具有3D CNN数据格式错误的Keras ImageDataGenerator流目录？

Question

我是Keras的新手，我有一个包含多个文件夹的数据集，每个文件夹都指向一个特定的类。我使用ImageDataGenerator从子文件夹中读取数据。我试图使用大小为80x100的16个连续帧，因此input_shape为（16,80,100,1）。当我进行训练时，有一个关于网络输入的错误，我知道输入应该是3D CNN的5d张量，但我不确定我是否正确这样做。

我正在使用spyder来编写和实现代码：

from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten, Conv3D, MaxPooling3D
from keras.layers.advanced_activations import LeakyReLU
from keras.optimizers import SGD, RMSprop
from keras.utils import np_utils, generic_utils
from keras.losses import categorical_crossentropy
from keras.optimizers import Adam
import os
import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import cv2
from sklearn.cross_validation import train_test_split
from sklearn import cross_validation
from sklearn import preprocessing

datagen = ImageDataGenerator()
train_data=datagen.flow_from_directory('C:\\Users\\AA\\Data\\Training', target_size=(80, 100), color_mode='grayscale', classes=None, class_mode='categorical', batch_size=32, interpolation='nearest')
test_data=datagen.flow_from_directory('C:\\Users\\AA\\Data\\Testing', target_size=(80, 100), color_mode='grayscale', classes=None, class_mode='categorical', batch_size=32, interpolation='nearest')



    ins = (16, 80, 100, 1)
    model = Sequential()
    model.add(Conv3D(32, kernel_size=(3, 3, 3), input_shape=ins, border_mode='same'))
    model.add(Activation('relu'))
    model.add(Conv3D(32, kernel_size=(3, 3, 3), border_mode='same'))
    model.add(Activation('softmax'))
    model.add(MaxPooling3D(pool_size=(3, 3, 3), border_mode='same'))
    model.add(Dropout(0.25))

    model.add(Conv3D(64, kernel_size=(3, 3, 3), border_mode='same'))
    model.add(Activation('relu'))
    model.add(Conv3D(64, kernel_size=(3, 3, 3), border_mode='same'))
    model.add(Activation('softmax'))
    model.add(MaxPooling3D(pool_size=(3, 3, 3), border_mode='same'))
    model.add(Dropout(0.25))

    model.add(Flatten())
    model.add(Dense(512, activation='sigmoid'))
    model.add(Dropout(0.5))
    model.add(Dense(8, activation='softmax'))


 model.compile(loss=categorical_crossentropy, optimizer=Adam(), metrics=['accuracy'])    


   model.fit_generator(train_data,
        steps_per_epoch=2000,
        epochs=50,
        validation_data=test_data,
        validation_steps=800)

错误说：

File "C:\Users\AA\AppData\Local\Continuum\Anaconda3\lib\site-packages\keras\engine\training.py", line 113, in _standardize_input_data
    'with shape ' + str(data_shape))

ValueError: Error when checking input: expected conv3d_24_input to have 5 dimensions, but got array with shape (32, 80, 100, 1)

Answer 1

我认为问题是ImageDataGenerator。

它只适用于图像，而不适用于视频（我得到了相同的错误，在https://github.com/keras-team/keras/issues/10150，他们也声称ImageDataGenerator只适用于图像形状的张量。在那里他们还建议实现你自己的数据生成，如https://stanford.edu/~shervine/blog/keras-how-to-generate-data-on-the-fly.html中描述的，但我我自己没试过）

Answer 2

https://gist.github.com/Emadeldeen-24/736c33ac2af0c00cc48810ad62e1f54a

这是一个自定义imagedatagenerator，用于Conv3D网络的5D输入。希望能帮助到你。

from tweaked_ImageGenerator_v2 import ImageDataGenerator
datagen = ImageDataGenerator()
train_data=datagen.flow_from_directory('path/to/data', target_size=(x, y), batch_size=32, frames_per_step=4)

Answer 3

该模型将您的输入解释为16个80x100灰度图像样本。您必须将输入重新整形为

(no_of_samples,16,80,100,1)

这里有16个是你的时间步