Getting train test data from Keras ImageDataGenerator

Question

I am using ImageDataGenerator from Keras as follows.

datagen = ImageDataGenerator(samplewise_center=True,
    samplewise_std_normalization=True, validation_split=0.30
                               )

Then .flow statement to obtain train and test split as follows.

train_iterator = datagen.flow(x, y, subset='training')
test_iterator = datagen.flow(x, y, subset='validation')

Here x represents images with a shape (588, 120, 120, 1) and y represents multiclass output (588, 4) .

In (588, 120, 120, 1) shape input data, there are total 588 samples each with a shape of (120, 120, 1) . The output is having 4 classes.

Then I train and test my CNN with the following code.

history =model.fit_generator(train_iterator,
                             
                              epochs=10,
                              validation_data=test_iterator,
                              callbacks=callbacks_list) 


pred_test = model.predict(test_iterator, steps=len(test_iterator), verbose=0)

My question is: How can I access the test data (both x and y) which test_iterator uses for prediction.

Answer 1

flow()返回产生 (x, y) 元组的迭代器，您可以使用test_iterator.next()访问元素。

Answer 2

Just an FYI. In the ImageDataGenerator you have set samplewise_center=True, samplewise_std_normalization=True,. If this is what you want you must FIRST fit the generator to accumulate the statistics of the input data so first do

datagen.fit(x)

Documentation is here.