为什么在 Keras 训练期间由 model.evaluate() 计算的指标与跟踪的指标不同？

Question

我正在使用 Keras 2.0.4（TensorFlow 后端）执行图像分类任务（基于预训练模型）。 在训练/调整期间，我使用CSVLogger 跟踪所有使用的指标（例如categorical_accuracy、categorical crossentropy）——包括与验证集相关联的相应指标（即val_categorical_accuracy、val_categorical_crossentropy）。

通过回调ModelCheckpoint，我正在跟踪权重的最佳配置（save_best_only=True）。为了评估验证集上的模型，我使用model.evaluate()。

我的期望是：CSVLogger（“最佳”时期）跟踪的指标等于model.evaluate() 计算的指标。 不幸的是，这种情况并非如此。指标相差 +- 5%。 这种行为有原因吗？

---

编辑：

经过一些测试，我可以获得一些见解：

1. 如果我不使用生成器来生成训练和验证数据（因此没有model.fit_generator()），问题就不会发生。 --> 使用ImageDataGenerator 进行训练和验证数据是差异的来源。 （请注意，对于 evaluate 的计算，我不使用生成器，但我确实使用相同的验证数据（至少如果 DataImageGenerator 可以作为预计...）。
   我认为， ImageDataGenerator 不能正常工作（请， 也可以看看this)。
2. 如果我根本不使用生成器，就不会有这个问题。 Id est 由CSVLogger（“最佳”时期）跟踪的指标等于model.evaluate() 计算的指标。
   有趣的是，还有另一个问题：如果您使用相同的数据进行训练和验证，那么在每个 epoch 结束时，训练指标（例如 loss）和验证指标（例如 val_loss）之间会存在差异。
   (A similar problem)

使用代码：

############################ import section ############################
from __future__ import print_function # perform like in python 3.x
from keras.datasets import mnist
from keras.utils import np_utils # numpy utils for to_categorical()
from keras.models import Model, load_model
from keras.layers import Dense, GlobalAveragePooling2D, Dropout, GaussianDropout, Conv2D, MaxPooling2D
from keras.optimizers import SGD, Adam
from keras import backend as K
from keras.preprocessing.image import ImageDataGenerator 
from keras import metrics
import os
import sys
from scipy import misc
import numpy as np
from keras.applications.vgg16 import preprocess_input as vgg16_preprocess_input
from keras.applications import VGG16
from keras.callbacks import CSVLogger, ModelCheckpoint


############################ manual settings ###########################
# general settings
seed = 1337

loss_function = 'categorical_crossentropy'

learning_rate = 0.001

epochs = 10

batch_size = 20

nb_classes = 5 

img_width, img_height = 400, 400 # >= 48 necessary, as VGG16 is used

chosen_optimizer = SGD(lr=learning_rate, momentum=0.0, decay=0.0, nesterov=False)

steps_per_epoch = 40 // batch_size  # 40 train samples in 5 classes
validation_steps = 40 // batch_size # 40 train samples in 5 classes

data_dir = # TODO: set path where data is stored (folders: 'train', 'val', 'test'; within each folder are folders named by classes)

# callbacks: CSVLogger & ModelCheckpoint
filepath = # TODO: set path, where you want to store files generated by the callbacks
file_best_checkpoint= 'best_epoch.hdf5'
file_csvlogger = 'logged_metrics.txt'

modelcheckpoint_best_epoch= ModelCheckpoint(filepath=os.path.join(filepath, file_best_checkpoint), 
                                  monitor = 'val_loss' , verbose = 1, 
                                  save_best_only = True, 
                                  save_weights_only=False, mode='auto', 
                                  period=1) # every epoch executed
csvlogger = CSVLogger(os.path.join(filepath, file_csvlogger) , separator=',', append=False)



############################ prepare data ##############################
# get validation data (for evaluation)
X_val, Y_val = # TODO: load train data (4darray, samples, img_width, img_height, nb_channels) IMPORTANT: 5 classes with 8 images each.

# preprocess data
my_preprocessing_function = mf.my_vgg16_preprocess_input

# 'augmentation' configuration we will use for training
train_datagen = ImageDataGenerator(preprocessing_function = my_preprocessing_function) # only preprocessing; static data set

# 'augmentation' configuration we will use for validation
val_datagen = ImageDataGenerator(preprocessing_function = my_preprocessing_function) # only preprocessing; static data set

train_data_dir = os.path.join(data_dir, 'train')
validation_data_dir = os.path.join(data_dir, 'val')
train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    shuffle = True,
    seed = seed, # random seed for shuffling and transformations
    class_mode='categorical')  # label type (categorical = one-hot vector)

validation_generator = val_datagen.flow_from_directory(
    validation_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    shuffle = True,
    seed = seed, # random seed for shuffling and transformations
    class_mode='categorical')  # label type (categorical = one-hot vector)



############################## training ###############################
print("\n---------------------------------------------------------------")
print("------------------------ training model -----------------------")
print("---------------------------------------------------------------")
# create the base pre-trained model
base_model = VGG16(include_top=False, weights = None, input_shape=(img_width, img_height, 3), pooling = 'max', classes = nb_classes)
model_name =  "VGG_modified"

# do not freeze any layers --> all layers trainable
for layer in base_model.layers:
    layer.trainable = True

# define topping of base_model
x = base_model.output # get the last layer of our base_model
x = Dense(1024, activation='relu', name='fc1')(x)
x = Dense(1024, activation='relu', name='fc2')(x)
predictions = Dense(nb_classes, activation='softmax', name='predictions')(x)

# finally, stack model together
model = Model(outputs=predictions, name= model_name, inputs=base_model.input) #Keras 1.x.x: model = Model(input=base_model.input, output=predictions) 
print(model.summary())

# compile the model (should be done *after* setting layers to non-trainable)
model.compile(optimizer = chosen_optimizer, loss=loss_function, 
            metrics=['categorical_accuracy','kullback_leibler_divergence'])

# train the model on your data
model.fit_generator(
    train_generator,
    steps_per_epoch=steps_per_epoch,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=validation_steps,
    callbacks = [csvlogger, modelcheckpoint_best_epoch])



############################## evaluation ##############################
print("\n\n---------------------------------------------------------------")
print("------------------ Evaluation of Best Epoch -------------------")
print("---------------------------------------------------------------")
# load model (corresponding to best training epoch)
model = load_model(os.path.join(filepath, file_best_checkpoint))

# evaluate model on validation data (in test mode!)
list_of_metrics = model.evaluate(X_val, Y_val, batch_size=batch_size, verbose=1, sample_weight=None)
index = 0
print('\nMetrics:')
for metric in model.metrics_names:
    print(metric+ ':' , str(list_of_metrics[index]))
    index += 1

---

E D I T 2
参考 E D I T 的 1. 如果我在训练和评估期间对验证数据使用相同的生成器（使用evaluate_generator()），问题仍然存在。 因此，这肯定是由生成器引起的问题......

Answer 1

这仅适用于验证数据集上的指标评估。

在训练期间在训练数据集上计算的指标并不反映模型在 epoch 结束时的真实指标，因为模型将在每个批次中更新（修改）。

这有帮助吗？