task.py

#!/usr/bin/env python
# Copyright 2019 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""This is the Chicago Taxi example for AI Platform training."""

import argparse
import logging
import os
import sys

from datetime import datetime
import tensorflow as tf

from . import model
from . import experiment


def get_args():
    """Define the task arguments with the default values.

    Returns:
        experiment parameters
    """

    args_parser = argparse.ArgumentParser()

    # Data files arguments
    args_parser.add_argument(
        '--train-files',
        help='GCS or local paths to training data.',
        nargs='+',
        required=True)
    args_parser.add_argument(
        '--eval-files',
        help='GCS or local paths to evaluation data.',
        nargs='+',
        required=True)

    # Experiment arguments
    args_parser.add_argument(
        '--train-steps',
        help="""
        Steps to run the training job for.
        If --num-epochs and --train-size are not specified,
        this must be specified;
        otherwise the training job will run indefinitely.
        if --num-epochs and --train-size are specified,
        then --train-steps will default to:
            (train-size/train-batch-size) * num-epochs
        """,
        default=0,
        type=int)
    args_parser.add_argument(
        '--eval-steps',
        help="""
        Number of steps to run evaluation for at each checkpoint.,
        Set to None to evaluate on the whole evaluation data.
        """,
        default=None,
        type=int)
    args_parser.add_argument(
        '--batch-size',
        help='Batch size for each training and evaluation step.',
        type=int,
        default=128)
    args_parser.add_argument(
        '--train-size',
        help="""
        Size of the training data (instance count).
        If both --train-size and --num-epochs are specified,
        --train-steps will default to:
            (train-size/train-batch-size) * num-epochs.
        """,
        type=int,
        default=None)
    args_parser.add_argument(
        '--num-epochs',
        help="""
        Maximum number of training data epochs on which to train.
        If both --train-size and --num-epochs are specified,
        --train-steps will default to:
            (train-size/train-batch-size) * num-epochs.
        """,
        default=100,
        type=int,
    )
    args_parser.add_argument(
        '--eval-frequency-secs',
        help='How many seconds to wait before running the next evaluation.',
        default=15,
        type=int)

    # Feature columns arguments
    args_parser.add_argument(
        '--embed-categorical-columns',
        help="""
        If set to True, the categorical columns will be embedded
        and used in the deep part of the model.
        The embedding size = sqrt(vocab_size).
        """,
        action='store_true',
        default=True,
    )
    args_parser.add_argument(
        '--use-indicator-columns',
        help="""
        If set to True, the categorical columns will be encoded
        as One-Hot indicators in the deep part of the model.
        """,
        action='store_true',
        default=False,
    )
    args_parser.add_argument(
        '--use-wide-columns',
        help="""
        If set to True, the categorical columns will be used in the
        wide part of the model.
        """,
        action='store_true',
        default=False,
    )

    # Estimator arguments
    args_parser.add_argument(
        '--learning-rate',
        help='Learning rate value for the optimizers.',
        default=0.1,
        type=float)
    args_parser.add_argument(
        '--learning-rate-decay-factor',
        help="""
        The factor by which the learning rate should decay by the end of the
        training.
        decayed_learning_rate = learning_rate * decay_rate ^ (global_step /
        decay_steps).
        If set to 1.0 (default), then no decay will occur.
        If set to 0.5, then the learning rate should reach 0.5 of its original
        value at the end of the training.
        Note that decay_steps is set to train_steps.
        """,
        default=1.0,
        type=float)
    args_parser.add_argument(
        '--hidden-units',
        help="""
        Hidden layer sizes to use for DNN feature columns, provided in
        comma-separated layers.
        If --scale-factor > 0, then only the size of the first layer will be
        used to compute the sizes of subsequent layers.
        """,
        default='30,30,30')
    args_parser.add_argument(
        '--layer-sizes-scale-factor',
        help="""
        Determine how the size of the layers in the DNN are scaled per
        successive layer.
        If value = 0 then the provided --hidden-units will be the
        same per layer.
        """,
        default=0.7,
        type=float)
    args_parser.add_argument(
        '--num-layers',
        help="""
        Number of layers in the DNN. If --layer-sizes-scale-factor > 0,
        then this parameter is ignored.
        """,
        default=4,
        type=int)
    args_parser.add_argument(
        '--dropout-prob',
        help='The probability we will drop out a given coordinate.',
        default=None)

    # Saved model arguments
    args_parser.add_argument(
        '--job-dir',
        help='GCS location to write checkpoints and export models.',
        required=True)
    args_parser.add_argument(
        '--reuse-job-dir',
        action='store_true',
        default=False,
        help="""
        Flag to decide if the model checkpoint should be
        re-used from the job-dir.
        If set to False then the job-dir will be deleted.
        """)
    args_parser.add_argument(
        '--serving-export-format',
        help='The input format of the exported serving SavedModel.',
        choices=['JSON', 'CSV', 'EXAMPLE'],
        default='JSON')
    args_parser.add_argument(
        '--eval-export-format',
        help='The input format of the exported evaluating SavedModel.',
        choices=['CSV', 'EXAMPLE'],
        default='CSV')

    return args_parser.parse_args()


def _setup_logging():
    """Sets up logging."""
    root_logger = logging.getLogger()
    root_logger_previous_handlers = list(root_logger.handlers)
    for h in root_logger_previous_handlers:
        root_logger.removeHandler(h)
    root_logger.setLevel(logging.INFO)
    root_logger.propagate = False

    # Set tf logging to avoid duplicate logging. If the handlers are not removed
    # then we will have duplicate logging
    tf_logger = logging.getLogger('TensorFlow')
    while tf_logger.handlers:
        tf_logger.removeHandler(tf_logger.handlers[0])

    # Redirect INFO logs to stdout
    stdout_handler = logging.StreamHandler(sys.stdout)
    stdout_handler.setLevel(logging.INFO)
    root_logger.addHandler(stdout_handler)

    # Suppress C++ level warnings.
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'


def main():
    args = get_args()
    _setup_logging()

    # If job_dir_reuse is False then remove the job_dir if it exists.
    logging.info('Resume training: {}'.format(args.reuse_job_dir))
    if not args.reuse_job_dir:
        if tf.io.gfile.exists(args.job_dir):
            tf.io.gfile.rmtree(args.job_dir)
            logging.info(
                'Deleted job_dir {} to avoid re-use'.format(args.job_dir))
    else:
        logging.info('Reusing job_dir {} if it exists'.format(args.job_dir))

    run_config = experiment.create_run_config(args)
    logging.info('Job directory: {}'.format(run_config.model_dir))

    # Compute the number of Training steps.
    if args.train_size is not None and args.num_epochs is not None:
        args.train_steps = int(
            (args.train_size / args.batch_size) * args.num_epochs)
    else:
        args.train_steps = args.train_steps

    logging.info('Train size: {}.'.format(args.train_size))
    logging.info('Epoch count: {}.'.format(args.num_epochs))
    logging.info('Batch size: {}.'.format(args.batch_size))
    logging.info('Training steps: {} ({}).'.format(
        args.train_steps,
        'supplied' if args.train_size is None else 'computed'))
    logging.info('Evaluate every {} steps.'.format(args.eval_frequency_secs))

    # Create the Estimator
    estimator = model.create(args, run_config)
    logging.info('Creating an Estimator: {}'.format(type(estimator)))

    # Run the train and evaluate experiment
    time_start = datetime.utcnow()
    logging.info('Experiment started...')
    logging.info('.......................................')

    # Run experiment
    experiment.run(estimator, args)

    time_end = datetime.utcnow()
    logging.info('.......................................')
    logging.info('Experiment finished.')
    time_elapsed = time_end - time_start
    logging.info('Experiment elapsed time: {} seconds'.format(
        time_elapsed.total_seconds()))


if __name__ == '__main__':
    main()