{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 动静态图结合\n", "\n", "[![下载Notebook](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_notebook.png)](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r1.7/zh_cn/design/mindspore_dynamic_graph_and_static_graph.ipynb) [![下载样例代码](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_download_code.png)](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r1.7/zh_cn/design/mindspore_dynamic_graph_and_static_graph.py) [![查看源文件](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/resource/_static/logo_source.png)](https://gitee.com/mindspore/docs/blob/r1.7/docs/mindspore/source_zh_cn/design/dynamic_graph_and_static_graph.ipynb)\n", "\n", "## 静态图和动态图的概念\n", "\n", "目前主流的深度学习框架的执行模式有两种,分别为静态图模式和动态图模式。\n", "\n", "静态图模式下,程序在编译执行时先生成神经网络的图结构,然后再执行图中涉及的计算操作。因此,在静态图模式下,编译器利用图优化等技术对执行图进行更大程度的优化,从而获得更好的执行性能,有助于规模部署和跨平台运行。\n", "\n", "动态图模式下,程序按照代码的编写顺序执行,在执行正向过程中根据反向传播的原理,动态生成反向执行图。这种模式下,编译器将神经网络中的各个算子逐一下发执行,方便用户编写和调试神经网络模型。\n", "\n", "## MindSpore静态图\n", "\n", "在MindSpore中,静态图模式又被称为Graph模式,可以通过`context.set_context(mode=context.GRAPH_MODE)`来设置成静态图模式。静态图模式比较适合网络固定且需要高性能的场景。在静态图模式下,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,因此在静态图下能获得较好的性能,但是执行图是从源码转换而来,因此在静态图下不是所有的Python语法都能支持,详细请查看[语法支持](https://www.mindspore.cn/docs/zh-CN/r1.7/note/syntax_list.html)。\n", "\n", "### Graph模式执行原理\n", "\n", "在Graph模式下,MindSpore通过源码转换的方式,将Python的源码转换成IR,再在此基础上进行相关的图优化,最终在硬件设备上执行优化后的图。MindSpore使用的是一种基于图表示的函数式IR,即MindIR,采用了接近于ANF函数式的语义。Graph模式是基于MindIR进行编译优化的,使用Graph模式时,需要使用`nn.Cell`类并且在`construct`函数中编写执行代码, 或者调用`@ms_function`装饰器。\n", "\n", "Graph模式的代码用例如下所示:" ] }, { "cell_type": "code", "execution_count": 1, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.282871Z", "start_time": "2022-01-04T10:51:21.743620Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[ 4. 10. 18.]\n" ] } ], "source": [ "import numpy as np\n", "import mindspore.nn as nn\n", "import mindspore.ops as ops\n", "from mindspore import context, Tensor\n", "\n", "context.set_context(mode=context.GRAPH_MODE, device_target=\"CPU\")\n", "\n", "class Net(nn.Cell):\n", " def __init__(self):\n", " super(Net, self).__init__()\n", " self.mul = ops.Mul()\n", "\n", " def construct(self, x, y):\n", " return self.mul(x, y)\n", "\n", "x = Tensor(np.array([1.0, 2.0, 3.0]).astype(np.float32))\n", "y = Tensor(np.array([4.0, 5.0, 6.0]).astype(np.float32))\n", "\n", "net = Net()\n", "print(net(x, y))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Graph模式自动微分原理\n", "\n", "在MindSpore中,Graph模式下的自动微分原理可以参考[自动微分](https://www.mindspore.cn/tutorials/zh-CN/r1.7/beginner/autograd.html)。\n", "\n", "## MindSpore动态图\n", "\n", "在MindSpore中,动态图模式又被称为PyNative模式,可以通过`context.set_context(mode=context.PYNATIVE_MODE)`来设置成动态图模式。在脚本开发和网络流程调试中,推荐使用动态图模式进行调试,支持执行单算子、普通函数和网络、以及单独求梯度的操作。\n", "\n", "### PyNative模式执行原理\n", "\n", "在PyNative模式下,用户可以使用完整的Python API,此外针对使用MindSpore提供的API时,框架会根据用户选择的硬件平台(Ascend,GPU,CPU),将算子API的操作在对应的硬件平台上执行,并返回相应的结果。框架整体的执行过程如下:\n", "\n", "![process](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/docs/mindspore/source_zh_cn/design/images/framework.png)\n", "\n", "通过前端的Python API,调用到框架层,最终到相应的硬件设备上进行计算。例如:完成一个加法" ] }, { "cell_type": "code", "execution_count": 2, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.292278Z", "start_time": "2022-01-04T10:51:23.284465Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[[[[2. 2. 2. 2.]\n", " [2. 2. 2. 2.]\n", " [2. 2. 2. 2.]]\n", "\n", " [[2. 2. 2. 2.]\n", " [2. 2. 2. 2.]\n", " [2. 2. 2. 2.]]\n", "\n", " [[2. 2. 2. 2.]\n", " [2. 2. 2. 2.]\n", " [2. 2. 2. 2.]]]]\n" ] } ], "source": [ "import numpy as np\n", "import mindspore.context as context\n", "import mindspore.nn as nn\n", "from mindspore import Tensor\n", "import mindspore.ops as ops\n", "\n", "context.set_context(mode=context.PYNATIVE_MODE, device_target=\"CPU\")\n", "x = Tensor(np.ones([1, 3, 3, 4]).astype(np.float32))\n", "y = Tensor(np.ones([1, 3, 3, 4]).astype(np.float32))\n", "output = ops.add(x, y)\n", "print(output.asnumpy())" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "此例中,当调用到Python接口ops.add(x, y)时,会将Python的接口调用通过Pybind11调用到框架的C++层,转换成C++的调用,接着框架会根据用户设置的device_target选择对应的硬件设备,在该硬件设备上执行add这个操作。\n", "\n", "从上述原理可以看到,在PyNative模式下,Python脚本代码会根据Python的语法进行执行,而执行过程中涉及到MindSpore的API,会根据用户设置在不同的硬件上进行执行,从而进行加速。因此,在PyNative模式下,用户可以随意使用Python的语法以及调试方法。例如可以使用常见的PyCharm、VS Code等IDE进行代码的调试。\n", "\n", "### PyNative模式自动微分原理\n", "\n", "在前面的介绍中,我们可以看出,在PyNative下执行正向过程完全是按照Python的语法进行执行。在PyNative下是基于Tensor进行实现反向传播的,我们在执行正向过程中,将所有应用于Tensor的操作记录下来,并针对每个操作求取其反向,并将所有反向过程串联起来形成整体反向传播图(简称反向图)。最终,将反向图在设备上进行执行计算出梯度。\n", "\n", "反向构图过程示例,如下代码,对矩阵x乘上固定参数z,然后与y进行矩阵乘法,最终对x进行求导。" ] }, { "cell_type": "code", "execution_count": 3, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.439867Z", "start_time": "2022-01-04T10:51:23.293334Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[[9.02 5.4 7.2000003]\n", " [9.02 5.4 7.2000003]]\n" ] } ], "source": [ "import numpy as np\n", "import mindspore.nn as nn\n", "import mindspore.ops as ops\n", "from mindspore import Tensor, context\n", "from mindspore import Parameter\n", "from mindspore import dtype as mstype\n", "\n", "context.set_context(mode=context.PYNATIVE_MODE, device_target=\"CPU\")\n", "\n", "class Net(nn.Cell):\n", " def __init__(self):\n", " super(Net, self).__init__()\n", " self.matmul = ops.MatMul()\n", " self.z = Parameter(Tensor(np.array([2.0], np.float32)), name='z')\n", "\n", " def construct(self, x, y):\n", " x = x * self.z\n", " out = self.matmul(x, y)\n", " return out\n", "\n", "class GradNetWrtX(nn.Cell):\n", " def __init__(self, net):\n", " super(GradNetWrtX, self).__init__()\n", " self.net = net\n", " self.grad_op = ops.GradOperation()\n", "\n", " def construct(self, x, y):\n", " gradient_function = self.grad_op(self.net)\n", " return gradient_function(x, y)\n", "\n", "x = Tensor([[0.8, 0.6, 0.2], [1.8, 1.3, 1.1]], dtype=mstype.float32)\n", "y = Tensor([[0.11, 3.3, 1.1], [1.1, 0.2, 1.4], [1.1, 2.2, 0.3]], dtype=mstype.float32)\n", "output = GradNetWrtX(Net())(x, y)\n", "print(output)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "![forward](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/docs/mindspore/source_zh_cn/design/images/forward.png) ![backward](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/docs/mindspore/source_zh_cn/design/images/backward.png)\n", "\n", "根据上述PyNative下构图原理,我们可以看到,在正向传播过程中,我们记录了Mul的计算过程,根据Mul对应的反向bprop的定义,得到了反向的MulGrad算子,根据Mul算子的bprop定义,如下:" ] }, { "cell_type": "code", "execution_count": 4, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.445921Z", "start_time": "2022-01-04T10:51:23.441876Z" } }, "outputs": [], "source": [ "from mindspore.ops._grad.grad_base import bprop_getters\n", "\n", "@bprop_getters.register(ops.Mul)\n", "def get_bprop_mul(self):\n", " \"\"\"Grad definition for `Mul` operation.\"\"\"\n", " mul_func = P.Mul()\n", "\n", " def bprop(x, y, out, dout):\n", " bc_dx = mul_func(y, dout)\n", " bc_dy = mul_func(x, dout)\n", " return binop_grad_common(x, y, bc_dx, bc_dy)\n", "\n", " return bprop" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "可以看到对Mul的输入求反向,需要两个输入和输出的反向传播梯度值,此时根据实际的输入值,可以将z连接到MulGrad。以此类推,对下一个算子Matmul,相应的得到MatmulGrad信息,再根据bprop的输入输出,将上下文梯度传播连接起来。\n", "\n", "最终,对于初始传播,在MindSpore中使用[sens](https://mindspore.cn/docs/zh-CN/r1.7/api_python/ops/mindspore.ops.GradOperation.html#mindspore.ops.GradOperation)进行缩放,默认值为1。同理对于输入y求导,可以使用同样的过程进行推导。\n", "\n", "### PyNative模式下的控制流\n", "\n", "在PyNative模式下,脚本按照Python的语法执行,因此在MindSpore中,针对控制流语法并没有做特殊处理,直接按照Python的语法直接展开执行,进而对展开的执行算子进行自动微分的操作。例如,对于for循环,在PyNative下会根据具体的循环次数,不断的执行for循环中的语句,并对其算子进行自动微分的操作。\n", "\n", "## 动静统一\n", "\n", "### 概述\n", "\n", "当前在业界支持动态图和静态图两种模式,动态图通过解释执行,具有动态语法亲和性,表达灵活;静态图使用jit编译优化执行,偏静态语法,在语法上有较多限制。动态图和静态图的编译流程不一致,语法约束不一致。MindSpore针对动态图和静态图模式,首先统一API表达,在两种模式下使用相同的API;其次统一动态图和静态图的底层微分机制。\n", "\n", "![dynamic](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.7/docs/mindspore/source_zh_cn/design/images/dynamic.png)\n", "\n", "### 动态图和静态图互相转换\n", "\n", "在MindSpore中,我们可以通过控制模式输入参数来切换执行使用动态图还是静态图。例如:" ] }, { "cell_type": "code", "execution_count": 5, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.461198Z", "start_time": "2022-01-04T10:51:23.447508Z" } }, "outputs": [], "source": [ "context.set_context(mode=context.PYNATIVE_MODE)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "由于在静态图下,对于Python语法有所限制,因此从动态图切换成静态图时,需要符合静态图的语法限制,才能正确使用静态图来进行执行。更多静态图的语法限制可以参考[静态图语法限制](https://www.mindspore.cn/docs/zh-CN/r1.7/note/static_graph_syntax_support.html)。\n", "\n", "### 动静结合\n", "\n", "MindSpore支持在动态图下使用静态编译的方式来进行混合执行,通过使用ms_function修饰需要用静态图来执行的函数对象,即可实现动态图和静态图的混合执行,更多ms_function的使用可参考[ms_function文档](https://www.mindspore.cn/tutorials/zh-CN/r1.7/advanced/pynative_graph/combine.html#ms-function装饰器)。\n", "\n", "例如:" ] }, { "cell_type": "code", "execution_count": 6, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:23.514919Z", "start_time": "2022-01-04T10:51:23.462207Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[[[[15.99984]]\n", "\n", " [[15.99984]]]]\n" ] } ], "source": [ "import numpy as np\n", "import mindspore as ms\n", "import mindspore.nn as nn\n", "from mindspore import Tensor\n", "from mindspore import ms_function\n", "from mindspore import Parameter\n", "import mindspore.context as context\n", "\n", "class AddMulMul(nn.Cell):\n", " def __init__(self):\n", " super(AddMulMul, self).__init__()\n", " self.param = Parameter(Tensor(0.5, ms.float32))\n", "\n", " @ms_function\n", " def construct(self, x):\n", " x = x + x\n", " x = x * self.param\n", " x = x * x\n", " return x\n", "\n", "class CellCallSingleCell(nn.Cell):\n", " def __init__(self):\n", " super(CellCallSingleCell, self).__init__()\n", " self.conv = nn.Conv2d(1, 2, kernel_size=2, stride=1, padding=0, weight_init=\"ones\", pad_mode=\"valid\")\n", " self.bn = nn.BatchNorm2d(2, momentum=0.99, eps=0.00001, gamma_init=\"ones\")\n", " self.relu = nn.ReLU()\n", " self.add_mul_mul = AddMulMul()\n", "\n", " def construct(self, x):\n", " x = self.conv(x)\n", " x = self.bn(x)\n", " x = self.add_mul_mul(x)\n", " x = self.relu(x)\n", " return x\n", "\n", "context.set_context(mode=context.PYNATIVE_MODE, device_target=\"CPU\")\n", "inputs = Tensor(np.ones([1, 1, 2, 2]).astype(np.float32))\n", "net = CellCallSingleCell()\n", "out = net(inputs)\n", "print(out)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### JIT Fallback\n", "\n", "JIT Fallback是为了实现动静统一提出的功能特性。通过JIT Fallback等特性,静态图可以支持尽量多的动态图语法,使得静态图提供接近动态图的语法使用体验。\n", "\n", "JIT Fallback是从静态图的角度出发考虑静态图和动态图的统一。MindSpore默认使用静态图模式,用户编写程序时需要遵循MindSpore[静态图语法支持](https://www.mindspore.cn/docs/zh-CN/r1.7/note/static_graph_syntax_support.html),语法使用存在约束限制。而在动态图模式下,Python脚本代码会根据Python语法进行执行,用户可以使用任意Python语法。可以看出,静态图和动态图的语法约束限制是不同的。JIT Fallback特性可以使得静态图支持尽量多的动态图语法,用户能够灵活地进行静态图和动态图的切换。\n", "\n", "当前JIT Fallback支持静态图模式的部分常量场景,包括调用第三方库、创建及使用Tensor、调用Python的print打印等。更多JIT Fallback的说明和使用,请参考[JIT Fallback文档](https://www.mindspore.cn/docs/zh-CN/r1.7/design/jit_fallback.html)。\n", "\n", "代码用例如下,其中,MindSpore静态图模式不支持在construct中调用NumPy第三方库和创建Tensor对象,因此用例中的`x = np.array([1, 2, 3])`和`y = Tensor(x)`将会通过JIT Fallback特性进行支持。" ] }, { "cell_type": "code", "execution_count": 7, "metadata": { "ExecuteTime": { "end_time": "2022-01-04T10:51:24.264847Z", "start_time": "2022-01-04T10:51:23.515957Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[1 2 3]" ] } ], "source": [ "import numpy as np\n", "import mindspore.nn as nn\n", "from mindspore import context, Tensor\n", "\n", "context.set_context(mode=context.GRAPH_MODE, device_target=\"CPU\")\n", "\n", "class Net(nn.Cell):\n", " def construct(self):\n", " x = np.array([1, 2, 3])\n", " y = Tensor(x)\n", " return y\n", "\n", "net = Net()\n", "print(net())" ] } ], "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": "3.7.5" } }, "nbformat": 4, "nbformat_minor": 4 }