AutoKeras 1.0 much slower than 0.4 on Google Colab

When I try to run a simple MNIST example on Google Colab with GPU with Autokeras 0.4 it runs very fast (1 epoch of the first model takes < 2 s) but with 1.0 it runs much slower (1 epoch of the first model takes > 80 s). When I disable the GPU 0.4 runs as slow as 1.0 which suggests that 1.0 isn’t using the GPU. How can I make Autokeras 1.0 run as fast as 0.4 with GPU?

To reproduce go to colab.research.google.com, choose a Python 3 runtime with GPU accelerator, and execute the following

0.4 code

%tensorflow_version 1.x

!pip install autokeras

import autokeras
import tensorflow

( ( x, y ), validation_data ) = tensorflow.keras.datasets.mnist.load_data( )
model = autokeras.ImageClassifier( verbose = True )
model.fit( x, y )

1.0 code

%tensorflow_version 2.x

!pip install git+git://github.com/keras-team/keras-tuner@master#egg=keras-tuner
!pip install git+git://github.com/keras-team/autokeras@master#egg=autokeras

import tensorflow
import autokeras

( ( x, y ), validation_data ) = tensorflow.keras.datasets.mnist.load_data( )
model = autokeras.ImageClassifier( )
model.fit( x, y, validation_data = validation_data )

The issue is breakdown to the following issues. After solving them, the speed should be improved.

#906, #907, #908, #909, #910.

[pezdorado]

I’ve tried the functional API with an ImageBlock( augment = False ) but it’s as slow as the task API. I hope I applied the functional API correctly since the tutorial is quite sparse

!pip install git+git://github.com/keras-team/autokeras@master#egg=autokeras

import tensorflow
import autokeras

( ( x, y ), validation_data ) = tensorflow.keras.datasets.mnist.load_data( )

inputs = autokeras.ImageInput( )
blocks = autokeras.hypermodel.hyperblock.ImageBlock( augment = False )( inputs )
outputs = autokeras.ClassificationHead( )( blocks )

model = autokeras.GraphAutoModel(
    inputs = inputs,
    outputs = outputs
)

model.fit( x, y, validation_split = 0.1, batch_size = 128 )

Did I apply the functional API correctly and did I correctly turn off augmentation? If so, then the problem does not seem to be caused by augmentation.

[haifeng-jin]

This issue has been breakdown to multiple parts in the AutoKeras Project. After adopting the preprocessing layers the issue should be resolved.