文档反馈
使用MindSpore Golden Stick进行模型转换
有三种方式可以进行模型转换导出MindIR：
训练后导出MindIR；
从ckpt导出MindIR；
训练前配置算法自动导出MindIR。
必要前提
首先下载数据集并构建Lenet网络，同时为了演示方便，我们实现了一个最简单的金箍棒算法，名为FooAlgo。
[ ]:
import os
import numpy as np
from download import download
import mindspore
from mindspore import nn, Model, Tensor, export
from mindspore.train import Accuracy
from mindspore.train import ModelCheckpoint
import mindspore.dataset as ds
import mindspore.dataset.vision as vision
import mindspore.dataset.transforms as transforms
from mindspore.dataset.vision import Inter
from mindspore import dtype as mstype
from mindspore.common.initializer import Normal
from mindspore_gs import CompAlgo
# Download data from open datasets
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/" \
"notebook/datasets/MNIST_Data.zip"
path = download(url, "./", kind="zip", replace=True)
def create_dataset(data_path, batch_size=32, num_parallel_workers=1):
"""
create dataset for train or test
"""
# define dataset
mnist_ds = ds.MnistDataset(data_path)
resize_height, resize_width = 32, 32
rescale = 1.0 / 255.0
rescale_nml = 1 / 0.3081
shift_nml = -1 * 0.1307 / 0.3081
# define map operations
resize_op = vision.Resize((resize_height, resize_width), interpolation=Inter.LINEAR) # Bilinear mode
rescale_nml_op = vision.Rescale(rescale_nml * rescale, shift_nml)
hwc2chw_op = vision.HWC2CHW()
type_cast_op = transforms.TypeCast(mstype.int32)
# apply map operations on images
mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers)
mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers)
mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers)
mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers)
# apply DatasetOps
mnist_ds = mnist_ds.shuffle(buffer_size=1024)
mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True)
return mnist_ds
train_dataset = create_dataset("MNIST_Data/train", 32, 1)
print("train dataset output shape: ", train_dataset.output_shapes())
# initial network
class LeNet5(nn.Cell):
def __init__(self, num_class=10, num_channel=1, include_top=True):
super(LeNet5, self).__init__()
self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid')
self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid')
self.relu = nn.ReLU()
self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
self.include_top = include_top
if self.include_top:
self.flatten = nn.Flatten()
self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02))
self.fc2 = nn.Dense(120, 84, weight_init=Normal(0.02))
self.fc3 = nn.Dense(84, num_class, weight_init=Normal(0.02))
def construct(self, x):
x = self.conv1(x)
x = self.relu(x)
x = self.max_pool2d(x)
x = self.conv2(x)
x = self.relu(x)
x = self.max_pool2d(x)
if not self.include_top:
return x
x = self.flatten(x)
x = self.relu(self.fc1(x))
x = self.relu(self.fc2(x))
x = self.fc3(x)
return x
# set graph mode
mindspore.set_context(mode=mindspore.GRAPH_MODE)
# for demonstration convenience, we implemented one simplest MindSpore Golden Stick algorithm, called FooAlgo
class FooAlgo(CompAlgo):
def apply(self, network: nn.Cell) -> nn.Cell:
return network
print("init ok.")
train dataset output shape: [[32, 1, 32, 32], [32]]
init ok.
训练后导出MindIR
MindSpore Golden Stick各种算法均提供统一的 convert 接口对网络进行模型转换，转换后的网络使用 mindspore.export 接口导出MindIR。
[9]:
## 1) Create network and dataset.
network = LeNet5(10)
train_dataset = create_dataset("MNIST_Data/train", 32, 1)
## 2) Create an algorithm instance.
algo = FooAlgo()
## 3) Apply MindSpore Golden Stick algorithm to origin network.
network_opt = algo.apply(network)
## 4) Set up Model.
net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
net_opt = nn.Momentum(network_opt.trainable_params(), 0.01, 0.9)
model = Model(network_opt, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
cbs = [ModelCheckpoint(prefix='network', directory='ckpt/')]
## 5) Config callback in model.train, start training.
cbs.extend(algo.callbacks())
model.train(1, train_dataset, callbacks=cbs)
## 6) Convert network.
net_deploy = algo.convert(network_opt)
## 7) Export MindIR
inputs = Tensor(np.ones([32, 1, 32, 32]).astype(np.float32)) # user define
export(net_deploy, inputs, file_name="net_1.mindir", file_format="MINDIR")
## 8) Test MindIR
file_path = "./net_1.mindir"
file_path = os.path.realpath(file_path)
if not os.path.exists(file_path):
print("Export MindIR failed!!!")
else:
print("Export MindIR success! MindIR path is: ", file_path)
test_inputs = Tensor(np.ones([32, 1, 32, 32]).astype(np.float32))
graph = mindspore.load(file_path)
net = nn.GraphCell(graph)
output = net(test_inputs)
print("Test output MindIR success, result shape is: ", output.shape)
Export MindIR success! MindIR path is: /home/workspace/golden_stick/net_1.mindir
Test output MindIR success, result shape is: (32, 10)
从ckpt导出
使用训练后得到的ckpt文件，调用 convert 和 mindspore.export 接口导出MindIR。
请先运行上一节示例代码，此小节需用到上节训练生成的ckpt文件。
[ ]:
## 1) Create network and dataset.
network = LeNet5(10)
train_dataset = create_dataset("MNIST_Data/train", 32, 1)
## 2) Create an algorithm instance.
algo = FooAlgo()
## 3) Apply MindSpore Golden Stick algorithm to origin network.
network_opt = algo.apply(network)
## 4) Convert network.
net_deploy = algo.convert(network_opt, ckpt_path="ckpt/network-1_1875.ckpt") # ckpt from previous section
## 5) Export MindIR
inputs = Tensor(np.ones([32, 1, 32, 32]).astype(np.float32)) # user define
export(net_deploy, inputs, file_name="net_2.mindir", file_format="MINDIR")
## 6) Test MindIR
file_path = "./net_2.mindir"
file_path = os.path.realpath(file_path)
if not os.path.exists(file_path):
print("Export MindIR failed!!!")
else:
print("Export MindIR success! MindIR path is: ", file_path)
test_inputs = Tensor(np.ones([32, 1, 32, 32]).astype(np.float32))
graph = mindspore.load(file_path)
net = nn.GraphCell(graph)
output = net(test_inputs)
print("Test output MindIR success, result shape is: ", output.shape)
Export MindIR success! MindIR path is: /home/workspace/golden_stick/net_2.mindir
Test output MindIR success, result shape is: (32, 10)
配置算法自动导出MindIR
在训练前配置算法set_save_mindir 接口，在训练后自动生成MindIR。
此种方式生成的MindIR在推理时，模型的输入shape需与训练时输入的数据集shape保持一致。
配置算法自动导出MindIR有两个必要的操作，
set_save_mindir(True)及在model.train中配回调函数时加入算法回调函数callbacks=algo.callbacks()。MindIR输出路径save_mindir_path若未配置则默认保存为./network.mindir。
[ ]:
## 1) Create network and dataset.
network = LeNet5(10)
train_dataset = create_dataset("MNIST_Data/train", 32, 1)
## 2) Create an algorithm instance.
algo = FooAlgo()
## 3) Enable automatically export MindIR after training.
algo.set_save_mindir(save_mindir=True)
## 4) Set MindIR output path, the default value for the path is 'network.mindir'.
algo.set_save_mindir_path(save_mindir_path="net_3.mindir")
## 5) Apply MindSpore Golden Stick algorithm to origin network.
network_opt = algo.apply(network)
## 6) Set up Model.
net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
net_opt = nn.Momentum(network_opt.trainable_params(), 0.01, 0.9)
model = Model(network_opt, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
## 7) Config callback in model.train, start training, then MindIR will be exported.
model.train(1, train_dataset, callbacks=algo.callbacks())
## 8) Test MindIR
file_path = "./net_3.mindir"
file_path = os.path.realpath(file_path)
if not os.path.exists(file_path):
print("Export MindIR failed!!!")
else:
print("Export MindIR success! MindIR path is: ", file_path)
test_inputs = Tensor(np.ones([32, 1, 32, 32]).astype(np.float32))
graph = mindspore.load(file_path)
net = nn.GraphCell(graph)
output = net(test_inputs)
print("Test output MindIR success, result shape is: ", output.shape)
Export MindIR success! MindIR path is: /home/workspace/golden_stick/net_3.mindir
Test output MindIR success, result shape is: (32, 10)
