official/nlp/continuous_finetune_lib.py

# Copyright 2023 The TensorFlow Authors. All Rights Reserved.
#
# 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.

"""TFM continuous finetuning+eval training driver library."""
import gc
import os
import time
from typing import Any, Mapping, Optional

from absl import logging
import tensorflow as tf, tf_keras

from official.common import distribute_utils
from official.core import config_definitions
from official.core import task_factory
from official.core import train_lib
from official.core import train_utils
from official.modeling import performance
from official.modeling.multitask import configs
from official.modeling.multitask import train_lib as multitask_train_lib


def _flatten_dict(xs):
  """Flatten a nested dictionary.

  The nested keys are flattened to a tuple.

  Example::

    xs = {'foo': 1, 'bar': {'a': 2, 'b': {}}}
    flat_xs = flatten_dict(xs)
    print(flat_xs)
    # {
    #   ('foo',): 1,
    #   ('bar', 'a'): 2,
    # }

  Note that empty dictionaries are ignored and
  will not be restored by `unflatten_dict`.

  Args:
    xs: a nested dictionary

  Returns:
    The flattened dictionary.
  """
  assert isinstance(xs, dict), 'input is not a dict'

  def _flatten(xs, prefix):
    if not isinstance(xs, dict):
      return {prefix: xs}
    result = {}
    for key, value in xs.items():
      path = prefix + (key,)
      result.update(_flatten(value, path))
    return result

  return _flatten(xs, ())


def run_continuous_finetune(
    mode: str,
    params: config_definitions.ExperimentConfig,
    model_dir: str,
    run_post_eval: bool = False,
    pretrain_steps: Optional[int] = None,
) -> Mapping[str, Any]:
  """Run modes with continuous training.

  Currently only supports continuous_train_and_eval.

  Args:
    mode: A 'str', specifying the mode. continuous_train_and_eval - monitors a
      checkpoint directory. Once a new checkpoint is discovered, loads the
      checkpoint, finetune the model by training it (probably on another dataset
      or with another task), then evaluate the finetuned model.
    params: ExperimentConfig instance.
    model_dir: A 'str', a path to store model checkpoints and summaries.
    run_post_eval: Whether to run post eval once after training, metrics logs
      are returned.
    pretrain_steps: Optional, the number of total training steps for the
      pretraining job.

  Returns:
    eval logs: returns eval metrics logs when run_post_eval is set to True,
      othewise, returns {}.
  """

  assert mode == 'continuous_train_and_eval', (
      'Only continuous_train_and_eval is supported by continuous_finetune. '
      'Got mode: {}'.format(mode))

  # Sets mixed_precision policy. Using 'mixed_float16' or 'mixed_bfloat16'
  # can have significant impact on model speeds by utilizing float16 in case of
  # GPUs, and bfloat16 in the case of TPUs. loss_scale takes effect only when
  # dtype is float16
  if params.runtime.mixed_precision_dtype:
    performance.set_mixed_precision_policy(params.runtime.mixed_precision_dtype)
  distribution_strategy = distribute_utils.get_distribution_strategy(
      distribution_strategy=params.runtime.distribution_strategy,
      all_reduce_alg=params.runtime.all_reduce_alg,
      num_gpus=params.runtime.num_gpus,
      tpu_address=params.runtime.tpu)

  retry_times = 0
  while not tf.io.gfile.isdir(params.task.init_checkpoint):
    # Wait for the init_checkpoint directory to be created.
    if retry_times >= 60:
      raise ValueError(
          'ExperimentConfig.task.init_checkpoint must be a directory for '
          'continuous_train_and_eval mode.')
    retry_times += 1
    time.sleep(60)

  summary_writer = tf.summary.create_file_writer(
      os.path.join(model_dir, 'eval'))

  global_step = 0

  def timeout_fn():
    if pretrain_steps and global_step < pretrain_steps:
      # Keeps waiting for another timeout period.
      logging.info(
          'Continue waiting for new checkpoint as current pretrain '
          'global_step=%d and target is %d.', global_step, pretrain_steps)
      return False
    # Quits the loop.
    return True

  for pretrain_ckpt in tf.train.checkpoints_iterator(
      checkpoint_dir=params.task.init_checkpoint,
      min_interval_secs=10,
      timeout=params.trainer.continuous_eval_timeout,
      timeout_fn=timeout_fn):

    # If there are checkpoints, they might be the finetune checkpoint of a
    # different pretrained checkpoint. So we just remove all checkpoints.
    train_utils.remove_ckpts(model_dir)

    with distribution_strategy.scope():
      global_step = train_utils.read_global_step_from_checkpoint(pretrain_ckpt)
    # Replaces params.task.init_checkpoint to make sure that we load
    # exactly this pretrain checkpoint.
    if params.trainer.best_checkpoint_export_subdir:
      best_ckpt_subdir = '{}_{}'.format(
          params.trainer.best_checkpoint_export_subdir, global_step)
      params_replaced = params.replace(
          task={'init_checkpoint': pretrain_ckpt},
          trainer={'best_checkpoint_export_subdir': best_ckpt_subdir})
    else:
      params_replaced = params.replace(task={'init_checkpoint': pretrain_ckpt})
    params_replaced.lock()
    logging.info('Running finetuning with params: %s', params_replaced)

    with distribution_strategy.scope():
      if isinstance(params, configs.MultiEvalExperimentConfig):
        task = task_factory.get_task(params_replaced.task)
        eval_tasks = [
            task_factory.get_task(config.task_config, name=config.task_name)
            for config in params.eval_tasks
        ]
        (_,
         eval_metrics) = multitask_train_lib.run_experiment_with_multitask_eval(
             distribution_strategy=distribution_strategy,
             train_task=task,
             eval_tasks=eval_tasks,
             mode='train_and_eval',
             params=params_replaced,
             model_dir=model_dir,
             run_post_eval=True,
             save_summary=False)
      else:
        task = task_factory.get_task(
            params_replaced.task, logging_dir=model_dir)
        _, eval_metrics = train_lib.run_experiment(
            distribution_strategy=distribution_strategy,
            task=task,
            mode='train_and_eval',
            params=params_replaced,
            model_dir=model_dir,
            run_post_eval=True,
            save_summary=False)
    logging.info('Evaluation finished. Pretrain global_step: %d', global_step)
    train_utils.write_json_summary(model_dir, global_step, eval_metrics)

    if not os.path.basename(model_dir):  # if model_dir.endswith('/')
      summary_grp = os.path.dirname(model_dir) + '_' + task.name
    else:
      summary_grp = os.path.basename(model_dir) + '_' + task.name
    summaries = {}
    for name, value in _flatten_dict(eval_metrics).items():
      summaries[summary_grp + '/' + '-'.join(name)] = value
    train_utils.write_summary(summary_writer, global_step, summaries)

    train_utils.remove_ckpts(model_dir)
    # In TF2, the resource life cycle is bound with the python object life
    # cycle. Force trigger python garbage collection here so those resources
    # can be deallocated in time, so it doesn't cause OOM when allocating new
    # objects.
    # TODO(b/169178664): Fix cycle reference in Keras model and revisit to see
    # if we need gc here.
    gc.collect()

  if run_post_eval:
    return eval_metrics
  return {}