为什么我的模型会预测相同的标签? [英] Why does my model predict the same label?

问题描述

我正在训练一个小型网络，训练似乎进行得很顺利，val损失减少，我的验证精度达到80，并且在没有更多改善时实际上停止了训练(耐心= 10).它训练了40个纪元.但是，它始终为每个测试图像仅预测一个类！我尝试随机初始化conv层，添加了正则化器，从Adam切换到SGD，添加了clipvalue，添加了辍学.我也切换到softmax(我只有两个标签，但是我看到了一些有关使用带有2个神经元的softmax和Dense层的建议).其中一些或其中一项有助于过度拟合，但对于预测问题却无济于事. 数据是均衡的，尽管它是一个很小的数据集，所以如果它也为评估集预测相同的标签，则达到80％并不有意义.

I am training a small network and the training seems to go fine, the val loss decreases, I reach validation accuracy around 80, and it actually stops training once there is no more improvement (patience=10). It trained for 40 epochs. However, it keeps predicting only one class for every test image! I tried to initialize the conv layers randomly, I added regularizers, I switched from Adam to SGD, I added clipvalue, I added dropouts. I also switched to softmax (I have only two labels but I saw some recommendation on using softmax and Dense layer with 2 neurons). Some or one of these helped with the overfitting, but nothing worked for the prediction problem. The data is balanced, though it is a small dataset, so it doesn't make sense that it reaches 80% if it predicts the same labels for evaluation set as well.

我的模型有什么问题，我该如何解决?欢迎任何评论.

What is wrong with my model and how can I fix it? Any comments are welcome.

#Import some packages to use
import cv2
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from keras.preprocessing.image import ImageDataGenerator
import os
from keras.regularizers import l2
from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
from keras.layers.core import Dense, Dropout, Flatten
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.initializers import RandomNormal

os.environ["CUDA_VISIBLE_DEVICES"]="0"

epochs = 200
callbacks = []
#schedule = None
decay = 0.0

earlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')
mcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')
reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')

train_dir = '/home/d/Desktop/s/data/train'
eval_dir = '/home/d/Desktop/s/data/eval'
test_dir = '/home/d/Desktop/s/data/test'



# create a data generator
train_datagen = ImageDataGenerator(rescale=1./255,   #Scale the image between 0 and 1
                                    rotation_range=40,
                                    width_shift_range=0.2,
                                    height_shift_range=0.2,
                                    shear_range=0.2,
                                    zoom_range=0.2,
                                    horizontal_flip=True,)

val_datagen = ImageDataGenerator(rescale=1./255)  #We do not augment validation data. we only perform rescale

test_datagen = ImageDataGenerator(rescale=1./255)  #We do not augment validation data. we only perform rescale

# load and iterate training dataset
train_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='categorical', batch_size=16, shuffle='True', seed=42)
# load and iterate validation dataset
val_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='categorical', batch_size=16, shuffle='True', seed=42)
# load and iterate test dataset
test_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)
#We will use a batch size of 32. Note: batch size should be a factor of 2.***4,8,16,32,64...***
#batch_size = 4



#from keras import layers
from keras import models
from keras import optimizers
#from keras.layers import Dropout
#from keras.preprocessing.image import ImageDataGenerator
from keras.preprocessing.image import img_to_array, load_img

model = models.Sequential()
model.add(Conv2D(64, (3, 3), activation='relu', name='block1_conv1', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05), input_shape=(224, 224, 3)))
model.add(Conv2D(64, (3, 3), activation='relu', name='block1_conv2', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(MaxPooling2D((2, 2)))
model.add(Dropout(0.2))
model.add(Conv2D(128, (3, 3), activation='relu', name='block2_conv1', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(Conv2D(128, (3, 3), activation='relu', name='block2_conv2',kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(MaxPooling2D((2, 2), name='block2_pool'))
model.add(Dropout(0.2))
model.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv1', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv2', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(Conv2D(256, (3, 3), activation='relu', name='block3_conv3', kernel_initializer=RandomNormal(
        mean=0.0, stddev=0.05), bias_initializer=RandomNormal(mean=0.0, stddev=0.05)))
model.add(MaxPooling2D((2, 2), name='block3_pool'))
model.add(Dropout(0.2))
#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv1'))
#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv2'))
#model.add(layers.Conv2D(512, (3, 3), activation='relu', name='block4_conv3'))
#model.add(layers.MaxPooling2D((2, 2), name='block4_pool'))
model.add(Flatten())
model.add(Dense(256, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='relu', kernel_initializer='he_uniform'))
model.add(Dropout(0.5))
model.add(Dense(2, kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01), activation='softmax'))

#Lets see our model
model.summary()

#We'll use the RMSprop optimizer with a learning rate of 0.0001
#We'll use binary_crossentropy loss because its a binary classification
#model.compile(loss='binary_crossentropy', optimizer=optimizers.SGD(lr=1e-5, momentum=0.9), metrics=['acc'])
model.compile(loss='categorical_crossentropy',
                   #optimizer=optimizers.Adadelta(lr=1.0, rho=0.95, epsilon=1e-08, decay=decay),
                    optimizer=optimizers.SGD(lr= 0.0001, clipvalue = 0.5, decay=1e-6, momentum=0.9, nesterov=True),
              metrics=['accuracy'])



#The training part
#We train for 64 epochs with about 100 steps per epoch
history = model.fit_generator(train_generator,
                              steps_per_epoch=train_generator.n // train_generator.batch_size,
                              epochs=epochs,
                              validation_data=val_generator,
                              validation_steps=val_generator.n // val_generator.batch_size,
                              callbacks=[earlyStopping, mcp_save]) #, reduce_lr_loss])


#Save the model
model.save_weights('/home/d/Desktop/s/categorical_weights.h5')
model.save('/home/d/Desktop/s/categorical_model_keras.h5')

#lets plot the train and val curve
#get the details form the history object
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(acc) + 1)

#Train and validation accuracy
plt.plot(epochs, acc, 'b', label='Training accuracy')
plt.plot(epochs, val_acc, 'r', label='Validation accuracy')
plt.title('Training and Validation accuracy')
plt.legend()

plt.figure()
#Train and validation loss
plt.plot(epochs, loss, 'b', label='Training loss')
plt.plot(epochs, val_loss, 'r', label='Validation loss')
plt.title('Training and Validation loss')
plt.legend()

plt.show()

model.evaluate_generator(generator=val_generator, steps=val_generator.n // val_generator.batch_size)

STEP_SIZE_TEST=test_generator.n//test_generator.batch_size
test_generator.reset()
pred=model.predict_generator(test_generator,
steps=STEP_SIZE_TEST,
verbose=1)

predicted_class_indices=np.argmax(pred,axis=1)

labels = (train_generator.class_indices)
np.save('/home/d/Desktop/s/classes', labels)

labels = dict((v,k) for k,v in labels.items())
predictions = [labels[k] for k in predicted_class_indices]

filenames=test_generator.filenames
results=pd.DataFrame({"Filename":filenames,
                      "Predictions":predictions})
results.to_csv("categorical_results.csv",index=False)

推荐答案

可能导致此类行为的问题之一是数据集不平衡.您的模型发现，如果每次都能预测到主导阶层，那将会得到很好的结果.

One of the problems that could lead to such behavior is imbalanced dataset. Your model found out that if it predicts the dominant class each time, it would get a good results.

有许多方法可以解决不平衡数据集.这是一个很好的教程. 最简单但功能强大的解决方案之一是，如果错误地预测了较小的类别，则对您的损失施加更高的罚款.可以通过在fit或fit_generator函数中设置参数class_weight在keras中实现.

There are many ways to tackle an imbalance dataset. Here is a good tutorial. One of the easiest yet powerful solution is to apply higher penalty to your loss if it wrongly predicted the smaller class. This can be implemented in keras by setting the parameter class_weight in the fitor fit_generator function.

它可以是示例字典:

class_weight = {0: 0.75, 1: 0.25}  # does not necessarily add to up 1.
history = model.fit_generator(train_generator,
                              steps_per_epoch=train_generator.n // train_generator.batch_size,
                              epochs=epochs,
                              class_weight= class_weight,  # this is the important part
                              validation_data=val_generator,
                              validation_steps=val_generator.n // val_generator.batch_size,
                              callbacks=[earlyStopping, mcp_save]) #, reduce_lr_loss])

这篇关于为什么我的模型会预测相同的标签?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！