ResNet18代码实现（tensorflow，80行）

前言：

ResNet18的实现还是相对比较简单的，一共就18次卷积，从数据加载到最后完成验证不到80行

import tensorflow as tf
import pandas as pd
import numpy as np
from tensorflow.keras import layers
from sklearn.metrics import classification_report
import warnings
warnings.filterwarnings('ignore')

#数据导入
train_data = pd.read_csv('sign_train.csv')
test_data = pd.read_csv('sign_test.csv')

train_x = train_data.drop('label',axis=1).values
train_y = train_data['label'].values

test_x = test_data.drop('label',axis=1).values
test_y = test_data['label'].values

#数据归一化
train_x = train_x / 255
test_x = test_x / 255

train_x = train_x.reshape((-1, 28, 28, 1))
test_x = test_x.reshape((-1, 28, 28, 1))

#模型建立
def conv2d_bn(inpt, filters=64, kernel_size=(3,3), strides=1, padding='same'):
    '''卷积、归一化和relu三合一'''
    x = layers.Conv2D(filters=filters, kernel_size=kernel_size, strides=strides, padding=padding)(inpt)
    x = layers.BatchNormalization()(x)
    x = tf.keras.layers.Activation('relu')(x)
    return x

def basic_bottle(inpt, filters=64, kernel_size=(3,3), strides=1, padding='same', if_baisc=False):
    '''18中的4个basic_bottle'''
    x = conv2d_bn(inpt, filters=filters, kernel_size=kernel_size, strides=strides, padding=padding)
    x = conv2d_bn(x, filters=filters)
    if if_baisc==True:
        temp = conv2d_bn(inpt, filters=filters, kernel_size=(1,1), strides=2, padding='same')
        outt = layers.add([x, temp])
    else:
        outt = layers.add([x, inpt])
    return outt

def resnet18(class_nums):
    '''主模型'''
    inpt = layers.Input(shape=(28,28,1))
    #layer 1
    x = conv2d_bn(inpt, filters=64, kernel_size=(7,7), strides=2, padding='valid')
    x = layers.MaxPool2D(pool_size=(3,3), strides=2)(x)
    #layer 2
    x = basic_bottle(x, filters=64, kernel_size=(3,3), strides=1, padding='same', if_baisc=False)
    x = basic_bottle(x, filters=64, kernel_size=(3,3), strides=1, padding='same', if_baisc=False)
    #layer 3
    x = basic_bottle(x, filters=128, kernel_size=(3, 3), strides=2, padding='same', if_baisc=True)
    x = basic_bottle(x, filters=128, kernel_size=(3, 3), strides=1, padding='same', if_baisc=False)
    # layer 4
    x = basic_bottle(x, filters=256, kernel_size=(3, 3), strides=2, padding='same', if_baisc=True)
    x = basic_bottle(x, filters=256, kernel_size=(3, 3), strides=1, padding='same', if_baisc=False)
    # layer 5
    x = basic_bottle(x, filters=512, kernel_size=(3, 3), strides=2, padding='same', if_baisc=True)
    x = basic_bottle(x, filters=512, kernel_size=(3, 3), strides=1, padding='same', if_baisc=False)
    #GlobalAveragePool
    x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dense(class_nums, activation='softmax')(x)
    model = tf.keras.Model(inputs=inpt, outputs=x)
    return model

#开始训练
class_num = len(set(train_y))+1
model = resnet18(class_num)
model.summary()
model.compile(optimizer=tf.keras.optimizers.Adam(),loss=tf.keras.losses.SparseCategoricalCrossentropy(), metrics=['accuracy'])
model.fit(train_x, train_y, batch_size=64, epochs=10, validation_split=0.2) #epochs可以调的大一点，如果你机器好的话

#验证结果
pred_y = model.predict(test_x)
last_pred = []
for i in range(pred_y.shape[0]):
    last_pred.append(pred_y[i].argmax())
last_pred = np.array(last_pred)
print(classification_report(test_y, last_pred))