diff --git a/experiment_1/1-LeNet5_MNIST.ipynb b/experiment_1/1-LeNet5_MNIST.ipynb index 7295dd5f11a2bd4b4389228c4c9c356c966e9ac2..1b385da662a797b916fc1357f59dd2f21a74511b 100644 --- a/experiment_1/1-LeNet5_MNIST.ipynb +++ b/experiment_1/1-LeNet5_MNIST.ipynb @@ -20,7 +20,7 @@ "\n", "- 熟练使用Python，了解Shell及Linux操作系统基本知识。\n", "- 具备一定的深度学习理论知识，如卷积神经网络、损失函数、优化器，训练策略等。\n", - "- 了解华为云的基本使用方法，包括[OBS（对象存储）](https://www.huaweicloud.com/product/obs.html)、[ModelArts（AI开发平台）](https://www.huaweicloud.com/product/modelarts.html)、[Notebook（开发工具）](https://support.huaweicloud.com/engineers-modelarts/modelarts_23_0033.html)、[训练作业](https://support.huaweicloud.com/engineers-modelarts/modelarts_23_0046.html)等功能。华为云官网：https://www.huaweicloud.com\n", + "- 了解华为云的基本使用方法，包括[OBS（对象存储）](https://www.huaweicloud.com/product/obs.html)、[ModelArts（AI开发平台）](https://www.huaweicloud.com/product/modelarts.html)、[Notebook（开发工具）](https://support.huaweicloud.com/engineers-modelarts/modelarts_23_0033.html)、[训练作业](https://support.huaweicloud.com/engineers-modelarts/modelarts_23_0046.html)等服务。华为云官网：https://www.huaweicloud.com\n", "- 了解并熟悉MindSpore AI计算框架，MindSpore官网：https://www.mindspore.cn\n", "\n", "## 实验环境\n", @@ -76,7 +76,8 @@ "│   └── train\n", "│   ├── train-images-idx3-ubyte\n", "│   └── train-labels-idx1-ubyte\n", - "└── 脚本等文件\n", + "├── *.ipynb\n", + "└── main.py\n", "```\n", "\n", "## 实验步骤（方案一）\n", @@ -122,6 +123,7 @@ "import mindspore.dataset.transforms.vision.c_transforms as CV\n", "\n", "from mindspore import nn\n", + "from mindspore.model_zoo.lenet import LeNet5\n", "from mindspore.train import Model\n", "from mindspore.train.callback import LossMonitor\n", "\n", @@ -149,6 +151,7 @@ "def create_dataset(training=True, num_epoch=1, batch_size=32, resize=(32, 32),\n", " rescale=1/(255*0.3081), shift=-0.1307/0.3081, buffer_size=64):\n", " ds = ms.dataset.MnistDataset(DATA_DIR_TRAIN if training else DATA_DIR_TEST)\n", + " \n", " ds = ds.map(input_columns=\"image\", operations=[CV.Resize(resize), CV.Rescale(rescale, shift), CV.HWC2CHW()])\n", " ds = ds.map(input_columns=\"label\", operations=C.TypeCast(ms.int32))\n", " ds = ds.shuffle(buffer_size=buffer_size).batch(batch_size, drop_remainder=True).repeat(num_epoch)\n", @@ -201,53 +204,20 @@ "source": [ "### 定义模型\n", "\n", - "MindSpore model_zoo中提供了现成的LeNet5模型，但当前ModelArts平台上暂未集成该模块。模型结构如下图所示：\n", + "MindSpore model_zoo中提供了多种常见的模型，可以直接使用。这里使用其中的LeNet5模型，模型结构如下图所示：\n", "\n", "\n", "\n", "[1] 图片来源于http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf" ] }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "class LeNet5(nn.Cell):\n", - " def __init__(self):\n", - " super(LeNet5, self).__init__()\n", - " self.relu = nn.ReLU()\n", - " self.conv1 = nn.Conv2d(1, 6, 5, stride=1, pad_mode='valid')\n", - " self.conv2 = nn.Conv2d(6, 16, 5, stride=1, pad_mode='valid')\n", - " self.pool = nn.MaxPool2d(kernel_size=2, stride=2)\n", - " self.flatten = nn.Flatten()\n", - " self.fc1 = nn.Dense(400, 120)\n", - " self.fc2 = nn.Dense(120, 84)\n", - " self.fc3 = nn.Dense(84, 10)\n", - " \n", - " def construct(self, input_x):\n", - " output = self.conv1(input_x)\n", - " output = self.relu(output)\n", - " output = self.pool(output)\n", - " output = self.conv2(output)\n", - " output = self.relu(output)\n", - " output = self.pool(output)\n", - " output = self.flatten(output)\n", - " output = self.fc1(output)\n", - " output = self.fc2(output)\n", - " output = self.fc3(output)\n", - " \n", - " return output" - ] - }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 训练\n", "\n", - "使用MNIST数据集对上述定义的LeNet模型进行训练。训练策略如下表所示，可以调整训练策略并查看训练效果，要求验证精度大于95%。\n", + "使用MNIST数据集对上述定义的LeNet5模型进行训练。训练策略如下表所示，可以调整训练策略并查看训练效果，要求验证精度大于95%。\n", "\n", "| batch size | number of epochs | learning rate | optimizer |\n", "| -- | -- | -- | -- |\n", @@ -292,6 +262,14 @@ "source": [ "## 实验步骤（方案二）\n", "\n", + "除了Notebook，ModelArts还提供了训练作业服务。相比Notebook，训练作业资源池更大，且具有作业排队等功能，适合大规模并发使用。使用训练作业时，也会有修改代码和调试的需求，有如下三个方案：\n", + "\n", + "1. 在本地修改代码后重新上传；\n", + "\n", + "2. 使用[PyCharm ToolKit](https://support.huaweicloud.com/tg-modelarts/modelarts_15_0001.html)配置一个本地Pycharm+ModelArts的开发环境，便于上传代码、提交训练作业和获取训练日志。\n", + "\n", + "3. 在ModelArts上创建Notebook，然后设置[Sync OBS功能](https://support.huaweicloud.com/engineers-modelarts/modelarts_23_0038.html)，可以在线修改代码并自动同步到OBS中。因为只用Notebook来编辑代码，所以创建CPU类型最低规格的Notebook就行。\n", + "\n", "### 代码梳理\n", "\n", "创建训练作业时，运行参数会通过脚本传参的方式输入给脚本代码，脚本必须解析传参才能在代码中使用相应参数。如data_url和train_url，分别对应数据存储路径(OBS路径)和训练输出路径(OBS路径)。脚本对传参进行解析后赋值到`args`变量里，在后续代码里可以使用。\n", @@ -351,7 +329,7 @@ "source": [ "## 实验小结\n", "\n", - "本实验展示了如何使用MindSpore进行手写数字识别，以及开发、训练和使用LeNet模型。通过对LeNet模型做几代的训练，然后使用训练后的LeNet模型对手写数字进行识别，识别准确率大于95%。即LeNet学习到了如何进行手写数字识别。" + "本实验展示了如何使用MindSpore进行手写数字识别，以及开发和训练LeNet5模型。通过对LeNet5模型做几代的训练，然后使用训练后的LeNet5模型对手写数字进行识别，识别准确率大于95%。即LeNet5学习到了如何进行手写数字识别。" ] } ], diff --git a/experiment_1/main.py b/experiment_1/main.py index e98ef908d23455222a3c254b117945341e6ce08a..bf236977b988793c8331547a33c90500abcc955d 100644 --- a/experiment_1/main.py +++ b/experiment_1/main.py @@ -9,6 +9,7 @@ import mindspore.dataset.transforms.c_transforms as C import mindspore.dataset.transforms.vision.c_transforms as CV from mindspore import nn +from mindspore.model_zoo.lenet import LeNet5 from mindspore.train import Model from mindspore.train.callback import LossMonitor @@ -21,6 +22,7 @@ DATA_DIR_TEST = "MNIST/test" # 测试集信息 def create_dataset(training=True, num_epoch=1, batch_size=32, resize=(32, 32), rescale=1/(255*0.3081), shift=-0.1307/0.3081, buffer_size=64): ds = ms.dataset.MnistDataset(DATA_DIR_TRAIN if training else DATA_DIR_TEST) + ds = ds.map(input_columns="image", operations=[CV.Resize(resize), CV.Rescale(rescale, shift), CV.HWC2CHW()]) ds = ds.map(input_columns="label", operations=C.TypeCast(ms.int32)) ds = ds.shuffle(buffer_size=buffer_size).batch(batch_size, drop_remainder=True).repeat(num_epoch) @@ -28,33 +30,6 @@ def create_dataset(training=True, num_epoch=1, batch_size=32, resize=(32, 32), return ds -class LeNet5(nn.Cell): - def __init__(self): - super(LeNet5, self).__init__() - self.relu = nn.ReLU() - self.conv1 = nn.Conv2d(1, 6, 5, stride=1, pad_mode='valid') - self.conv2 = nn.Conv2d(6, 16, 5, stride=1, pad_mode='valid') - self.pool = nn.MaxPool2d(kernel_size=2, stride=2) - self.flatten = nn.Flatten() - self.fc1 = nn.Dense(400, 120) - self.fc2 = nn.Dense(120, 84) - self.fc3 = nn.Dense(84, 10) - - def construct(self, input_x): - output = self.conv1(input_x) - output = self.relu(output) - output = self.pool(output) - output = self.conv2(output) - output = self.relu(output) - output = self.pool(output) - output = self.flatten(output) - output = self.fc1(output) - output = self.fc2(output) - output = self.fc3(output) - - return output - - def test_train(lr=0.01, momentum=0.9, num_epoch=3, ckpt_name="a_lenet"): ds_train = create_dataset(num_epoch=num_epoch) ds_eval = create_dataset(training=False)