Is it possible to support TPU?
TPU can be used in colab https://colab.research.google.com/github/tensorflow/tpu/blob/master/tools/colab/fashion_mnist.ipynb
vincent-hui
All 5 comments
@vincent-hui Yes, we are working on this issue.
Please refer to #105 .
Thanks.
haifeng-jin
on 23 Oct 2018
@haifeng-jin so is there TPU support in Colab for AutoKeras? Or do we have to stick with GPU鈥檚?
neel04
on 19 Dec 2020
This is a really old issue.
I expect we already support TPU,
but having some performance issues.
Should be solved when we support compile function for AutoModels.
haifeng-jin
on 11 Feb 2021
If anyone is having any error caused by running AutoKeras on TPU.
Please paste it here.
haifeng-jin
on 11 Feb 2021
Looks like TPU is supported as KerasTuner now supports TPU, here is an example: https://colab.research.google.com/drive/1AaFDTQaq8B5p3-FKtNDdMdhlmUHtwOIM?authuser=1#scrollTo=HUJ0LxZnbQnc
and it's quite simple actually with only 2 lines of code:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=tpu_address)
strategy = tf.distribute.TPUStrategy(resolver)
clf = ak.ImageClassifier(max_trials=3, distribution_strategy=strategy)
Note: It looks like using with strategy.scope() will not work (same as KerasTuner) and the TPU strategy have to be passed down this way. I think it will be great if this is added somewhere in the documentation (I'm happy to contribute if needed), also thanks for the awesome package!
Edit: Actually there are some models generate that are not compatible with TPU will break, see the same colab above.
shun-lin
on 16 Feb 2021
Related issues
xuzhang5788
路
4Comments
michelleowen
路
6Comments
mlaradji
路
4Comments
nectario
路
5Comments
max1563
路
4Comments
Most helpful comment
Looks like TPU is supported as KerasTuner now supports TPU, here is an example: https://colab.research.google.com/drive/1AaFDTQaq8B5p3-FKtNDdMdhlmUHtwOIM?authuser=1#scrollTo=HUJ0LxZnbQnc
and it's quite simple actually with only 2 lines of code:
Note: It looks like using
with strategy.scope()will not work (same as KerasTuner) and the TPU strategy have to be passed down this way. I think it will be great if this is added somewhere in the documentation (I'm happy to contribute if needed), also thanks for the awesome package!Edit: Actually there are some models generate that are not compatible with TPU will break, see the same colab above.