Quick Start
Since developers may perform custom model and case development in OrangePi AIpro (hereinafter: OrangePi Development Board), this chapter illustrates the development considerations in the OrangePi Development Board through a handwritten digit recognition case based on MindSpore.
from mindspore import nn
from mindspore.dataset import vision, transforms
from mindspore.dataset import MnistDataset
Setting Running Environment
Since Mindspore 2.3 and earlier versions do not support dynamic memory request on demand, and CANN lacks corresponding dynamic operators, we need to set the environment via set_context before executing the cases. The above problems will be solved in the future with the continuous updating of the version, so that the cases can be executed directly without the need to configure the environment.
max_device_memory="2GB" : Set the maximum memory available to the device to 2GB.
mode=mindspore.GRAPH_MODE : Indicates running in GRAPH_MODE mode.
device_target="Ascend" : Indicates that the target device to be run is Ascend.
jit_config={"jit_level":"O2"} : The compilation optimization level turns on extreme performance optimization and uses sinking execution.
scend_config={"precision_mode":"allow_mix_precision"} : Auto mixed-precision, which automatically reduces the precision of some operators to float16 or bfloat16.
import mindspore
mindspore.set_context(max_device_memory="2GB", mode=mindspore.GRAPH_MODE, device_target="Ascend", jit_config={"jit_level":"O2"}, ascend_config={"precision_mode":"allow_mix_precision"})
Processing Dataset
MindSpore provides a Pipeline-based data engine to realize efficient data preprocessing through data loading and processing to realize efficient data preprocessing. In this case, we use the Mnist dataset, which is automatically downloaded and then preprocessed using the data transforms provided by mindspore.dataset.
# Download data from open datasets
from download import download
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/" \
"notebook/datasets/MNIST_Data.zip"
path = download(url, "./", kind="zip", replace=True)
Downloading data from https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/MNIST_Data.zip (10.3 MB)
file_sizes: 100%|██████████████████████████| 10.8M/10.8M [00:01<00:00, 7.63MB/s]
Extracting zip file...
Successfully downloaded / unzipped to ./
The MNIST dataset catalog is structured as follows:
MNIST_Data
└── train
├── train-images-idx3-ubyte (60000 training images)
├── train-labels-idx1-ubyte (60000 training labels)
└── test
├── t10k-images-idx3-ubyte (10000 test images)
├── t10k-labels-idx1-ubyte (10000 test labels)
After the data download is complete, the dataset object is obtained.
train_dataset = MnistDataset('MNIST_Data/train')
test_dataset = MnistDataset('MNIST_Data/test')
Prints the names of the data columns contained in the dataset for dataset preprocessing.
print(train_dataset.get_col_names())
['image', 'label']
MindSpore dataset using data processing pipeline needs to specify map, batch, shuffle and other operations. Here we use map to transform the image data and labels by scaling the input image to 1/255, normalizing it according to the mean value of 0.1307 and standard deviation value of 0.3081, and then packing the processed dataset into a batch of size 64.
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)
The dataset can be accessed iteratively using create_tuple_iterator or create_dict_iterator to see the shape and datatype of the data and labels.
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
Shape of image [N, C, H, W]: (64, 1, 28, 28) Float32
Shape of label: (64,) Int32
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
Shape of image [N, C, H, W]: (64, 1, 28, 28) Float32
Shape of label: (64,) Int32
# 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)
Network<
(flatten): Flatten<>
(dense_relu_sequential): SequentialCell<
(0): Dense<input_channels=784, output_channels=512, has_bias=True>
(1): ReLU<>
(2): Dense<input_channels=512, output_channels=512, has_bias=True>
(3): ReLU<>
(4): Dense<input_channels=512, output_channels=10, has_bias=True>
>
>
Model Training
In model training, a complete training process (STEP) requires the realization of the following three steps:
Forward computation: the model predicts the results, and with the correct label to find the predicted loss.
Backpropagation: automatically solves for the gradients of the model parameters with respect to the loss, using an automatic differentiation mechanism.
Parameter optimization: update the gradients to the parameters.
MindSpore uses a functional automatic differentiation mechanism, so for the above steps need to be implemented:
Define the forward computation function.
Use value_and_grad to obtain the gradient computation function by functional transformation.
Define the training function and use set_train to set to training mode, perform forward computation, backpropagation and parameter optimization.
# 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 = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)
# 3. 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}]")
In addition to training, we define test functions to evaluate the performance of the model.
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")
The training process requires multiple iterations of the dataset, and a complete iteration is called a round (epoch). In each round, the training set is traversed for training and at the end the test set is used for prediction. Printing the loss value and prediction accuracy for each round, we can see that the loss is decreasing and Accuracy is increasing.
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!")
Epoch 1
-------------------------------
loss: 2.302898 [ 0/938]
loss: 1.729961 [100/938]
loss: 0.865714 [200/938]
loss: 0.782822 [300/938]
loss: 0.389282 [400/938]
loss: 0.293149 [500/938]
loss: 0.474819 [600/938]
loss: 0.242542 [700/938]
loss: 0.542277 [800/938]
loss: 0.342929 [900/938]
Test:
Accuracy: 90.7%, Avg loss: 0.321954
Epoch 2
-------------------------------
loss: 0.249492 [ 0/938]
loss: 0.347967 [100/938]
loss: 0.220382 [200/938]
loss: 0.308149 [300/938]
loss: 0.353044 [400/938]
loss: 0.392116 [500/938]
loss: 0.396438 [600/938]
loss: 0.231412 [700/938]
loss: 0.194819 [800/938]
loss: 0.228290 [900/938]
Test:
Accuracy: 93.0%, Avg loss: 0.249993
Epoch 3
-------------------------------
loss: 0.343888 [ 0/938]
loss: 0.307786 [100/938]
loss: 0.153425 [200/938]
loss: 0.254917 [300/938]
loss: 0.198072 [400/938]
loss: 0.108963 [500/938]
loss: 0.202033 [600/938]
loss: 0.340418 [700/938]
loss: 0.144911 [800/938]
loss: 0.175447 [900/938]
Test:
Accuracy: 93.7%, Avg loss: 0.212180
Done!
Saving Models
Once the model is trained, its parameters need to be saved.
# Save checkpoint
mindspore.save_checkpoint(model, "model.ckpt")
print("Saved Model to model.ckpt")
Saved Model to model.ckpt
Loading Models
Loading the saved weights is a two-step process:
re-instantiate the model object and construct the model.
load the model parameters and load them onto the model.
# 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)
[]
param_not_loadis the list of parameters that were not loaded. Being empty means all parameters were loaded successfully.
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
Predicted: "[1 2 0 4 6 4 9 0 2 2]", Actual: "[1 2 0 4 6 9 9 0 2 2]"