{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Cell\n", "\n", "[![](https://gitee.com/mindspore/docs/raw/r1.3/resource/_static/logo_source.png)](https://gitee.com/mindspore/docs/blob/r1.3/docs/mindspore/programming_guide/source_zh_cn/cell.ipynb) [![](https://gitee.com/mindspore/docs/raw/r1.3/resource/_static/logo_notebook.png)](https://obs.dualstack.cn-north-4.myhuaweicloud.com/mindspore-website/notebook/r1.3/programming_guide/zh_cn/mindspore_cell.ipynb) [![](https://gitee.com/mindspore/docs/raw/r1.3/resource/_static/logo_modelarts.png)](https://authoring-modelarts-cnnorth4.huaweicloud.com/console/lab?share-url-b64=aHR0cHM6Ly9vYnMuZHVhbHN0YWNrLmNuLW5vcnRoLTQubXlodWF3ZWljbG91ZC5jb20vbWluZHNwb3JlLXdlYnNpdGUvbm90ZWJvb2svbW9kZWxhcnRzL3Byb2dyYW1taW5nX2d1aWRlL21pbmRzcG9yZV9jZWxsLmlweW5i&imageid=65f636a0-56cf-49df-b941-7d2a07ba8c8c)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 概述\n", "\n", "MindSpore的`Cell`类是构建所有网络的基类，也是网络的基本单元。当用户需要自定义网络时，需要继承`Cell`类，并重写`__init__`方法和`construct`方法。\n", "\n", "损失函数、优化器和模型层等本质上也属于网络结构，也需要继承`Cell`类才能实现功能，同样用户也可以根据业务需求自定义这部分内容。\n", "\n", "本节内容介绍`Cell`类的关键成员函数，“构建网络”中将介绍基于`Cell`实现的MindSpore内置损失函数、优化器和模型层及使用方法，以及通过实例介绍如何利用`Cell`类构建自定义网络。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 关键成员函数\n", "\n", "### construct方法\n", "\n", "`Cell`类重写了`__call__`方法，在`Cell`类的实例被调用时，会执行`construct`方法。网络结构在`construct`方法里面定义。\n", "\n", "下面的样例中，我们构建了一个简单的网络实现卷积计算功能。构成网络的算子在`__init__`中定义，在`construct`方法里面使用，用例的网络结构为`Conv2d` -> `BiasAdd`。\n", "\n", "在`construct`方法中，`x`为输入数据，`output`是经过网络结构计算后得到的计算结果。" ] }, { "cell_type": "code", "execution_count": 1, "metadata": { "ExecuteTime": { "end_time": "2021-02-08T01:01:31.855049Z", "start_time": "2021-02-08T01:01:31.084345Z" } }, "outputs": [], "source": [ "import mindspore.nn as nn\n", "import mindspore.ops as ops\n", "from mindspore import Parameter\n", "from mindspore.common.initializer import initializer\n", "\n", "class Net(nn.Cell):\n", " def __init__(self, in_channels=10, out_channels=20, kernel_size=3):\n", " super(Net, self).__init__()\n", " self.conv2d = ops.Conv2D(out_channels, kernel_size)\n", " self.bias_add = ops.BiasAdd()\n", " self.weight = Parameter(\n", " initializer('normal', [out_channels, in_channels, kernel_size, kernel_size]),\n", " name='conv.weight')\n", "\n", " def construct(self, x):\n", " output = self.conv2d(x, self.weight)\n", " output = self.bias_add(output, self.bias)\n", " return output" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### parameters_dict\n", "\n", "`parameters_dict`方法识别出网络结构中所有的参数，返回一个以key为参数名，value为参数值的`OrderedDict`。\n", "\n", "`Cell`类中返回参数的方法还有许多，例如`get_parameters`、`trainable_params`等，具体使用方法可以参见[API文档](https://www.mindspore.cn/docs/api/zh-CN/r1.3/api_python/nn/mindspore.nn.Cell.html)。\n", "\n", "代码样例如下：" ] }, { "cell_type": "code", "execution_count": 2, "metadata": { "ExecuteTime": { "end_time": "2021-02-08T01:01:31.867924Z", "start_time": "2021-02-08T01:01:31.856066Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "odict_keys(['conv.weight'])\n", "Parameter (name=conv.weight, shape=(20, 10, 3, 3), dtype=Float32, requires_grad=True)\n" ] } ], "source": [ "net = Net()\n", "result = net.parameters_dict()\n", "print(result.keys())\n", "print(result['conv.weight'])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "样例中的`Net`采用上文构造网络的用例，打印了网络中所有参数的名字和`weight`参数的结果。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### cells_and_names\n", "\n", "`cells_and_names`方法是一个迭代器，返回网络中每个`Cell`的名字和它的内容本身。\n", "\n", "用例简单实现了获取与打印每个`Cell`名字的功能，其中根据网络结构可知，存在1个`Cell`为`nn.Conv2d`。\n", "\n", "其中`nn.Conv2d`是`MindSpore`以Cell为基类封装好的一个卷积层，其具体内容将在“模型层”中进行介绍。\n", "\n", "代码样例如下：" ] }, { "cell_type": "code", "execution_count": 3, "metadata": { "ExecuteTime": { "end_time": "2021-02-08T01:01:31.893191Z", "start_time": "2021-02-08T01:01:31.870508Z" } }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "('', Net1<\n", " (conv): Conv2d\n", " >)\n", "('conv', Conv2d)\n", "-------names-------\n", "['conv']\n" ] } ], "source": [ "import mindspore.nn as nn\n", "\n", "class Net1(nn.Cell):\n", " def __init__(self):\n", " super(Net1, self).__init__()\n", " self.conv = nn.Conv2d(3, 64, 3, has_bias=False, weight_init='normal')\n", "\n", " def construct(self, x):\n", " out = self.conv(x)\n", " return out\n", "\n", "net = Net1()\n", "names = []\n", "for m in net.cells_and_names():\n", " print(m)\n", " names.append(m[0]) if m[0] else None\n", "print('-------names-------')\n", "print(names)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### set_grad\n", "\n", "`set_grad`接口功能是使用户构建反向网络，在不传入参数调用时，默认设置`requires_grad`为True，需要在计算网络反向的场景中使用。\n", "\n", "以`TrainOneStepCell`为例，其接口功能是使网络进行单步训练，需要计算网络反向，因此初始化方法里需要使用`set_grad`。\n", "\n", "`TrainOneStepCell`部分代码如下：" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "```python\n", "class TrainOneStepCell(Cell):\n", " def __init__(self, network, optimizer, sens=1.0):\n", " super(TrainOneStepCell, self).__init__(auto_prefix=False)\n", " self.network = network\n", " self.network.set_grad()\n", " ......\n", "```" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "如果用户使用`TrainOneStepCell`等类似接口无需使用`set_grad`， 内部已封装实现。\n", "\n", "若用户需要自定义此类训练功能的接口，需要在其内部调用，或者在外部设置`network.set_grad`。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## nn模块与ops模块的关系\n", "\n", "MindSpore的nn模块是Python实现的模型组件，是对低阶API的封装，主要包括各种模型层、损失函数、优化器等。\n", "\n", "同时nn也提供了部分与`Primitive`算子同名的接口，主要作用是对`Primitive`算子进行进一步封装，为用户提供更友好的API。\n", "\n", "重新分析上文介绍`construct`方法的用例，此用例是MindSpore的`nn.Conv2d`源码简化内容，内部会调用`ops.Conv2D`。`nn.Conv2d`卷积API增加输入参数校验功能并判断是否`bias`等，是一个高级封装的模型层。" ] }, { "cell_type": "code", "execution_count": 4, "metadata": { "ExecuteTime": { "end_time": "2021-02-08T01:01:31.916550Z", "start_time": "2021-02-08T01:01:31.894206Z" } }, "outputs": [], "source": [ "import mindspore.nn as nn\n", "import mindspore.ops as ops\n", "from mindspore import Parameter\n", "from mindspore.common.initializer import initializer\n", "\n", "class Net(nn.Cell):\n", " def __init__(self, in_channels=10, out_channels=20, kernel_size=3):\n", " super(Net, self).__init__()\n", " self.conv2d = ops.Conv2D(out_channels, kernel_size)\n", " self.bias_add = ops.BiasAdd()\n", " self.weight = Parameter(\n", " initializer('normal', [out_channels, in_channels, kernel_size, kernel_size]),\n", " name='conv.weight')\n", "\n", " def construct(self, x):\n", " output = self.conv2d(x, self.weight)\n", " output = self.bias_add(output, self.bias)\n", " return output" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "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 }