{
"cells": [
{
"cell_type": "markdown",
"id": "69a92ef2",
"metadata": {},
"source": [
"[](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r2.4.1/tutorials/zh_cn/beginner/mindspore_accelerate_with_static_graph.ipynb) [](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r2.4.1/tutorials/zh_cn/beginner/mindspore_accelerate_with_static_graph.py) [](https://gitee.com/mindspore/docs/blob/r2.4.1/tutorials/source_zh_cn/beginner/accelerate_with_static_graph.ipynb)\n",
"\n",
"[基本介绍](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/introduction.html) || [快速入门](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/quick_start.html) || [张量 Tensor](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/tensor.html) || [数据加载与处理](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/dataset.html) || [网络构建](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/model.html) || [函数式自动微分](https://gitee.com/mindspore/docs/blob/r2.4.1/tutorials/source_zh_cn/beginner/autograd.ipynb) || [模型训练](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/train.html) || [保存与加载](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/save_load.html) || **使用静态图加速** || [自动混合精度](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/mixed_precision.html) ||\n",
"\n",
"# 使用静态图加速\n",
"\n",
"## 背景介绍\n",
"\n",
"AI编译框架分为两种运行模式，分别是动态图模式以及静态图模式。MindSpore默认情况下是以动态图模式运行，但也支持手工切换为静态图模式。两种运行模式的详细介绍如下：\n",
"\n",
"### 动态图模式\n",
"\n",
"动态图的特点是计算图的构建和计算同时发生（Define by run），其符合Python的解释执行方式，在计算图中定义一个Tensor时，其值就已经被计算且确定，因此在调试模型时较为方便，能够实时得到中间结果的值，但由于所有节点都需要被保存，导致难以对整个计算图进行优化。\n",
"\n",
"在MindSpore中，动态图模式又被称为PyNative模式。由于动态图的解释执行特性，在脚本开发和网络流程调试过程中，推荐使用动态图模式进行调试。\n",
"如需要手动控制框架采用PyNative模式，可以通过以下代码进行网络构建："
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "6dc8d4c5",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" ...\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]\n",
" [-0.00134926 -0.13563682 -0.02863023 -0.05452826 0.03290743 -0.12423715\n",
" -0.0582641 -0.10854103 -0.08558805 0.06099342]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"ms.set_context(mode=ms.PYNATIVE_MODE) # 使用set_context进行动态图模式的配置\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"model = Network()\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"output = model(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "6a9a195c",
"metadata": {},
"source": [
"### 静态图模式\n",
"\n",
"相较于动态图而言，静态图的特点是将计算图的构建和实际计算分开（Define and run）。有关静态图模式的运行原理，可以参考[静态图语法支持](https://www.mindspore.cn/docs/zh-CN/r2.4.1/model_train/program_form/static_graph.html#概述)。\n",
"\n",
"在MindSpore中，静态图模式又被称为Graph模式，在Graph模式下，基于图优化、计算图整图下沉等技术，编译器可以针对图进行全局的优化，获得较好的性能，因此比较适合网络固定且需要高性能的场景。\n",
"\n",
"如需要手动控制框架采用静态图模式，可以通过以下代码进行网络构建："
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "2476e059",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" ...\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]\n",
" [ 0.05363735 0.05117104 -0.03343301 0.06347139 0.07546629 0.03263091\n",
" 0.02790363 0.06269836 0.01838502 0.04387159]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"ms.set_context(mode=ms.GRAPH_MODE) # 使用set_context进行运行静态图模式的配置\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"model = Network()\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"output = model(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "89825897",
"metadata": {},
"source": [
"## 静态图模式的使用场景\n",
"\n",
"MindSpore编译器重点面向Tensor数据的计算以及其微分处理。因此使用MindSpore API以及基于Tensor对象的操作更适合使用静态图编译优化。其他操作虽然可以部分入图编译，但实际优化作用有限。另外，静态图模式先编译后执行的模式导致其存在编译耗时。因此，如果函数无需反复执行，那么使用静态图加速也可能没有价值。\n",
"\n",
"有关使用静态图来进行网络编译的示例，请参考[网络构建](https://www.mindspore.cn/tutorials/zh-CN/r2.4.1/beginner/model.html)。\n",
"\n",
"## 静态图模式开启方式\n",
"\n",
"通常情况下，由于动态图的灵活性，我们会选择使用PyNative模式来进行自由的神经网络构建，以实现模型的创新和优化。但是当需要进行性能加速时，我们需要对神经网络部分或整体进行加速。MindSpore提供了两种切换为图模式的方式，分别是基于装饰器的开启方式以及基于全局context的开启方式。\n",
"\n",
"### 基于装饰器的开启方式\n",
"\n",
"MindSpore提供了jit装饰器，可以通过修饰Python函数或者Python类的成员函数使其被编译成计算图，通过图优化等技术提高运行速度。此时我们可以简单的对想要进行性能优化的模块进行图编译加速，而模型其他部分，仍旧使用解释执行方式，不丢失动态图的灵活性。无论全局context是设置成静态图模式还是动态图模式，被jit修饰的部分始终会以静态图模式进行运行。\n",
"\n",
"在需要对Tensor的某些运算进行编译加速时，可以在其定义的函数上使用jit修饰器，在调用该函数时，该模块自动被编译为静态图。需要注意的是，jit装饰器只能用来修饰函数，无法对类进行修饰。jit的使用示例如下："
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "d97a2f69",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" ...\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]\n",
" [-0.12126954 0.06986676 -0.2230821 -0.07087803 -0.01003947 0.01063392\n",
" 0.10143848 -0.0200909 -0.09724037 0.0114444 ]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"\n",
"@ms.jit # 使用ms.jit装饰器，使被装饰的函数以静态图模式运行\n",
"def run(x):\n",
" model = Network()\n",
" return model(x)\n",
"\n",
"output = run(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "0c5adf4d",
"metadata": {},
"source": [
"除使用修饰器外，也可使用函数变换方式调用jit方法，示例如下："
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "7337a0bb",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" ...\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]\n",
" [ 0.11027216 -0.09628229 0.0457969 0.05396656 -0.06958974 0.0428197\n",
" -0.1572069 -0.14151613 -0.04531277 0.07521383]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"\n",
"def run(x):\n",
" model = Network()\n",
" return model(x)\n",
"\n",
"run_with_jit = ms.jit(run) # 通过调用jit将函数转换为以静态图方式执行\n",
"output = run_with_jit(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "df543f53",
"metadata": {},
"source": [
"当我们需要对神经网络的某部分进行加速时，可以直接在construct方法上使用jit修饰器，在调用实例化对象时，该模块自动被编译为静态图。示例如下："
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "e51b32f5",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" ...\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]\n",
" [ 0.10522258 0.06597593 -0.09440921 -0.04883489 0.07194916 0.1343117\n",
" -0.06813788 0.01986085 0.0216996 -0.05345828]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" @ms.jit # 使用ms.jit装饰器，使被装饰的函数以静态图模式运行\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"model = Network()\n",
"output = model(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "d8aef1df",
"metadata": {},
"source": [
"### 基于context的开启方式\n",
"\n",
"context模式是一种全局的设置模式。代码示例如下："
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "bb1d7b61",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[[ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" ...\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]\n",
" [ 0.08501796 -0.04404321 -0.05165704 0.00357929 0.00051521 0.00946456\n",
" 0.02748473 -0.19415936 -0.00278988 0.04024826]]\n"
]
}
],
"source": [
"import numpy as np\n",
"import mindspore as ms\n",
"from mindspore import nn, Tensor\n",
"ms.set_context(mode=ms.GRAPH_MODE) # 使用set_context进行运行静态图模式的配置\n",
"\n",
"class Network(nn.Cell):\n",
" def __init__(self):\n",
" super().__init__()\n",
" self.flatten = nn.Flatten()\n",
" self.dense_relu_sequential = nn.SequentialCell(\n",
" nn.Dense(28*28, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 512),\n",
" nn.ReLU(),\n",
" nn.Dense(512, 10)\n",
" )\n",
"\n",
" def construct(self, x):\n",
" x = self.flatten(x)\n",
" logits = self.dense_relu_sequential(x)\n",
" return logits\n",
"\n",
"model = Network()\n",
"input = Tensor(np.ones([64, 1, 28, 28]).astype(np.float32))\n",
"output = model(input)\n",
"print(output)"
]
},
{
"cell_type": "markdown",
"id": "b863f0f7",
"metadata": {},
"source": [
"## 静态图的语法约束\n",
"\n",
"在Graph模式下，Python代码并不是由Python解释器去执行，而是将代码编译成静态计算图，然后执行静态计算图。因此，编译器无法支持全量的Python语法。MindSpore的静态图编译器维护了Python常用语法子集，以支持神经网络的构建及训练。详情可参考[静态图语法支持](https://www.mindspore.cn/docs/zh-CN/r2.4.1/model_train/program_form/static_graph.html)。\n",
"\n",
"## 静态图高级编程技巧\n",
"\n",
"使用静态图高级编程技巧可以有效地提高编译效率以及执行效率，并可以使程序运行的更加稳定。详情可参考[静态图高级编程技巧](https://www.mindspore.cn/docs/zh-CN/r2.4.1/model_train/program_form/static_graph_syntax/static_graph_expert_programming.html)。"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "MindSpore",
"language": "python",
"name": "mindspore"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.10"
}
},
"nbformat": 4,
"nbformat_minor": 5
}