python - 使用 CNN 的时尚 MNIST，过度拟合？

Question

cnn_model = Sequential()

cnn_model.add(Conv2D(64,3, 3, input_shape = (28,28,1), activation='relu'))
cnn_model.add(MaxPooling2D(pool_size = (2, 2)))

cnn_model.add(Dropout(0.25))

cnn_model.add(Flatten())
cnn_model.add(Dense(output_dim = 32, activation = 'relu'))
cnn_model.add(Dense(output_dim = 10, activation = 'sigmoid'))

cnn_model.compile(loss ='sparse_categorical_crossentropy', optimizer=Adam(lr=0.001),metrics =['accuracy'])

epochs = 50

history = cnn_model.fit(X_train,
                        y_train,
                        batch_size = 512,
                        nb_epoch = epochs,
                        verbose = 1,
                        validation_data = (X_validate, y_validate))

最后我得到了这个结果:

Epoch 50/50
48000/48000 [==============================] - 35s 728us/step - loss: 0.1265 - accuracy: 0.9537 - val_loss: 0.2425 - val_accuracy: 0.9167
training loss=0.125 ,validation loss=0.2425
training accuracy=95.3% ,validation accuracy=91.67

我的问题如下:

1. 模型是过拟合还是欠拟合？
2. 我应该增加吗？时代？

Graph of Losses

Answer 1

由于模型过拟合，你可以

1. Shuffle 数据，通过在 cnn_model.fit 中使用 shuffle=True。
2. 使用提前停止
3. 使用正则化

完整代码，使用与您相同的架构，减少过拟合，如下所示。损失有所增加，但我们可以通过添加更多的卷积层和池化层来改进它。

# To support both python 2 and python 3
from __future__ import division, print_function, unicode_literals
from io import open

# Common imports
import numpy as np
import os
import tensorflow as tf
from tensorflow.keras.layers import Input, Conv2D, MaxPool2D, Dense, Dropout, Flatten
from tensorflow.keras.models import Sequential
from tensorflow.keras.optimizers import Adam
import matplotlib.pyplot as plt
from tensorflow.keras.regularizers import l2
from matplotlib.pyplot import axis as ax

tf.__version__

X = tf.placeholder(tf.float32, shape=[None, 784], name="X")
X_reshaped = tf.reshape(X, shape=[-1, 28, 28, 1])
y = tf.placeholder(tf.int32, shape=[None], name="y")

# instantiate regularizer
Regularizer = l2(0.001)

cnn_model = Sequential()

cnn_model.add(Conv2D(64,3, 3, input_shape = (28,28,1), activation='relu', data_format='channels_last', 
                    activity_regularizer=Regularizer, kernel_regularizer=Regularizer))

cnn_model.add(MaxPool2D(pool_size = (2, 2)))

cnn_model.add(Dropout(0.25))

cnn_model.add(Flatten())

cnn_model.add(Dense(units = 32, activation = 'relu', 
                    activity_regularizer=Regularizer, kernel_regularizer=Regularizer))

cnn_model.add(Dense(units = 10, activation = 'sigmoid', 
                    activity_regularizer=Regularizer, kernel_regularizer=Regularizer))

cnn_model.compile(loss ='sparse_categorical_crossentropy', optimizer=Adam(lr=0.001),metrics =['accuracy'])

epochs = 50

cnn_model.summary()

(X_train, y_train), (X_test, y_test) = tf.keras.datasets.fashion_mnist.load_data()
X_train = X_train.astype(np.float32).reshape(-1, 28*28) / 255.0
X_train_reshaped = tf.reshape(X_train, shape=[-1, 28, 28, 1])
X_test = X_test.astype(np.float32).reshape(-1, 28*28) / 255.0
X_test_reshaped = tf.reshape(X_test, shape=[-1, 28, 28, 1])

y_train = tf.cast(y_train, dtype = tf.int32)

y_test = tf.cast(y_test, dtype = tf.int32)

def shuffle_batch(X, y, batch_size):
    rnd_idx = np.random.permutation(len(X))
    n_batches = len(X) // batch_size
    for batch_idx in np.array_split(rnd_idx, n_batches):
        X_batch, y_batch = X[batch_idx], y[batch_idx]
        yield X_batch, y_batch

#steps_per_epoch = X_train_reshaped.shape[0]//512
steps_per_epoch = X_train_reshaped.shape[0].value//512

epochs = 50

callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=15)

history = cnn_model.fit(x = X_train_reshaped,
                        y = y_train,
                        batch_size = 512,
                        epochs = epochs, callbacks=[callback],
                        verbose = 1, validation_data = (X_test_reshaped, y_test),
                        validation_steps = 10, steps_per_epoch=steps_per_epoch, shuffle = True)

print(history.history.keys())

#  "Accuracy"
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
# "Loss"
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
axes = plt.axes()
axes.set_ylim([0.2, 1.5])
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()