Ascend910与GPU推理

Ascend GPU 进阶 推理应用

image0image1image2image3

本文将介绍如何在Ascend910和GPU硬件环境中,利用MindIR和Checkpoint执行推理。MindIR是MindSpore的统一模型文件,同时存储了网络结构和权重参数值,定义了可扩展的图结构以及算子的IR表示,消除了不同后端的模型差异,一般用于跨硬件平台执行推理任务。Checkpoint是训练参数,采用了Protocol Buffers格式,一般用于训练任务中断后恢复训练,或训练后的微调(Fine Tune)任务。

下面将针对这两种情况,介绍如何使用MindSpore进行单卡推理。

使用checkpoint格式文件单卡推理

使用本地模型推理

用户可以通过load_checkpointload_param_into_net接口从本地加载模型与参数,传入验证数据集后使用model.eval即可进行模型验证,使用model.predict可进行模型推理。在这里我们下载MindSpore Hub中已经预训练好的LeNet和MINIST数据集进行推理演示:

[ ]:
!mkdir -p ./datasets/MNIST_Data/test
!wget -NP ./datasets/MNIST_Data/test https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/mnist/t10k-labels-idx1-ubyte
!wget -NP ./datasets/MNIST_Data/test https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/mnist/t10k-images-idx3-ubyte
[ ]:
!mkdir -p ./checkpoint
!wget -NP ./checkpoint https://download.mindspore.cn/model_zoo/r1.1/lenet_ascend_v111_offical_cv_mnist_bs32_acc98/lenet_ascend_v111_offical_cv_mnist_bs32_acc98.ckpt

配置运行所需信息,进行推理的数据处理:

如果在Ascend910环境中运行,下述配置中device_target=“GPU”的GPU需改为Ascend

[4]:
import os
import argparse

from mindspore import context
import mindspore.dataset as ds
import mindspore.dataset.transforms.c_transforms as C
import mindspore.dataset.vision.c_transforms as CV
from mindspore.dataset.vision import Inter
from mindspore import dtype as mstype

context.set_context(mode=context.GRAPH_MODE, device_target="GPU")

def create_dataset(data_path, batch_size=32, repeat_size=1,
                   num_parallel_workers=1):
    # 定义数据集
    mnist_ds = ds.MnistDataset(data_path)
    resize_height, resize_width = 32, 32
    rescale = 1.0 / 255.0
    shift = 0.0
    rescale_nml = 1 / 0.3081
    shift_nml = -1 * 0.1307 / 0.3081

    # 定义所需要操作的map映射
    resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR)
    rescale_nml_op = CV.Rescale(rescale_nml, shift_nml)
    rescale_op = CV.Rescale(rescale, shift)
    hwc2chw_op = CV.HWC2CHW()
    type_cast_op = C.TypeCast(mstype.int32)

    # 使用map映射函数,将数据操作应用到数据集
    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_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)

    # 进行shuffle、batch操作
    buffer_size = 10000
    mnist_ds = mnist_ds.shuffle(buffer_size=buffer_size)
    mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True)

    return mnist_ds

创建LeNet模型:

[5]:
import mindspore.nn as nn
from mindspore.common.initializer import Normal

class LeNet5(nn.Cell):
    """
    Lenet网络结构
    """
    def __init__(self, num_class=10, num_channel=1):
        super(LeNet5, self).__init__()
        # 定义所需要的运算
        self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid')
        self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid')
        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))
        self.relu = nn.ReLU()
        self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
        self.flatten = nn.Flatten()

    def construct(self, x):
        # 使用定义好的运算构建前向网络
        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)
        x = self.flatten(x)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.fc3(x)
        return x

# 实例化网络
net = LeNet5()

在推理进行前,需要使用load_checkpointload_param_into_net接口从本地加载模型与参数。这样一来就可以使用本地模型完成后面的推理过程。

[6]:
from mindspore import load_checkpoint, load_param_into_net
ckpt_file_name = "./checkpoint/lenet_ascend_v111_offical_cv_mnist_bs32_acc98.ckpt"
param_dict = load_checkpoint(ckpt_file_name)
load_param_into_net(net, param_dict)

设置损失函数与优化器,并调用model接口创建对象:

[7]:
import numpy as np
from mindspore.nn import Accuracy
from mindspore import Model, Tensor

net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
net_opt = nn.Momentum(net.trainable_params(), learning_rate=0.01, momentum=0.9)
model = Model(net, net_loss, net_opt, metrics={"Accuracy": Accuracy()})

下面调用model.eval接口执行验证过程:

[8]:
mnist_path = "./datasets/MNIST_Data/test"
ds_eval = create_dataset(mnist_path)
acc = model.eval(ds_eval, dataset_sink_mode=False)
print("============== {} ==============".format(acc))
============== {'Accuracy': 0.9846754807692307} ==============

推理完整样例代码参见https://gitee.com/mindspore/models/blob/r1.5/official/cv/lenet/eval.py

调用model.predict接口执行验证过程,这里选取数据集中的一张图片进行预测:

被预测的图片数据为随机抽取,实际执行结果可能与文本示例不一致,预测分类与实际分类一致即表示预测结果正确。

[9]:
ds_eval = ds_eval.create_dict_iterator()
data = next(ds_eval)

# images为测试图片,labels为测试图片的实际分类
images = data["image"].asnumpy()
labels = data["label"].asnumpy()

# 使用函数model.predict预测image对应分类
output = model.predict(Tensor(data['image']))
predicted = np.argmax(output.asnumpy(), axis=1)

# 输出预测分类与实际分类
print(f'Predicted: "{predicted[0]}", Actual: "{labels[0]}"')
Predicted: "6", Actual: "6"

加载MindSpore Hub模型执行推理

除了使用load_checkpointload_param_into_net从本地加载模型之外,也可以通过安装MindSpore Hub,通过mindspore_hub.load从云端加载模型参数执行推理。

之前使用的加载本地模型的方法为:

from mindspore import load_checkpoint, load_param_into_net
ckpt_file_name = "./checkpoint/lenet_ascend_v111_offical_cv_mnist_bs32_acc98.ckpt"
param_dict = load_checkpoint(ckpt_file_name)
load_param_into_net(net, param_dict)

可替换为mindspore_hub.load方法:

import mindspore_hub
model_uid = "mindspore/ascend/1.2/lenet_v1.2_mnist"
net = mindspore_hub.load(model_uid)

在Ascend中使用C++接口推理MindIR格式文件

本小节将介绍如何使用C++接口推理MINDIR格式的模型。完整代码可参考ascend910_resnet50_preprocess_sample

本小节内容及代码仅适用于Ascend环境。

推理代码介绍

完成的推理代码为main.cc文件,现在对其中的功能实现进行说明。

引用mindsporemindspore::dataset的名字空间。

namespace ms = mindspore;
namespace ds = mindspore::dataset;

环境初始化,指定硬件为Ascend 910,DeviceID为0:

auto context = std::make_shared<ms::Context>();
auto ascend910_info = std::make_shared<ms::Ascend910DeviceInfo>();
ascend910_info->SetDeviceID(0);
context->MutableDeviceInfo().push_back(ascend910_info);

加载模型文件:

// 加载 MindIR 模型
ms::Graph graph;
ms::Status ret = ms::Serialization::Load(resnet_file, ms::ModelType::kMindIR, &graph);
// 进行图编译
ms::Model resnet50;
ret = resnet50.Build(ms::GraphCell(graph), context);

获取模型所需输入信息:

std::vector<ms::MSTensor> model_inputs = resnet50.GetInputs();

加载图片文件:

ms::MSTensor ReadFile(const std::string &file);
auto image = ReadFile(image_file);

图片预处理:

// 对图片进行解码,变为RGB格式,并重设尺寸
std::shared_ptr<ds::TensorTransform> decode(new ds::vision::Decode());
std::shared_ptr<ds::TensorTransform> resize(new ds::vision::Resize({256}));
// 输入归一化
std::shared_ptr<ds::TensorTransform> normalize(new ds::vision::Normalize(
    {0.485 * 255, 0.456 * 255, 0.406 * 255}, {0.229 * 255, 0.224 * 255, 0.225 * 255}));
// 剪裁图片
std::shared_ptr<ds::TensorTransform> center_crop(new ds::vision::CenterCrop({224, 224}));
// shape (H, W, C) 变为 shape (C, H, W)
std::shared_ptr<ds::TensorTransform> hwc2chw(new ds::vision::HWC2CHW());

// 定义preprocessor
ds::Execute preprocessor({decode, resize, normalize, center_crop, hwc2chw});

// 调用函数,获取处理后的图像
ret = preprocessor(image, &image);

执行推理:

// 创建输入输出向量
std::vector<ms::MSTensor> outputs;
std::vector<ms::MSTensor> inputs;

inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
                    image.Data().get(), image.DataSize());
// 执行推理
ret = resnet50.Predict(inputs, &outputs);

获取推理结果:

// 获取推理结果的最大可能性
std::cout << "Image: " << image_file << " infer result: " << GetMax(outputs[0]) << std::endl;

构建脚本介绍

构建脚本用于构建用户程序,完整代码位于CMakeLists.txt ,下面进行解释说明。

为编译器添加头文件搜索路径:

option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)

在MindSpore中查找所需动态库:

find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)

使用指定的源文件生成目标可执行文件,并为目标文件链接MindSpore库:

add_executable(resnet50_sample main.cc)
target_link_libraries(resnet50_sample ${MS_LIB} ${MD_LIB})

编译并执行推理代码

可选择将实验的脚本下载至Ascend910环境中编译并执行。

进入工程目录ascend910_resnet50_preprocess_sample,执行cmake命令,其中pip3需要按照实际情况修改::

cmake . -DMINDSPORE_PATH=`pip3 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`

再执行make命令编译即可。

make

编译成功后,会获得resnet50_sample可执行文件。在工程目录ascend910_resnet50_preprocess_sample下创建model目录放置MindIR文件resnet50_imagenet.mindir。此外,创建test_data目录用于存放待分类的图片,图片可来自ImageNet等各类开源数据集,输入执行命令即可获取推理结果:

./resnet50_sample