GCP.md

GCP Setup

Provisioning

export PROJECT_ID=...
export BUCKET=...

gcloud config set project $PROJECT_ID
gcloud config set compute/zone us-central1-b
gcloud config set compute/region us-central1

gsutil mb -c standard -l us-central1 -b on gs://$BUCKET
ctpu up --name ml-1 --preemptible 

gcloud compute ssh ml-1

Training

export MODEL=lstm_v3 && \
export BUCKET=... && \
export TPU_NAME=ml-1 && \
export STORAGE_BUCKET=gs://$BUCKET && \
export MODEL_DIR=$STORAGE_BUCKET/$MODEL && \
export DATA_DIR=$STORAGE_BUCKET/data && \
export PYTHONPATH="$PYTHONPATH:/usr/share/models"

python3 $MODEL.py \
--model_dir=$MODEL_DIR \
--data_dir=$DATA_DIR \
--distribution_strategy=tpu \
--tpu=$TPU_NAME \
--train_epochs=100

Resources

• console
  • https://console.cloud.google.com
• tutorials
  • https://cloud.google.com/tpu/docs/using-estimator-api
  • https://cloud.google.com/tpu/docs/tutorials
  • https://blog.tensorflow.org/2018/08/code-with-eager-execution-run-with-graphs.html
  • https://www.tensorflow.org/guide/migrate
  • https://github.com/tensorflow/tpu/tree/master/models/official
  • https://github.com/tensorflow/models/tree/master/official
  • https://blog.tensorflow.org/2019/01/what-are-symbolic-and-imperative-apis.html
  • https://developers.google.com/machine-learning/guides/text-classification/step-3
  • https://pbpython.com/categorical-encoding.html