View Source on Gitee

Advanced Encapsulation: Model

Generally, defining a training and evaluation network and running it directly can meet basic requirements.

On the one hand, Model can simplify code to some extent. For example, you do not need to manually traverse datasets. If you do not need to customize nn.TrainOneStepCell, you can use Model to automatically build a training network. You can use the eval API of Model to evaluate the model and directly output the evaluation result. You do not need to manually invoke the clear, update, and eval functions of evaluation metrics.

On the other hand, Model provides many high-level functions, such as data offloading and mixed precision. Without the help of Model, it takes a long time to customize these functions by referring to Model.

The following describes MindSpore models and how to use Model for model training, evaluation, and inference.

import mindspore
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset
from mindspore.train import Model, CheckpointConfig, ModelCheckpoint, LossMonitor

Introduction to Model

Model is a high-level API provided by MindSpore for model training, evaluation, and inference. The common parameters of the API are as follows:

  • network: neural network used for training or inference.

  • loss_fn: used loss function.

  • optimizer: used optimizer.

  • metrics: evaluation function used for model evaluation.

  • eval_network: network used for model evaluation. If the network is not defined, Model uses network and loss_fn for encapsulation.

Model provides the following APIs for model training, evaluation, and inference:

  • fit: Evaluate the model while training.

  • train: used for model training on the training set.

  • eval: used to evaluate the model on the evaluation set.

  • predict: performs inference on a group of input data and outputs the prediction result.

Using the Model API

For a neural network in a simple scenario, you can specify the feedforward network network, loss function loss_fn, optimizer optimizer, and evaluation function metrics when defining Model.

Downloading and Processing Dataset

# Download data from open datasets
from download import download

url = "" \
path = download(url, "./", kind="zip", replace=True)

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

    dataset = MnistDataset(path)
    dataset =, 'image')
    dataset =, 'label')
    dataset = dataset.batch(batch_size)
    return dataset

train_dataset = datapipe('MNIST_Data/train', 64)
test_dataset = datapipe('MNIST_Data/test', 64)

Defining Model

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

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

model = Network()

Defining Loss Function and Optimizer

To train neural network model, loss function and optimizer function need to be defined.

  • The loss function here uses CrossEntropy Loss .

  • The optimizer uses SGD here.

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

Training and Saving Model

Before starting the training, MindSpore needs to state in advance whether the network model needs to save the intermediate process and results during the training process. Therefore, ModelCheckpoint is used to save the network model and parameters for subsequent fine tuning.

steps_per_epoch = train_dataset.get_dataset_size()
config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch)

ckpt_callback = ModelCheckpoint(prefix="mnist", directory="./checkpoint", config=config)
loss_callback = LossMonitor(steps_per_epoch)

The interface provided by MindSpore makes it easy to train and evaluate the network, and LossMonitor can monitor the changes of loss values during the training process.

trainer = Model(model, loss_fn=loss_fn, optimizer=optimizer, metrics={'accuracy'}), train_dataset, test_dataset, callbacks=[ckpt_callback, loss_callback])
epoch: 1 step: 938, loss is 0.602992594242096
Eval result: epoch 1, metrics: {'accuracy': 0.8435}
epoch: 2 step: 938, loss is 0.2797124981880188
Eval result: epoch 2, metrics: {'accuracy': 0.9003}
epoch: 3 step: 938, loss is 0.32015785574913025
Eval result: epoch 3, metrics: {'accuracy': 0.9179}
epoch: 4 step: 938, loss is 0.17153620719909668
Eval result: epoch 4, metrics: {'accuracy': 0.9308}
epoch: 5 step: 938, loss is 0.18772485852241516
Eval result: epoch 5, metrics: {'accuracy': 0.9382}
epoch: 6 step: 938, loss is 0.45641791820526123
Eval result: epoch 6, metrics: {'accuracy': 0.946}
epoch: 7 step: 938, loss is 0.11519066989421844
Eval result: epoch 7, metrics: {'accuracy': 0.9506}
epoch: 8 step: 938, loss is 0.43486487865448
Eval result: epoch 8, metrics: {'accuracy': 0.9555}
epoch: 9 step: 938, loss is 0.1941455900669098
Eval result: epoch 9, metrics: {'accuracy': 0.9588}
epoch: 10 step: 938, loss is 0.13441434502601624
Eval result: epoch 10, metrics: {'accuracy': 0.9632}

During training, the loss value will be printed, and the loss value will fluctuate, but in general, the loss value will gradually decrease and the accuracy will gradually improve. The loss values run by each person are random and not necessarily identical.

The results obtained by running the test dataset of the model verify the generalization ability of the model:

  1. Use model.eval to read in the test dataset.

  2. Use the saved model parameters for reasoning.

acc = trainer.eval(test_dataset)
{'accuracy': 0.9632}

The model accuracy data can be seen from the print information. In the example, the accuracy data reaches more than 95%, and the model quality is good. As the number of network iterations increases, the accuracy of the model will be further improved.