在GCP TPU上训练我的模型

Question

What architecture of program can run in TPU?

Is it must to use tf.contrib.tpu.TPUEstimator ?

Is there any tpu-program example for reference except the tensorflow official model?

Answer 1

Google Cloud TPUs uses TensorFlow [1] which supports Python, C++, Java, Go and Swift APIs [2]. Here is the link to Google Cloud TPU System Architecture [1]. TPUEstimator handle low-level, hardware-specific details and it makes it easier to achieve maximum performance by applying performance optimizations automatically on your behalf [3]. Also, if you are using for neural network, then you should use TPUEstimator. Here are some steps to convert from TensorFlow Estimator to TPUEstimator [4]. Additional tutorials are located on Google Cloud TPU Tutorials page [5] and also on GitHub [6].

[1] https://cloud.google.com/tpu/docs/system-architecture

[2] https://www.tensorflow.org/api_docs/

[3] https://cloud.google.com/tpu/docs/using-estimator-api

[4] https://cloud.google.com/tpu/docs/using-estimator-api#converting_from_tensorflow_estimator_to_tpuestimator

[5] https://cloud.google.com/tpu/docs/tutorials

[6] https://github.com/tensorflow/tpu