# cnn-hello

**Repository Path**: fengyefly/cnn-hello

## Basic Information

- **Project Name**: cnn-hello
- **Description**: cnn训练的简单示例
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-04-01
- **Last Updated**: 2025-04-01

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## 前言
本仓库中代码提供一个从在windows11上安装tensorflow训练环境，并且进行cnn训练的简单示例，采用mnist提供的数据集。

## 安装
### 1. 创建新的conda环境，conda环境名称为t2，python版本为3.10
`conda create -n t2 python=3.10`

### 2. 进入刚创建conda环境：t1
`conda activate t2`

### 3. 设置conda清华源
```
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/
conda config --show channels
```

### 4. 安装cudatoolkit的11.2版本和cudnn的8.1版本
`conda install cudatoolkit=11.2 cudnn=8.1`

以下是tensorflow-gpu与cuda和cudnn的版本对照：
![tensorflow-gpu-version.png](imgs/tensorflow-gpu-version.png)

### 5. 安装tensorflow的2.10.0版本（因为pip源是单独控制的，所以这里临时指定）
`pip install tensorflow==2.10.0 -i https://pypi.tuna.tsinghua.edu.cn/simple`

### 6. 安装tensorflow-gpu的2.10.0版本
`pip install tensorflow-gpu==2.10.0 -i https://pypi.tuna.tsinghua.edu.cn/simple`

### 7. 降低numpy的版本到1.22.4
`pip install numpy==1.22.4 -i https://pypi.tuna.tsinghua.edu.cn/simple`


## 注意事项
1. 在程序中`numpy<2`报错，就是因为numpy版本不匹配导致，降低numpy的版本到1.22.4可解决。
2. 由tensorflow官网可知，tensorflow 2.11开始windows不支持CUDA构建，要在 Windows 上使用 TensorFlow GPU，您需要在 WSL2 中构建/安装 TensorFlow 或将 tensorflow-cpu 与 TensorFlow-DirectML-Plugin 一起使用。（后续会研究“在WSL2中构建/安装TensorFlow”）


## 训练
运行代码文件`cnn-hello-train.py`。
### 1、使用cpu训练
```
Epoch 1/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 5s 16ms/step - accuracy: 0.5552 - loss: 6.6226 - val_accuracy: 0.9600 - val_loss: 0.1341
Epoch 2/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 15ms/step - accuracy: 0.9106 - loss: 0.3238 - val_accuracy: 0.9789 - val_loss: 0.0774
Epoch 3/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 16ms/step - accuracy: 0.9420 - loss: 0.2108 - val_accuracy: 0.9833 - val_loss: 0.0619
Epoch 4/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 15ms/step - accuracy: 0.9541 - loss: 0.1685 - val_accuracy: 0.9842 - val_loss: 0.0576
Epoch 5/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 3s 14ms/step - accuracy: 0.9613 - loss: 0.1431 - val_accuracy: 0.9845 - val_loss: 0.0543
Epoch 6/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 16ms/step - accuracy: 0.9654 - loss: 0.1260 - val_accuracy: 0.9861 - val_loss: 0.0533
Epoch 7/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 17ms/step - accuracy: 0.9689 - loss: 0.1126 - val_accuracy: 0.9865 - val_loss: 0.0481
Epoch 8/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 16ms/step - accuracy: 0.9701 - loss: 0.1101 - val_accuracy: 0.9871 - val_loss: 0.0542
Epoch 9/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 15ms/step - accuracy: 0.9716 - loss: 0.1022 - val_accuracy: 0.9873 - val_loss: 0.0526
Epoch 10/10
240/240 ━━━━━━━━━━━━━━━━━━━━ 4s 16ms/step - accuracy: 0.9739 - loss: 0.0922 - val_accuracy: 0.9871 - val_loss: 0.0459
313/313 ━━━━━━━━━━━━━━━━━━━━ 1s 2ms/step - accuracy: 0.9843 - loss: 0.0517
测试集预测准确率为： [0.039050184190273285, 0.9883000254631042]
```
共计：约40s

cpu占用如图：
![cpu-train.png](imgs/cpu-train.png)


### 2、使用gpu训练
```
Epoch 1/10
2025-04-01 00:20:35.986852: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8100
2025-04-01 00:20:37.184923: I tensorflow/stream_executor/cuda/cuda_blas.cc:1614] TensorFloat-32 will be used for the matrix multiplication. This will only be logged once.
240/240 [==============================] - 4s 4ms/step - loss: 1.7912 - accuracy: 0.6880 - val_loss: 0.1344 - val_accuracy: 0.9595
Epoch 2/10
240/240 [==============================] - 1s 4ms/step - loss: 0.2980 - accuracy: 0.9168 - val_loss: 0.0778 - val_accuracy: 0.9776
Epoch 3/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1997 - accuracy: 0.9456 - val_loss: 0.0596 - val_accuracy: 0.9820
Epoch 4/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1590 - accuracy: 0.9566 - val_loss: 0.0528 - val_accuracy: 0.9849
Epoch 5/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1371 - accuracy: 0.9627 - val_loss: 0.0549 - val_accuracy: 0.9849
Epoch 6/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1270 - accuracy: 0.9656 - val_loss: 0.0464 - val_accuracy: 0.9877
Epoch 7/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1143 - accuracy: 0.9684 - val_loss: 0.0494 - val_accuracy: 0.9883
Epoch 8/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1063 - accuracy: 0.9710 - val_loss: 0.0508 - val_accuracy: 0.9887
Epoch 9/10
240/240 [==============================] - 1s 4ms/step - loss: 0.1025 - accuracy: 0.9720 - val_loss: 0.0463 - val_accuracy: 0.9883
Epoch 10/10
240/240 [==============================] - 1s 4ms/step - loss: 0.0984 - accuracy: 0.9730 - val_loss: 0.0452 - val_accuracy: 0.9887
训练耗时：12.295096635818481s
313/313 [==============================] - 1s 1ms/step - loss: 0.0437 - accuracy: 0.9875
测试集预测准确率为： [0.04365367442369461, 0.987500011920929]
```
共计：约12.3s

gpu占用如图：
![gpu-train1.png](imgs/gpu-train1.png)
![gpu-train.png](imgs/gpu-train.png)