从TFLearn转换代码到Keras工作

Question

我使用TFLearn编写的一些代码作为参考，并尝试使用Keras重新编写代码。我对这两个软件包都比较陌生，但我不确定自己是否正确编写了它。

我已经试过我的代码 - 它的工作原理 - 但我没有得到预期的结果（准确度没有提高20多个时代），我想知道我是否在某个地方犯了错误。

就我的数据而言，我有一个'数据'目录，其中有'训练'和'验证'目录。每个内部都有3个目录，分别用于我的3个图像类。

原始TFLearn代码：

import numpy as np 
import tflearn 
from tflearn.layers.conv import conv_2d, max_pool_2d 
from tflearn.layers.core import input_data, dropout, fully_connected 
from tflearn.layers.estimator import regression 

def createModel(nbClasses,imageSize): 

    convnet = input_data(shape=[None, imageSize, imageSize, 1], name='input') 

    convnet = conv_2d(convnet, 64, 2, activation='elu', weights_init="Xavier") 
    convnet = max_pool_2d(convnet, 2) 

    convnet = conv_2d(convnet, 128, 2, activation='elu', weights_init="Xavier") 
    convnet = max_pool_2d(convnet, 2) 

    convnet = conv_2d(convnet, 256, 2, activation='elu', weights_init="Xavier") 
    convnet = max_pool_2d(convnet, 2) 

    convnet = conv_2d(convnet, 512, 2, activation='elu', weights_init="Xavier") 
    convnet = max_pool_2d(convnet, 2) 

    convnet = fully_connected(convnet, 1024, activation='elu') 
    convnet = dropout(convnet, 0.5) 

    convnet = fully_connected(convnet, nbClasses, activation='softmax') 
    convnet = regression(convnet, optimizer='rmsprop', loss='categorical_crossentropy') 

    model = tflearn.DNN(convnet) 
return model 

我使用Keras代码：

from keras import backend as K 
from keras.layers.core import Flatten, Dense, Dropout, Activation 
from keras.optimizers import rmsprop 
from keras.models import Sequential 
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img 
from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D 
import numpy as np 

num_classes = 3 
image_size = 256 
nb_epoch = 80 
batch_size = 32 
nb_train_samples = 7994 
nb_validation_samples = 2000 

if K.image_data_format() == 'channels_first': 
    input_shape = (3, image_size, image_size) 
else: 
    input_shape = (image_size, image_size, 3) 

model = Sequential() 

model.add(ZeroPadding2D((1,1), input_shape=input_shape)) 

model.add(Conv2D(64, 2, activation='elu', kernel_initializer='glorot_normal')) 
model.add(MaxPooling2D((2, 2))) 

model.add(Conv2D(128, 2, activation='elu', kernel_initializer='glorot_normal')) 
model.add(MaxPooling2D((2, 2))) 

model.add(Conv2D(256, 2, activation='elu', kernel_initializer='glorot_normal')) 
model.add(MaxPooling2D((2, 2))) 

model.add(Conv2D(512, 2, activation='elu', kernel_initializer='glorot_normal')) 
model.add(MaxPooling2D((2, 2))) 

model.add(Flatten()) 
model.add(Dense(1024)) 
model.add(Activation('elu')) 
model.add(Dropout(0.5)) 

model.add(Dense(num_classes)) 
model.add(Activation('softmax')) 
opt = rmsprop() 
model.compile(loss='categorical_crossentropy', 
     optimizer = opt, 
     metrics = ['accuracy']) 

train_data_dir = 'data/train' 
validation_data_dir = 'data/validation' 
train_datagen = ImageDataGenerator(rescale= 1./255) 
validation_datagen = ImageDataGenerator(rescale=1./255) 

train_generator = train_datagen.flow_from_directory(
    train_data_dir, 
    target_size=(image_size, image_size), 
    batch_size=batch_size, 
    class_mode='categorical' 
    ) 

validation_generator = validation_datagen.flow_from_directory(
    validation_data_dir, 
    target_size=(image_size, image_size), 
    batch_size=batch_size, 
    class_mode='categorical' 
    ) 



model.fit_generator(train_generator, 
       steps_per_epoch=(nb_train_samples // batch_size), 
       epochs=nb_epoch, 
       validation_data=validation_generator, 
       validation_steps=(nb_validation_samples // batch_size) 
       ) 

model.save_weights('first_try.h5') 

Answer 1

我想你的代码有三个班MNIST数据集，并可以训练就好了。如预期的那样，准确度在第一时期增加。

至少对于MNIST来说，我可以通过仅使用前两个Conv图层和64的密集图层来训练得更快。根据您的数据，我建议您尝试使用更简单的模型（即2个Conv图层），检查模型正在学习，然后从那里改进。