Imageai: TPU Training

[flozi00]

Is it possible to add TPU support for training the custom modells ?

It should be easy using following code after model.compile function

try: model = tf.contrib.tpu.keras_to_tpu_model( model, strategy=tf.contrib.tpu.TPUDistributionStrategy( tf.contrib.cluster_resolver.TPUClusterResolver('grpc://' + TPU_ADDRESS))) except: pass

from:
https://nbviewer.jupyter.org/github/cedrickchee/data-science-notebooks/blob/master/notebooks/tensorflow/google_cloud_tpu/guide_to_tensorflow_keras_on_tpu_mnist.ipynb

[OlafenwaMoses]

@flozi00 Thanks for sharing this useful information. However, I have concerns on backward compatibility with previous versions of Tensorflow and Keras. I haven't work with the TPU before.

[flozi00]

Maybe it would be a nice feature in one of the future versions

[kettenbach-it]

TPU support would be a killer feature!
Consider that training on TPUs is much faster and cheaper then training on GPUs

[OlafenwaMoses]

@kettenbach-it Thanks for the comment. We will consider support for this in our future updates.