Question

我有一个图像分类任务，有一个庞大的数据集（800,000个标记数据，80,000个未标记数据和10个要分类的类）我已经建立了模型，但是训练过程非常耗时，所以我想要要知道我的模型是否足够强大以完成任务：

img_width, img_height = 150, 150
batch_size = 64
samples_per_epoch = 671429
nb_val_samples = 167852
nb_filters1 = 32
nb_filters2 = 64
nb_filters3 = 128
conv_size = 3
pool_size = 2
classes_num = 10

# Define class weight to handel imbalanced classes
def get_class_weights(counts, smooth_factor=0):

    if smooth_factor > 0:
        p = max(counts) * smooth_factor
        for k in counts:
            k += p

    majority = max(counts)

    return {cls: float(majority / count) for cls, count in enumerate(counts)}

class_weight = get_class_weights([37415, 9344, 128127, 80545, 60559, 64548, 31029, 105297, 309151, 13266], 0.1)

# Initialising the CNN
classifier = Sequential()

# Step 1 - Convolution
classifier.add(Convolution2D(nb_filters1, conv_size, conv_size, input_shape = (img_width, img_height, 3), activation = 'relu', border_mode='same'))

# Step 2 - Pooling
classifier.add(MaxPooling2D(pool_size = (pool_size, pool_size)))

# Adding a second convolutional layer
classifier.add(Convolution2D(nb_filters2, conv_size, conv_size, activation = 'relu', border_mode='same'))
classifier.add(MaxPooling2D(pool_size = (pool_size, pool_size)))

# Adding a third convolutional layer
classifier.add(Convolution2D(nb_filters3, conv_size, conv_size, activation = 'relu', border_mode='same'))
classifier.add(MaxPooling2D(pool_size = (pool_size, pool_size)))

# Step 3 - Flattening
classifier.add(Flatten())

# Step 4 - Full connection
classifier.add(Dense(output_dim = 1024, activation = 'relu'))
classifier.add(Dropout(0.2))

classifier.add(Dense(output_dim = 1024, activation = 'relu'))
classifier.add(Dropout(0.2))

classifier.add(Dense(output_dim = classes_num, activation = 'softmax'))

# Compiling the CNN
classifier.compile(optimizer = 'RMSprop', loss = 'categorical_crossentropy', metrics = ['accuracy'])

使用Keras，构建CNN模型以对图像进行分类

0 个答案: