In [ ]:

    

# Only for when running on Colab:
import sys
if 'google.colab' in sys.modules:
    # Get the dependency .py files, if any.
    ! git clone https://github.com/GoogleCloudPlatform/cloudml-samples.git
    ! cp cloudml-samples/tpu/templates/tpu_estimator/* .

    # Authenticate the user for better GCS access.
    # Copy verification code into the text field to continue.
    from google.colab import auth
    auth.authenticate_user()


    

In [ ]:

    

import argparse
import numpy as np
import tensorflow as tf


    

In [ ]:

    

def model_fn(features, labels, mode, params):
    # build model
    global_step = tf.train.get_global_step()
    hidden = tf.layers.dense(features, 10, activation=tf.nn.relu)
    output = tf.layers.dense(hidden, 1)

    predictions = output
    loss = None
    train_op = None

    if mode == tf.estimator.ModeKeys.TRAIN:
        # define loss
        loss = tf.nn.l2_loss(predictions - labels)

        # define train_op
        optimizer = tf.train.RMSPropOptimizer(learning_rate=0.05)

        # wrapper to make the optimizer work with TPUs
        if params['use_tpu']:
            optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)

        train_op = optimizer.minimize(loss, global_step=global_step)

    if params['use_tpu']:
        # TPU version of EstimatorSpec
        return tf.contrib.tpu.TPUEstimatorSpec(
            mode=mode,
            predictions=predictions,
            loss=loss,
            train_op=train_op)
    else:
        return tf.estimator.EstimatorSpec(
            mode=mode,
            predictions=predictions,
            loss=loss,
            train_op=train_op)


    

In [ ]:

    

def train_input_fn(params={}):
    # make some fake regression data
    x = np.random.rand(100, 5)
    w = np.random.rand(5)
    y = np.sum(x * w, axis=1)

    # TPUs currently do not support float64
    x_tensor = tf.constant(x, dtype=tf.float32)
    y_tensor = tf.constant(y, dtype=tf.float32)

    # create tf.data.Dataset
    dataset = tf.data.Dataset.from_tensor_slices((x_tensor, y_tensor))

    # TPUEstimator passes params when calling input_fn
    batch_size = params.get('batch_size', 16)

    dataset = dataset.repeat().shuffle(32).batch(batch_size, drop_remainder=True)

    # TPUs need to know all dimensions when the graph is built
    # Datasets know the batch size only when the graph is run
    def set_shapes(features, labels):
        features_shape = features.get_shape().merge_with([batch_size, None])
        labels_shape = labels.get_shape().merge_with([batch_size])

        features.set_shape(features_shape)
        labels.set_shape(labels_shape)

        return features, labels

    dataset = dataset.map(set_shapes)
    dataset = dataset.prefetch(tf.contrib.data.AUTOTUNE)

    return dataset


    

In [ ]:

    

def main(args):
    # pass the args as params so the model_fn can use
    # the TPU specific args
    params = vars(args)

    if args.use_tpu:
        # additional configs required for using TPUs
        tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver(args.tpu)
        tpu_config = tf.contrib.tpu.TPUConfig(
            num_shards=8, # using Cloud TPU v2-8
            iterations_per_loop=args.save_checkpoints_steps)

        # use the TPU version of RunConfig
        config = tf.contrib.tpu.RunConfig(
            cluster=tpu_cluster_resolver,
            model_dir=args.model_dir,
            tpu_config=tpu_config,
            save_checkpoints_steps=args.save_checkpoints_steps,
            save_summary_steps=100)

        # TPUEstimator
        estimator = tf.contrib.tpu.TPUEstimator(
            model_fn=model_fn,
            config=config,
            params=params,
            train_batch_size=args.train_batch_size,
            eval_batch_size=32,
            export_to_tpu=False)
    else:
        config = tf.estimator.RunConfig(model_dir=args.model_dir)

        estimator = tf.estimator.Estimator(
            model_fn,
            config=config,
            params=params)

    estimator.train(train_input_fn, max_steps=args.max_steps)


    

In [ ]:

    

parser = argparse.ArgumentParser()

parser.add_argument(
    '--model-dir',
    type=str,
    default='/tmp/tpu-template',
    help='Location to write checkpoints and summaries to.  Must be a GCS URI when using Cloud TPU.')
parser.add_argument(
    '--max-steps',
    type=int,
    default=1000,
    help='The total number of steps to train the model.')
parser.add_argument(
    '--train-batch-size',
    type=int,
    default=16,
    help='The training batch size.  The training batch is divided evenly across the TPU cores.')
parser.add_argument(
    '--save-checkpoints-steps',
    type=int,
    default=100,
    help='The number of training steps before saving each checkpoint.')
parser.add_argument(
    '--use-tpu',
    action='store_true',
    help='Whether to use TPU.')
parser.add_argument(
    '--tpu',
    default=None,
    help='The name or GRPC URL of the TPU node.  Leave it as `None` when training on AI Platform.')

args, _ = parser.parse_known_args()


    

In [ ]:

    

# colab.research.google.com specific
if 'google.colab' in sys.modules:
    import json
    import os

    # TODO(user): change this
    args.model_dir = 'gs://your-gcs-bucket'

    # When connected to the TPU runtime
    if 'COLAB_TPU_ADDR' in os.environ:
        tpu_grpc = 'grpc://{}'.format(os.environ['COLAB_TPU_ADDR'])

        args.tpu = tpu_grpc
        args.use_tpu = True

        # Upload credentials to the TPU
        with tf.Session(tpu_grpc) as sess:
            data = json.load(open('/content/adc.json'))
            tf.contrib.cloud.configure_gcs(sess, credentials=data)


    

In [ ]:

    

main(args)