Support python generators

I have a simple task to find the best CNN architecture for image regression. However, I have a large dataset, which cannot be loaded into memory at one time. It seems in the current release ImageRegressor only supports fit method requiring all the data (x and y) loaded in memory. How can I use generator in Autokeras? I have checked a closed issue #204, but it seems it was not solved.

I have already tried the tf.dataset by converting my generator to tf.dataset, but it didn’t work. For example,

    dataset = tf.data.Dataset.from_generator(generate_batch, (tf.float32, tf.float32))
    vq_predictor = ak.ImageRegressor()
    for i, (X, y) in enumerate(dataset):
        X_dataset = tf.data.Dataset.from_tensors(X)
        y_dataset = tf.data.Dataset.from_tensors(y)
        vq_predictor.fit(X_dataset, y_dataset, validation_split=0.2)

Then I got error:

File “C:\Users\junyong\AppData\Local\Continuum\anaconda3\envs\tensorflow2\lib\site-packages\autokeras\tasks\image.py”, line 222, in fit **kwargs) File “C:\Users\junyong\AppData\Local\Continuum\anaconda3\envs\tensorflow2\lib\site-packages\autokeras\auto_model.py”, line 231, in fit validation_split=validation_split) File “C:\Users\junyong\AppData\Local\Continuum\anaconda3\envs\tensorflow2\lib\site-packages\autokeras\auto_model.py”, line 313, in _prepare_data dataset, validation_data = utils.split_dataset(dataset, validation_split) File “C:\Users\junyong\AppData\Local\Continuum\anaconda3\envs\tensorflow2\lib\site-packages\autokeras\utils.py”, line 69, in split_dataset raise ValueError('The dataset should at least contain 2 ’ ValueError: The dataset should at least contain 2 instances to be split.

Any suggestions are highly appreciated.

[theGOTOguy]

@VictorReaver1999 I was also getting the “Cannot take the length of shape with unknown rank” error. The issue is that autokeras.utils.data_utils.batched doesn’t know what to do with the generator. Since the generators in the example are indeed outputting batches, just monkey-patching this function to return True will get you up and running for now.

train_dataset = tf.data.Dataset.from_generator(lambda: callable_iterator(train_generator),output_types=(tf.float32, tf.float32))
val_dataset = tf.data.Dataset.from_generator(lambda: callable_iterator(val_generator),output_types=(tf.float32, tf.float32))

# Autokeras data_utils gets confused by the generator.
# Just let it know that the data is indeed batched.
ak.utils.data_utils.batched = lambda _: True

clf = ak.ImageRegressor(
    max_trials=args.max_trials,
    directory=args.save_model_dir)

Update (above left for posterity):

I found that there was a place downstream that was also failing in spite of the monkey-patch. As I experimented with AutoKeras, I found a cleaner way that both clears the downstream error and avoids the ugly monkey-patching:

def callable_iterator(generator, expected_batch_size):
  for img_batch, targets_batch in generator:
    if img_batch.shape[0] == expected_batch_size:
      yield img_batch, targets_batch

train_dataset = tf.data.Dataset.from_generator(
    lambda: callable_iterator(train_generator, args.batch_size),
    output_types=(tf.float32, tf.float32),
    output_shapes=(tf.TensorShape([None, args.crop_size, args.crop_size, 1 if args.gray else 3]),
                   tf.TensorShape([None, 1])))
val_dataset = tf.data.Dataset.from_generator(
    lambda: callable_iterator(val_generator, args.batch_size),
    output_types=(tf.float32, tf.float32),
    output_shapes=(tf.TensorShape([None, args.crop_size, args.crop_size, 1 if args.gray else 3]),
                   tf.TensorShape([None, 1])))

Obviously the things starting with args. should be replaced with something appropriate to your code.

This surfaces the appropriate dimensions to the necessary places in AutoKeras. AutoKeras does not behave well if you give it any incomplete batches, so I had to modify callable_iterator as above to reject incomplete batches from the iterator.

[haruiz]

@ciberger, this code seems to work, you can give a shot:

import tensorflow as tf
import numpy as np
import autokeras as ak
from tensorflow.keras.preprocessing import image
import pathlib
import matplotlib.pylab as plt

BATCH_SIZE = 32
IMG_HEIGHT = 224
IMG_WIDTH = 224
data_dir = "dataset/train"
data_dir = pathlib.Path(data_dir)
#image_count = len(list(data_dir.glob('*/*.jpg')))
#STEPS_PER_EPOCH = np.ceil(image_count/BATCH_SIZE)

def preprocess(img):
    img = image.array_to_img(img, scale=False)
    img = img.resize((IMG_WIDTH, IMG_HEIGHT))
    img = image.img_to_array(img)
    return img / 255.0

image_generator = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255,
                                              horizontal_flip=True,
                                              validation_split=0.2,
                                              preprocessing_function=preprocess)

train_generator = image_generator.flow_from_directory(
    directory=str(data_dir),
     batch_size=BATCH_SIZE,
     shuffle=True,
     #class_mode="categorical",
     target_size=(IMG_HEIGHT, IMG_WIDTH),
    subset='training'
)

val_generator = image_generator.flow_from_directory(
    directory=str(data_dir),
     batch_size=BATCH_SIZE,
     shuffle=True,
     #class_mode="categorical",
     target_size=(IMG_HEIGHT, IMG_WIDTH),
    subset='validation'
)

def callable_iterator(generator):
    for img_batch, targets_batch in generator:
        yield img_batch, targets_batch

train_dataset = tf.data.Dataset.from_generator(lambda: callable_iterator(train_generator),output_types=(tf.float32, tf.float32))
val_dataset = tf.data.Dataset.from_generator(lambda: callable_iterator(val_generator),output_types=(tf.float32, tf.float32))

# for image, label in train_dataset.take(1):
#   print("Image shape: ", image.numpy().shape)
#   print("Label: ", label.numpy())
#   plt.imshow(image.numpy()[0] * 255)
#   plt.show()


clf = ak.ImageClassifier(max_trials=10)
#Feed the tensorflow Dataset to the classifier.
clf.fit(train_dataset, epochs=60)
#Evaluate the best model.
print(clf.evaluate(val_dataset))

Then you can feed your fit function with the tf.datasets