在 CIFAR 10 数据集上使用 VGG16 进行迁移学习：训练和测试精度非常高但预测错误

Question

我使用迁移学习在cifar10数据集上训练了vgg16 模型。 它在一个时期后达到约89% 的训练准确率和约89% 的测试准确率。 但是，使用经过训练的模型来预测数据集以外的图像标签会给出错误的答案。 甚至错误地标记非常清晰的图像。

我尝试将 epochs 增加到 20，这将训练和测试的准确率提高到 93-94% 左右，并尝试了许多不同的图像。 经过训练的模型可以正确预测数据集中的图像，但在处理新图像时会遇到问题。

#!/usr/bin/env python
# coding: utf-8

# In[1]:

from keras.models import load_model
import numpy as np
from tqdm import tqdm
from keras import models
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras.applications.vgg16 import VGG16,preprocess_input
from keras.optimizers import Adam
from keras.models import Sequential, Model
from keras.layers import Dense, Flatten, GlobalAveragePooling2D
import pandas as pd
from keras.utils import np_utils
np.random.seed(123) 


# In[2]:


from keras.datasets import cifar10

(X_train, y_train), (X_test, y_test) = cifar10.load_data()


# In[3]:


print (X_train.shape)
print (X_test.shape)
#print (X_train[:2])

# In[4]:


from matplotlib import pyplot as plt


# In[5]:


X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255


# In[6]:


print (y_train.shape)
print (y_test[:10])



# In[7]:


Y_train = np_utils.to_categorical(y_train, 10)
Y_test = np_utils.to_categorical(y_test, 10)


# In[8]:



# In[9]:


import cv2


# In[24]:


train_set_x= X_train[:500]

train_set_y= Y_train[:500]
test_set_x= X_test[:100]
test_set_y= Y_test[:100]

# In[33]:

plt.imshow(X_test[1])
plt.show()

#train_set_y.shape


# In[27]:


frozen = VGG16 (weights="imagenet", input_shape=(32,32,3), include_top=False)


# In[28]:


frozen.summary()


# In[36]:


trainable = frozen.output
trainable = GlobalAveragePooling2D()(trainable)
#print(trainable.shape)
trainable = Dense(128, activation="relu")(trainable)
trainable = Dense(32, activation="relu")(trainable)
trainable = Dense(10, activation="softmax")(trainable)


# In[37]:


model = Model(inputs=frozen.input, outputs=trainable)


# In[38]:


model.summary()


# In[16]:


model.layers


# In[18]:


for layer in model.layers[:-4]:
    layer.trainable = False


# In[19]:


for layer in model.layers:
    print(layer, layer.trainable)


# In[40]:


learning_rate = 0.0001
opt = Adam(lr=learning_rate)
model.compile(optimizer=opt,
              loss='binary_crossentropy',
              metrics=['accuracy'])


# In[41]:


def evaluate_this_model(model, epochs):

    np.random.seed(1)

    history = model.fit(train_set_x, train_set_y, epochs=epochs)
    results = model.evaluate(test_set_x, test_set_y)

    plt.plot(np.squeeze(history.history["loss"]))
    plt.ylabel('cost')
    plt.xlabel('iterations (per tens)')
    plt.title("Learning rate =" + str(learning_rate))
    plt.show()

    print("\n\nAccuracy on training set is {}".format(history.history["acc"][-1]))
    print("\nAccuracy on test set is {}".format(results[1]))


# In[42]:


train_set_x.shape


evaluate_this_model(model, 1)
model.save('vgg16.h5')

model1=load_model('vgg16.h5')




IMG_SIZE=32
path1='../input/ship.png'
img_data1 = cv2.imread(path1, cv2.IMREAD_COLOR)
img_data1 = cv2.resize(img_data1, (IMG_SIZE, IMG_SIZE))
data1 = img_data1.reshape(-1, IMG_SIZE, IMG_SIZE, 3)
model_out=model1.predict(data1)

if np.argmax(model_out) == 1:
    str_label = 'Automobile'
    print(str_label)
if np.argmax(model_out) == 2:
    str_label = 'Bird'
    print(str_label)
if np.argmax(model_out) == 3:
    str_label = 'Cat'
    print(str_label)
if np.argmax(model_out) == 4:
    str_label = 'Deer'
    print(str_label)
if np.argmax(model_out) == 0:
    str_label = 'Airplane'
    print(str_label)
if np.argmax(model_out) == 5:
    str_label = 'Dog'
    print(str_label)
if np.argmax(model_out) == 6:
    str_label = 'Frog'
    print(str_label)
if np.argmax(model_out) == 7:
    str_label = 'Horse'
    print(str_label)
if np.argmax(model_out) == 8:
    str_label = 'Ship'
    print(str_label)
if np.argmax(model_out) == 9:
    str_label = 'Truck'
    print(str_label)

经过训练的模型即使在一个时期后也能在数据集图像上正确预测和标记，但在处理新图像时会遇到问题，它会完全给出错误的标签。 例如：它将非常清晰的船图像标记为鹿。 其他班级也一样。

Answer 1

看起来您正在通过除以 255 来缩放训练和测试数据的颜色。我没有看到 ship.png 发生这种情况。 我建议创建一个函数来完成所有预处理，并确保运行它进行训练、测试和预测，这样您就可以确保对所有图像应用完全相同的清理。