Keras: How to use data (image) augmentation with fit_generator()?

[pengpaiSH]

@fchollet and other kerasors,
We all know that if we are dealing with large scale of data such as ImageNet, we could write a customized generator which produces batch data (often as numpy.array) from disk. Then we could train our model with model.fit_generator(). But, if we want to use ImageDataGenerator to do the online data augmentation at the same time, what is the simplest way to implement? Note that I would like to use its flow() method instead of flow_from_directory() method.

[tetmin]

It's possible to do this by doing the following:

datagen = ImageDataGenerator(...)
train_generator = datagen.flow(X_train, Y_train, batch_size=128)
model.fit_generator(train_generator, samples_per_epoch=len(X_train), ...)

What I'd like to know is how it's possible to combine ImageDataGenerator with a custom thread-safe data augmentation generator and flow from these?

[pengpaiSH]

@tetmin You are assuming that X_train and Y_train could be loaded into memory at once. My concern is that if not, then we have to write a customized generator which loads a batch of training data from disk. And how to combine the generator with online data augmentation.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.

[PeterStrom]

I too want to know. I have written my own batch generator that read batches from disk (all images do not fit in memory), and want to combine this with the ImageDataGenerator.

def batch_generator(df, batch_size, path_tiles, num_classes):
    """This generator use a pandas DataFrame to read images (df.tile_name) from disk.
    """
    N = df.shape[0]
    while True:
        for start in range(0, N, batch_size):
            x_batch = []
            y_batch = []
            end = min(start + batch_size, N)
            df_tmp = df[start:end]
            ids_batch = df_tmp.tile_name
            for id in ids_batch:
                img = cv2.imread(path_tiles+'/{}'.format(id))
                # [0] since duplicated names
                labelname=df_tmp['y'][df_tmp.tile_name == id].values[0]  
                labelname=np.asscalar(labelname)
                x_batch.append(img)
                y_batch.append(labelname)
            x_batch = np.array(x_batch, np.float32) / 255
            y_batch = utils.np_utils.to_categorical(y_batch, num_classes) 
            yield (x_batch, y_batch)

model.fit_generator(generator=batch_generator(df_train, 
                                              batch_size=batch_size,
                                              path_tiles=path_tiles,
                                              num_classes=num_classes), 
                    steps_per_epoch=len(df_train) // batch_size, 
                    epochs=epochs)

[harischeong]

I just put this in the training loop...

`

first load your batch from disk using custom functions

x= load_img (data_path,batch_numbatchSize,(batch_num+1)batchSize)
x = preprocess_img(x)

then augment your batch

for x in img_datagen.flow(x, batch_size=batchSize, seed =seed):
x=x
break

this generates one batch of augmented data

`

[was84san]

Is there any suggestion for this issue. I have the same of this issue