1. MindIR overview
MindSpore defines the logical structure of the network and the attributes of operators through unified IR, decouples the model file in MindIR format
from the hardware platform, and realizes multiple deployments for one training.
MindIR, as the unified model file of MindSpore, also stores the network structure and weight parameter values. At the same time, it supports
deployment to cloud Serving and end-side Lite platforms to perform inference tasks.
The same MindIR file supports the deployment of multiple hardware forms:-Serving
deployment reasoning
-End-side Lite reasoning deployment
Second, noun explanation
Checkpoint
• The Protocol Buffers format is used to store all parameter values in the network.
• Generally used to resume training after a training task is interrupted, or to fine-tune (Fine Tune) tasks after training.
•MindIR
• The full name is MindSpore IR, which is a functional IR of MindSpore based on graph representation. It defines a scalable graph
structure and IR representation of operators.
• It eliminates model differences between different backends, and is generally used to perform inference tasks across hardware platforms.
•ONNX
• The full name is Open Neural Network Exchange, which is a general expression for machine learning models.
• Generally used for model migration between different frameworks or used on the inference engine (TensorRT).
•AIR
• The full name is Ascend Intermediate Representation, which is an open
file format designed for machine learning defined by Huawei .
• It can better adapt to the Huawei AI processor and is generally used to perform inference tasks on the Ascend 310.
3. Export the MindIR format model of the LeNet network
1. Train LeNet model to generate checkpoint
LeNet training code uses the code in modelzool in the official MindSpore code warehouse
Address: https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/lenet
2. MNIST data download address
1) Official download address: http://yann.lecun.com/exdb/mnist/
2) Download from SkyDrive: Link: https://pan.baidu.com/s/1zX-OwL8bOgq4dhEuaRj2Xg Extraction code: zew6
After the MNIST data set is downloaded, unzip it to the code root directory
3. Execute training commands
python train.py --data_path ./MNIST_DATA/ --ckpt_path=./checkpoint/ --device_target AscendNote: I use the Shengteng platform for model training. If you use CPU or GPU for training, then the device_target will be changed to the corresponding one.
When epoch, loss and other values appear on the screen, the model starts training
4. Get the checkpoint
When the ckpt file appears in the ckpt_path folder specified in the model training parameters, the training is successful. Next, we can choose a ckpt file to convert to the mindir format. Here I choose checkpoint_lenet-10_1875.ckpt
5. Write model conversion code
import numpy as np
from mindspore import Tensor, export, load_checkpoint, load_param_into_net
from src.lenet import LeNet5
lenet = LeNet5()
# load the parameter into net
load_checkpoint("./checkpoint/checkpoint_lenet-10_1875.ckpt", net=lenet) #checkpoint_lenet-10_1875.ckpt更换成对应所需要转换的ckpt文件
input = np.random.uniform(0.0, 1.0, size=[32, 1, 32, 32]).astype(np.float32) #Lenet模型的size为32,1,32,32
export(lenet, Tensor(input), file_name='lenet-10_1875', file_format='MINDIR') #file_name指定转换后文件的文件名6. Execute model conversion code
python lenet_mindr.pyAfter execution, check the root directory of the code. If there is the lenet-10_1875.mindir file we specified earlier, it means the model conversion is successful!
Fourth, export the MindIR format model of the ResNet50 network
1. Train ResNet50 network to generate checkpoint
The ResNet50 model training code still uses the modelzoo code in the official MindSpore code warehouse, link: https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/resnet
2.Cifar10 data set
Because the ImageNet data set is relatively large, it takes a long time to train, so here we use the cifar10 data set for model training
1) Official download of Cifar10 dataset: http://www.cs.toronto.edu/~kriz/cifar.html
2) Baidu network disk download: link: https://pan.baidu.com/s/1CpgjFtZk2ZQsr_qUtc6z1g Extraction code: kjhi
3. Execute training commands
python train.py --net resnet50 --dataset cifar10 --dataset_path ./data/cifar10-bin/train/ --device_target AscendNote: I use the Shengteng platform for model training. If you use CPU or GPU for training, then the device_target will be changed to the corresponding one.
When epoch, loss and other values appear on the screen, the model starts training
4. Get checkpoint
When the ckpt file appears in the ckpt_path folder specified in the model training parameters, the training is successful. Next, we can choose a ckpt file to convert to the mindir format. Here I choose resnet-90_1562.ckpt
5. Write model conversion scripts
import numpy as np
from mindspore import Tensor, export, load_checkpoint, load_param_into_net
from src.resnet import resnet50 as ResNet50
resnet = ResNet50()
# load the parameter into net
load_checkpoint("./checkpoint/resnet-90_1562.ckpt", net=resnet)
input = np.random.uniform(0.0, 1.0, size=[32, 3, 224, 224]).astype(np.float32)
export(resnet, Tensor(input), file_name='resnet-90_162', file_format='MINDIR')6. Perform model conversion
python resnet_mindir.pyAfter execution, check the code root directory. If there is the resnet-90_1562.mindir file we specified earlier, it means that the model conversion is successful!