RT-Thread-CIFAR10分类CNN模型训练加保存RT-Thread问答社区

CIFAR10分类CNN模型训练加保存

发布于 2021-06-16 15:11:23 浏览：732 订阅该版

[tocm]

CIFAR10分类CNN模型搭建&数据集处理&训练&保存. 模型简单可以尝试将其移植在嵌入式端实验. 附件提供jupyter notebook文件供实验.
```
#!/usr/bin/env python
# coding: utf-8

# In[1]:

#cifar
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import backend
from tensorflow.keras import layers
import os
import numpy as np 
from matplotlib import pyplot as plt
# os.environ["CUDA_VISIBLE_DEVICES"] = "1"
get_ipython().run_line_magic('matplotlib', 'inline')

# In[2]:

# create CNN
def CNNmodel(input_shape,filters=64, kernel=(3,3),size=4,dropout=0.2,**kwargs):
    _inputs = layers.Input(shape=input_shape)
    x = layers.Conv2D(8,(3,3),padding='same',use_bias=False,strides=(2,2), name='conv_0')(_inputs)
    x = layers.BatchNormalization(axis=-1, name='conv_0_bn')(x)
    x = layers.ReLU(6., name='conv_0_relu')(x)
    
    x = layers.Conv2D(16,(3,3),padding='same',use_bias=False,strides=(2,2), name='conv_1')(_inputs)
    x = layers.BatchNormalization(axis=-1, name='conv_1_bn')(x)
    x = layers.ReLU(6., name='conv_1_relu')(x)

for block_id in range(2,size+2):
        x = layers.Conv2D(filters,kernel,padding='same',use_bias=False,strides=(1,1), name='conv_%d'%block_id)(x)
        x = layers.BatchNormalization(axis=-1, name='conv_%d_bn'%block_id)(x)
        x = layers.ReLU(6., name='conv_%d_relu'%block_id)(x)

x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dropout(dropout, name='dropout')(x)
    x = layers.Dense(10)(x)
    x = layers.Softmax()(x)
    return keras.Model(inputs=_inputs,outputs=x)

# In[3]:

# preprocess input
def preprocess_input(inputs, std=255. ,mean=0., expand_dims=None):
    inputs = tf.cast(inputs,tf.float32)
    inputs = (inputs - mean) / std
    if expand_dims is not None:
        np.expand_dims(inputs,expand_dims)
    return inputs

# dataset aug
def img_aug_fun(elem):
    elem = tf.image.random_flip_left_right(elem)#左右翻转
    elem = tf.image.random_brightness(elem, max_delta=0.5)#调亮度
    elem = tf.image.random_contrast(elem, lower=0.5, upper=1.5)#调对比度
    elem = preprocess_input(elem)
    return elem

# load CIFAR10 dataset, size(32,32,3)
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()
assert x_train.shape == (50000, 32, 32, 3)
assert x_test.shape == (10000, 32, 32, 3)
assert y_train.shape == (50000, 1)
assert y_test.shape == (10000, 1)

x_test = preprocess_input(x_test)
x_train_ds = tf.data.Dataset.from_tensor_slices(x_train).map(img_aug_fun)
y_train_ds = tf.data.Dataset.from_tensor_slices(y_train)
x_y_train_ds = tf.data.Dataset.zip((x_train_ds,y_train_ds))
x_y_train_ds = x_y_train_ds.batch(128)

# In[4]:

reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor='accuracy', factor=0.5, patience=4, min_lr=0.0001,verbose=1)
earlystop = keras.callbacks.EarlyStopping(monitor='val_accuracy', patience=8,verbose=1)

model = CNNmodel(input_shape=(32,32,3),filters=64, kernel=(3,3),size=9)
model.compile(optimizer='SGD',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
history = model.fit(x_y_train_ds,validation_data=(x_test,y_test),callbacks=[reduce_lr,earlystop],verbose=2,epochs=500)

# In[5]:

plt.plot(history.history['val_accuracy'],label='val_acc')
plt.legend()
plt.xlabel('Epochs')
plt.ylabel('Acc')
plt.show()

# In[6]:

model.save('checkpoint/Cifar10_CNN_%.3f'%history.history['val_accuracy'][-1]+'.h5')

[cifar_cnn.ipynb](https://oss-club.rt-thread.org/uploads/20210616/8ed570671e76b0075cd78b42a36f8545.ipynb)

```