")

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')
print(train_dataset.get_col_names())

def datapipe(dataset, batch_size):
    image_transforms = [
        vision.Rescale(1.0 / 255.0, 0),
        vision.Normalize(mean=(0.1307,), std=(0.3081,)),
        vision.HWC2CHW()
    ]
    label_transform = transforms.TypeCast(mindspore.int32)

    dataset = dataset.map(image_transforms, 'image')
    dataset = dataset.map(label_transform, 'label')
    dataset = dataset.batch(batch_size)
    return dataset

# Map vision transforms and batch dataset
train_dataset = datapipe(train_dataset, 64)
test_dataset = datapipe(test_dataset, 64)

for image, label in test_dataset.create_tuple_iterator():
    print(f"Shape of image [N, C, H, W]: {image.shape} {image.dtype}")
    print(f"Shape of label: {label.shape} {label.dtype}")
    break

for data in test_dataset.create_dict_iterator():
    print(f"Shape of image [N, C, H, W]: {data['image'].shape} {data['image'].dtype}")
    print(f"Shape of label: {data['label'].shape} {data['label'].dtype}")
    break

# 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()
print(model)

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

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

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

# Define function of one-step training
def train_step(data, label):
    (loss, _), grads = grad_fn(data, label)
    optimizer(grads)
    return loss

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)

        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")

start = time.time()

epochs = 3
for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train(model, train_dataset)
    test(model, test_dataset, loss_fn)
print("Done!")
print("epoch run time:", time.time() - start)

# Save checkpoint
mindspore.save_checkpoint(model, "model.ckpt")
print("Saved Model to model.ckpt")

# Instantiate a random initialized model
model = Network()
# Load checkpoint and load parameter to model
param_dict = mindspore.load_checkpoint("model.ckpt")
param_not_load = mindspore.load_param_into_net(model, param_dict)
print(param_not_load)

model.set_train(False)
for data, label in test_dataset:
    pred = model(data)
    predicted = pred.argmax(1)
    print(f'Predicted: "{predicted[:10]}", Actual: "{label[:10]}"')
    break

****************************************************解答*****************************************************

mindspore.set_context(mode=ms.GRAPH_MODE, device_target="Ascend", device_id=0)，我的是这样的device_id需要看一下自己的逻辑ID 不过一般都是0, 之前翻源码，老版 HBM训练 设置device_target="Ascend"；Ai-core训练 设置device_target="DaVinci"，新版里面貌似淘汰这种用法，只能设置device_target="Ascend"处理器会自动选择使用HBM还是Aicore。不过HBM处理加载不管多少数据每次跑满是真的。