{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 开发入门\n",
"\n",
"[](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r2.4.10/zh_cn/orange_pi/mindspore_dev_start.ipynb) [](https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/r2.4.10/zh_cn/orange_pi/mindspore_dev_start.py) [](https://gitee.com/mindspore/docs/blob/r2.4.10/docs/mindspore/source_zh_cn/orange_pi/dev_start.ipynb)\n",
"\n",
"因开发者可能会在OrangePi AIpro（下称：香橙派开发板）进行自定义模型和案例开发，本章节通过基于MindSpore的手写数字识别案例，说明香橙派开发板中的开发注意事项。"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"from mindspore import nn\n",
"from mindspore.dataset import vision, transforms\n",
"from mindspore.dataset import MnistDataset"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 设置运行环境\n",
"\n",
"> 由于Mindspore2.3及之前的版本暂未支持内存动态按需申请，以及CANN缺少对应的动态算子，所以执行案例之前需要通过set_context设置运行环境。后续会随着版本不断更新解决以上问题，从而在不需要进行环境配置的前提下可直接执行案例。\n",
"\n",
" max_device_memory=\"2GB\" : 设置设备可用的最大内存为2GB。\n",
"\n",
" mode=mindspore.GRAPH_MODE : 表示在GRAPH_MODE模式中运行。\n",
"\n",
" device_target=\"Ascend\" : 表示待运行的目标设备为Ascend。\n",
"\n",
" jit_config={\"jit_level\":\"O2\"} : 编译优化级别开启极致性能优化，使用下沉的执行方式。\n",
"\n",
" ascend_config={\"precision_mode\":\"allow_mix_precision\"} : 自动混合精度，自动将部分算子的精度降低到float16或bfloat16。"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"import mindspore\n",
"mindspore.set_context(max_device_memory=\"2GB\", mode=mindspore.GRAPH_MODE, device_target=\"Ascend\", jit_config={\"jit_level\":\"O2\"}, ascend_config={\"precision_mode\":\"allow_mix_precision\"})"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 处理数据集\n",
"\n",
"MindSpore提供基于Pipeline的[数据引擎](https://www.mindspore.cn/docs/zh-CN/r2.4.10/design/data_engine.html)，通过[数据加载与处理](https://www.mindspore.cn/tutorials/zh-CN/r2.4.10/beginner/dataset.html)实现高效的数据预处理。在本案例中，我们使用Mnist数据集，自动下载完成后，使用`mindspore.dataset`提供的数据变换进行预处理。\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Downloading data from https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/MNIST_Data.zip (10.3 MB)\n",
"\n",
"file_sizes: 100%|██████████████████████████| 10.8M/10.8M [00:01<00:00, 7.63MB/s]\n",
"Extracting zip file...\n",
"Successfully downloaded / unzipped to ./\n"
]
}
],
"source": [
"# Download data from open datasets\n",
"from download import download\n",
"\n",
"url = \"https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/\" \\\n",
" \"notebook/datasets/MNIST_Data.zip\"\n",
"path = download(url, \"./\", kind=\"zip\", replace=True)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"MNIST数据集目录结构如下：\n",
"\n",
"```text\n",
"MNIST_Data\n",
"└── train\n",
" ├── train-images-idx3-ubyte (60000个训练图片)\n",
" ├── train-labels-idx1-ubyte (60000个训练标签)\n",
"└── test\n",
" ├── t10k-images-idx3-ubyte (10000个测试图片)\n",
" ├── t10k-labels-idx1-ubyte (10000个测试标签)\n",
"\n",
"```\n",
"\n",
"数据下载完成后，获得数据集对象。"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"train_dataset = MnistDataset('MNIST_Data/train')\n",
"test_dataset = MnistDataset('MNIST_Data/test')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"打印数据集中包含的数据列名，用于dataset的预处理。"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"['image', 'label']\n"
]
}
],
"source": [
"print(train_dataset.get_col_names())"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"MindSpore的dataset使用数据处理流水线（Data Processing Pipeline），需指定map、batch、shuffle等操作。这里我们使用map对图像数据及标签进行变换处理，将输入的图像缩放为1/255，根据均值0.1307和标准差值0.3081进行归一化处理，然后将处理好的数据集打包为大小为64的batch。"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"def datapipe(dataset, batch_size):\n",
" image_transforms = [\n",
" vision.Rescale(1.0 / 255.0, 0),\n",
" vision.Normalize(mean=(0.1307,), std=(0.3081,)),\n",
" vision.HWC2CHW()\n",
" ]\n",
" label_transform = transforms.TypeCast(mindspore.int32)\n",
"\n",
" dataset = dataset.map(image_transforms, 'image')\n",
" dataset = dataset.map(label_transform, 'label')\n",
" dataset = dataset.batch(batch_size)\n",
" return dataset"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"# Map vision transforms and batch dataset\n",
"train_dataset = datapipe(train_dataset, 64)\n",
"test_dataset = datapipe(test_dataset, 64)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"可使用[create_tuple_iterator](https://www.mindspore.cn/docs/zh-CN/r2.4.10/api_python/dataset/dataset_method/iterator/mindspore.dataset.Dataset.create_tuple_iterator.html) 或[create_dict_iterator](https://www.mindspore.cn/docs/zh-CN/r2.4.10/api_python/dataset/dataset_method/iterator/mindspore.dataset.Dataset.create_dict_iterator.html)对数据集进行迭代访问，查看数据和标签的shape和datatype。"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Shape of image [N, C, H, W]: (64, 1, 28, 28) Float32\n",
"Shape of label: (64,) Int32\n"
]
}
],
"source": [
"for image, label in test_dataset.create_tuple_iterator():\n",
" print(f\"Shape of image [N, C, H, W]: {image.shape} {image.dtype}\")\n",
" print(f\"Shape of label: {label.shape} {label.dtype}\")\n",
" break"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Shape of image [N, C, H, W]: (64, 1, 28, 28) Float32\n",
"Shape of label: (64,) Int32\n"
]
}
],
"source": [
"for data in test_dataset.create_dict_iterator():\n",
" print(f\"Shape of image [N, C, H, W]: {data['image'].shape} {data['image'].dtype}\")\n",
" print(f\"Shape of label: {data['label'].shape} {data['label'].dtype}\")\n",
" break"
]
},
{
"cell_type": "code",
"execution_count": 11,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Network<\n",
" (flatten): Flatten<>\n",
" (dense_relu_sequential): SequentialCell<\n",
" (0): Dense<input_channels=784, output_channels=512, has_bias=True>\n",
" (1): ReLU<>\n",
" (2): Dense<input_channels=512, output_channels=512, has_bias=True>\n",
" (3): ReLU<>\n",
" (4): Dense<input_channels=512, output_channels=10, has_bias=True>\n",
" >\n",
" >\n"
]
}
],
"source": [
"# Define model\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",
"print(model)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 模型训练"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"在模型训练中，一个完整的训练过程（step）需要实现以下三步：\n",
"\n",
"1. **正向计算**：模型预测结果（logits），并与正确标签（label）求预测损失（loss）。\n",
"2. **反向传播**：利用自动微分机制，自动求模型参数（parameters）对于loss的梯度（gradients）。\n",
"3. **参数优化**：将梯度更新到参数上。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"MindSpore使用函数式自动微分机制，因此针对上述步骤需要实现：\n",
"\n",
"1. 定义正向计算函数。\n",
"2. 使用[value_and_grad](https://www.mindspore.cn/docs/zh-CN/r2.4.10/api_python/mindspore/mindspore.value_and_grad.html)通过函数变换获得梯度计算函数。\n",
"3. 定义训练函数，使用[set_train](https://www.mindspore.cn/docs/zh-CN/r2.4.10/api_python/nn/mindspore.nn.Cell.html#mindspore.nn.Cell.set_train)设置为训练模式，执行正向计算、反向传播和参数优化。"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
"# Instantiate loss function and optimizer\n",
"loss_fn = nn.CrossEntropyLoss()\n",
"optimizer = nn.SGD(model.trainable_params(), 1e-2)\n",
"\n",
"# 1. Define forward function\n",
"def forward_fn(data, label):\n",
" logits = model(data)\n",
" loss = loss_fn(logits, label)\n",
" return loss, logits\n",
"\n",
"# 2. Get gradient function\n",
"grad_fn = mindspore.value_and_grad(forward_fn, None, optimizer.parameters, has_aux=True)\n",
"\n",
"# 3. Define function of one-step training\n",
"def train_step(data, label):\n",
" (loss, _), grads = grad_fn(data, label)\n",
" optimizer(grads)\n",
" return loss\n",
"\n",
"def train(model, dataset):\n",
" size = dataset.get_dataset_size()\n",
" model.set_train()\n",
" for batch, (data, label) in enumerate(dataset.create_tuple_iterator()):\n",
" loss = train_step(data, label)\n",
"\n",
" if batch % 100 == 0:\n",
" loss, current = loss.asnumpy(), batch\n",
" print(f\"loss: {loss:>7f} [{current:>3d}/{size:>3d}]\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"除训练外，我们定义测试函数，用来评估模型的性能。"
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {},
"outputs": [],
"source": [
"def test(model, dataset, loss_fn):\n",
" num_batches = dataset.get_dataset_size()\n",
" model.set_train(False)\n",
" total, test_loss, correct = 0, 0, 0\n",
" for data, label in dataset.create_tuple_iterator():\n",
" pred = model(data)\n",
" total += len(data)\n",
" test_loss += loss_fn(pred, label).asnumpy()\n",
" correct += (pred.argmax(1) == label).asnumpy().sum()\n",
" test_loss /= num_batches\n",
" correct /= total\n",
" print(f\"Test: \\n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \\n\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"训练过程需多次迭代数据集，一次完整的迭代称为一轮（epoch）。在每一轮，遍历训练集进行训练，结束后使用测试集进行预测。打印每一轮的loss值和预测准确率（Accuracy），可以看到loss在不断下降，Accuracy在不断提高。"
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch 1\n",
"-------------------------------\n",
"loss: 2.302898 [ 0/938]\n",
"loss: 1.729961 [100/938]\n",
"loss: 0.865714 [200/938]\n",
"loss: 0.782822 [300/938]\n",
"loss: 0.389282 [400/938]\n",
"loss: 0.293149 [500/938]\n",
"loss: 0.474819 [600/938]\n",
"loss: 0.242542 [700/938]\n",
"loss: 0.542277 [800/938]\n",
"loss: 0.342929 [900/938]\n",
"Test: \n",
" Accuracy: 90.7%, Avg loss: 0.321954 \n",
"\n",
"Epoch 2\n",
"-------------------------------\n",
"loss: 0.249492 [ 0/938]\n",
"loss: 0.347967 [100/938]\n",
"loss: 0.220382 [200/938]\n",
"loss: 0.308149 [300/938]\n",
"loss: 0.353044 [400/938]\n",
"loss: 0.392116 [500/938]\n",
"loss: 0.396438 [600/938]\n",
"loss: 0.231412 [700/938]\n",
"loss: 0.194819 [800/938]\n",
"loss: 0.228290 [900/938]\n",
"Test: \n",
" Accuracy: 93.0%, Avg loss: 0.249993 \n",
"\n",
"Epoch 3\n",
"-------------------------------\n",
"loss: 0.343888 [ 0/938]\n",
"loss: 0.307786 [100/938]\n",
"loss: 0.153425 [200/938]\n",
"loss: 0.254917 [300/938]\n",
"loss: 0.198072 [400/938]\n",
"loss: 0.108963 [500/938]\n",
"loss: 0.202033 [600/938]\n",
"loss: 0.340418 [700/938]\n",
"loss: 0.144911 [800/938]\n",
"loss: 0.175447 [900/938]\n",
"Test: \n",
" Accuracy: 93.7%, Avg loss: 0.212180 \n",
"\n",
"Done!\n"
]
}
],
"source": [
"epochs = 3\n",
"for t in range(epochs):\n",
" print(f\"Epoch {t+1}\\n-------------------------------\")\n",
" train(model, train_dataset)\n",
" test(model, test_dataset, loss_fn)\n",
"print(\"Done!\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 保存模型\n",
"\n",
"模型训练完成后，需要将其参数进行保存。"
]
},
{
"cell_type": "code",
"execution_count": 15,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Saved Model to model.ckpt\n"
]
}
],
"source": [
"# Save checkpoint\n",
"mindspore.save_checkpoint(model, \"model.ckpt\")\n",
"print(\"Saved Model to model.ckpt\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 加载模型"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"加载保存的权重分为两步：\n",
"\n",
"1. 重新实例化模型对象，构造模型。\n",
"2. 加载模型参数，并将其加载至模型上。"
]
},
{
"cell_type": "code",
"execution_count": 16,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"[]\n"
]
}
],
"source": [
"# Instantiate a random initialized model\n",
"model = Network()\n",
"# Load checkpoint and load parameter to model\n",
"param_dict = mindspore.load_checkpoint(\"model.ckpt\")\n",
"param_not_load, _ = mindspore.load_param_into_net(model, param_dict)\n",
"print(param_not_load)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"> `param_not_load`是未被加载的参数列表，为空时代表所有参数均加载成功。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"加载后的模型可以直接用于预测推理。"
]
},
{
"cell_type": "code",
"execution_count": 17,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Predicted: \"[1 2 0 4 6 4 9 0 2 2]\", Actual: \"[1 2 0 4 6 9 9 0 2 2]\"\n"
]
}
],
"source": [
"model.set_train(False)\n",
"for data, label in test_dataset:\n",
" pred = model(data)\n",
" predicted = pred.argmax(1)\n",
" print(f'Predicted: \"{predicted[:10]}\", Actual: \"{label[:10]}\"')\n",
" break"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 更多基于昇思MindSpore的香橙派开发板案例详见：[GitHub链接](https://github.com/mindspore-courses/orange-pi-mindspore)\n"
]
}
],
"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": "3.9.18"
},
"vscode": {
"interpreter": {
"hash": "8c9da313289c39257cb28b126d2dadd33153d4da4d524f730c81a4aaccbd2ca7"
}
}
},
"nbformat": 4,
"nbformat_minor": 4
}