ResNet50 使用cifar-10数据集训练，出现了训练精度高，验证精度仅有75%的问题。还请各位同学不吝赐教！
ResNet50 使用cifar-10数据集训练，出现了训练精度高，验证精度仅有75%的问题。还请各位同学不吝赐教！
由于是第一次在电脑上跑实验，代码是从TensorFlow2.0卷积神经网络实战上的代码，经过自己调试之后程序可以正常执行，但是结果不是很好，还请各位道友不吝赐教！
下面抛出代码：
获取数据集的代码：
import pickle
import numpy as np
import os
from PIL import Image
import tensorflow as tf
def get_cifar10_train_data_and_label(root = ""):
print("-------------------------------------------------------------")
print("获取训练数据集和标签")
def load_file(filename):
with open(filename, 'rb') as fo:
data = pickle.load(fo, encoding='latin1')
return data
data_batch_1 = load_file(os.path.join(root, 'data_batch_1'))
data_batch_2 = load_file(os.path.join(root, 'data_batch_2'))
data_batch_3 = load_file(os.path.join(root, 'data_batch_3'))
data_batch_4 = load_file(os.path.join(root, 'data_batch_4'))
data_batch_5 = load_file(os.path.join(root, 'data_batch_5'))
dataset = []
labelset = []
for data in [data_batch_1,data_batch_2,data_batch_3,data_batch_4,data_batch_5]:
img_data = (data["data"])
img_label = (data["labels"])
for i in img_data:
i = i.reshape(1,3,32,32)
i = tf.transpose(i,[0,2,3,1])
dataset.append(i)
labelset.append(img_label)
dataset = np.concatenate(dataset)
labelset = np.concatenate(labelset)
print("-------------------------------------------------------------")
return dataset,labelset
def get_cifar10_test_data_and_label(root = ""):
print("-------------------------------------------------------------")
print("获取测试数据集和标签")
def load_file(filename):
with open(filename, 'rb') as fo:
data = pickle.load(fo, encoding='latin1')
return data
data_batch_1 = load_file(os.path.join(root, 'test_batch'))
dataset = []
labelset = []
for data in [data_batch_1]:
img_data = (data["data"])
img_label = (data["labels"])
for i in img_data:
i = i.reshape(1,3,32,32)
i = tf.transpose(i,[0,2,3,1])
dataset.append(i)
labelset.append(img_label)
dataset = np.concatenate(dataset)
labelset = np.concatenate(labelset)
print("-------------------------------------------------------------")
return dataset,labelset
def get_CIFAR100_dataset(root = ""):
train_dataset,label_dataset = get_cifar10_train_data_and_label(root=" ")
print(train_dataset.shape)
test_dataset,test_label_dataset = get_cifar10_test_data_and_label(root=root)
return train_dataset,label_dataset,test_dataset,test_label_dataset
if __name__ == "__main__":
get_CIFAR100_dataset(root="C:/Users/thinkpad/SpyderProjects/ResnetDemo/cifar-10-python/cifar-10-batches-py/")
下面是定义模型的代码：
import tensorflow as tf
def identity_block(input_tensor,out_dim):
conv1 = tf.keras.layers.Conv2D(out_dim // 4, kernel_size=1, padding="SAME", activation=tf.nn.relu)(input_tensor)
conv2 = tf.keras.layers.BatchNormalization()(conv1)
conv3 = tf.keras.layers.Conv2D(out_dim // 4, kernel_size=3, padding="SAME", activation=tf.nn.relu)(conv2)
conv4 = tf.keras.layers.BatchNormalization()(conv3)
conv5 = tf.keras.layers.Conv2D(out_dim, kernel_size=1, padding="SAME")(conv4)
#残差
out = tf.keras.layers.Add()([input_tensor, conv5])
out = tf.nn.relu(out)
return out
def resnet_Model():
input_xs = tf.keras.Input(shape=[32,32,3])#rgb图像32,32,3
conv_1 = tf.keras.layers.Conv2D(filters=32,kernel_size=3,padding="SAME",activation=tf.nn.relu)(input_xs)
"""--------第一层----------"""
out_dim = 32
identity_1 = tf.keras.layers.Conv2D(filters=out_dim, kernel_size=3, padding="SAME", activation=tf.nn.relu)(conv_1)
identity_1 = tf.keras.layers.BatchNormalization()(identity_1)
for _ in range(3):
identity_1 = identity_block(identity_1,out_dim)
"""--------第二层----------"""
out_dim = 64
identity_2 = tf.keras.layers.Conv2D(filters=out_dim, kernel_size=3, padding="SAME", activation=tf.nn.relu)(identity_1)
identity_2 = tf.keras.layers.BatchNormalization()(identity_2)
for _ in range(4):
identity_2 = identity_block(identity_2,out_dim)
"""--------第三层----------"""
out_dim = 128
identity_3 = tf.keras.layers.Conv2D(filters=out_dim, kernel_size=3, padding="SAME", activation=tf.nn.relu)(identity_2)
identity_3 = tf.keras.layers.BatchNormalization()(identity_3)
for _ in range(6):
identity_3 = identity_block(identity_3,out_dim)
"""--------第四层----------"""
out_dim = 256
identity_4 = tf.keras.layers.Conv2D(filters=out_dim, kernel_size=3, padding="SAME", activation=tf.nn.relu)(identity_3)
identity_4 = tf.keras.layers.BatchNormalization()(identity_4)
for _ in range(3):
identity_4 = identity_block(identity_4,out_dim)
flat = tf.keras.layers.Flatten()(identity_4)
flat = tf.keras.layers.Dropout(0.217)(flat)
dense = tf.keras.layers.Dense(1024,activation=tf.nn.relu)(flat)
dense = tf.keras.layers.BatchNormalization()(dense)
logits = tf.keras.layers.Dense(10,activation=tf.nn.softmax)(dense)
model = tf.keras.Model(inputs=input_xs, outputs=logits)
return model
if __name__ == "__main__":
resnet_model = resnet_Model()
print(resnet_model.summary())#.2.4、使用ResNet50实战CIFAR10
下面是训练的代码：
import tensorflow as tf
import numpy as np
import getDataset
batch_size = 64
study_rate = 1e-2
x_train, y_train,x_test, y_test = getDataset.get_CIFAR100_dataset(root="C:/Users/thinkpad/SpyderProjects/ResnetDemo/cifar-10-python/cifar-10-batches-py/")
y_train = np.float32(tf.keras.utils.to_categorical(y_train,num_classes=10))
y_test = np.float32(tf.keras.utils.to_categorical(y_test,num_classes=10))
train_data = tf.data.Dataset.from_tensor_slices((x_train,y_train)).shuffle(batch_size*10).batch(batch_size).repeat(3)
import ResnetModel
model = ResnetModel.resnet_Model()
model.compile(optimizer=tf.optimizers.Adam(study_rate), loss=tf.losses.categorical_crossentropy,metrics = ['accuracy'])
model.fit(train_data, epochs=10)
score = model.evaluate(x_test, y_test)
print("last score:",score)
训练结果和测试结果：
Epoch 100/100
782/782 [] - 136s 173ms/step - loss: 0.0035 - accuracy: 0.9989
313/313 [] - 8s 24ms/step - loss: 2.0742 - accuracy: 0.7598
last score: [2.074207067489624, 0.7598000168800354]
在使用测试数据集进行测试时，正确率只有75%左右，而ResNet的文章中标明只有6%左右的错误率，这个差距有点大，不知道是什么原因？
有的人说测试数据集少，但是也有5000个样本
原文章没有使用dropout，我的代码中有用，而且BN层也有不少