models/official/nlp/transformer/transformer_main.py

# Copyright 2018 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.
# ==============================================================================
"""Train and evaluate the Transformer model.

See README for description of setting the training schedule and evaluating the
BLEU score.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import tempfile

from absl import app
from absl import flags
from absl import logging
import tensorflow as tf

from official.modeling import performance
from official.nlp.transformer import compute_bleu
from official.nlp.transformer import data_pipeline
from official.nlp.transformer import metrics
from official.nlp.transformer import misc
from official.nlp.transformer import optimizer
from official.nlp.transformer import transformer
from official.nlp.transformer import translate
from official.nlp.transformer.utils import tokenizer
from official.utils.flags import core as flags_core
from official.utils.misc import distribution_utils
from official.utils.misc import keras_utils

INF = int(1e9)
BLEU_DIR = "bleu"
_SINGLE_SAMPLE = 1


def translate_and_compute_bleu(model,
                               params,
                               subtokenizer,
                               bleu_source,
                               bleu_ref,
                               distribution_strategy=None):
  """Translate file and report the cased and uncased bleu scores.

  Args:
    model: A Keras model, used to generate the translations.
    params: A dictionary, containing the translation related parameters.
    subtokenizer: A subtokenizer object, used for encoding and decoding source
      and translated lines.
    bleu_source: A file containing source sentences for translation.
    bleu_ref: A file containing the reference for the translated sentences.
    distribution_strategy: A platform distribution strategy, used for TPU based
      translation.

  Returns:
    uncased_score: A float, the case insensitive BLEU score.
    cased_score: A float, the case sensitive BLEU score.
  """
  # Create temporary file to store translation.
  tmp = tempfile.NamedTemporaryFile(delete=False)
  tmp_filename = tmp.name

  translate.translate_file(
      model,
      params,
      subtokenizer,
      bleu_source,
      output_file=tmp_filename,
      print_all_translations=False,
      distribution_strategy=distribution_strategy)

  # Compute uncased and cased bleu scores.
  uncased_score = compute_bleu.bleu_wrapper(bleu_ref, tmp_filename, False)
  cased_score = compute_bleu.bleu_wrapper(bleu_ref, tmp_filename, True)
  os.remove(tmp_filename)
  return uncased_score, cased_score


def evaluate_and_log_bleu(model,
                          params,
                          bleu_source,
                          bleu_ref,
                          vocab_file,
                          distribution_strategy=None):
  """Calculate and record the BLEU score.

  Args:
    model: A Keras model, used to generate the translations.
    params: A dictionary, containing the translation related parameters.
    bleu_source: A file containing source sentences for translation.
    bleu_ref: A file containing the reference for the translated sentences.
    vocab_file: A file containing the vocabulary for translation.
    distribution_strategy: A platform distribution strategy, used for TPU based
      translation.

  Returns:
    uncased_score: A float, the case insensitive BLEU score.
    cased_score: A float, the case sensitive BLEU score.
  """
  subtokenizer = tokenizer.Subtokenizer(vocab_file)

  uncased_score, cased_score = translate_and_compute_bleu(
      model, params, subtokenizer, bleu_source, bleu_ref, distribution_strategy)

  logging.info("Bleu score (uncased): %s", uncased_score)
  logging.info("Bleu score (cased): %s", cased_score)
  return uncased_score, cased_score


class TransformerTask(object):
  """Main entry of Transformer model."""

  def __init__(self, flags_obj):
    """Init function of TransformerMain.

    Args:
      flags_obj: Object containing parsed flag values, i.e., FLAGS.

    Raises:
      ValueError: if not using static batch for input data on TPU.
    """
    self.flags_obj = flags_obj
    self.predict_model = None

    # Add flag-defined parameters to params object
    num_gpus = flags_core.get_num_gpus(flags_obj)
    self.params = params = misc.get_model_params(flags_obj.param_set, num_gpus)

    params["num_gpus"] = num_gpus
    params["use_ctl"] = flags_obj.use_ctl
    params["data_dir"] = flags_obj.data_dir
    params["model_dir"] = flags_obj.model_dir
    params["static_batch"] = flags_obj.static_batch
    params["max_length"] = flags_obj.max_length
    params["decode_batch_size"] = flags_obj.decode_batch_size
    params["decode_max_length"] = flags_obj.decode_max_length
    params["padded_decode"] = flags_obj.padded_decode
    params["max_io_parallelism"] = (
        flags_obj.num_parallel_calls or tf.data.experimental.AUTOTUNE)

    params["use_synthetic_data"] = flags_obj.use_synthetic_data
    params["batch_size"] = flags_obj.batch_size or params["default_batch_size"]
    params["repeat_dataset"] = None
    params["dtype"] = flags_core.get_tf_dtype(flags_obj)
    params["enable_tensorboard"] = flags_obj.enable_tensorboard
    params["enable_metrics_in_training"] = flags_obj.enable_metrics_in_training
    params["steps_between_evals"] = flags_obj.steps_between_evals
    params["enable_checkpointing"] = flags_obj.enable_checkpointing

    self.distribution_strategy = distribution_utils.get_distribution_strategy(
        distribution_strategy=flags_obj.distribution_strategy,
        num_gpus=num_gpus,
        all_reduce_alg=flags_obj.all_reduce_alg,
        num_packs=flags_obj.num_packs,
        tpu_address=flags_obj.tpu or "")
    if self.use_tpu:
      params["num_replicas"] = self.distribution_strategy.num_replicas_in_sync
    else:
      logging.info("Running transformer with num_gpus = %d", num_gpus)

    if self.distribution_strategy:
      logging.info("For training, using distribution strategy: %s",
                   self.distribution_strategy)
    else:
      logging.info("Not using any distribution strategy.")

    performance.set_mixed_precision_policy(
        params["dtype"],
        flags_core.get_loss_scale(flags_obj, default_for_fp16="dynamic"))

  @property
  def use_tpu(self):
    if self.distribution_strategy:
      return isinstance(self.distribution_strategy,
                        tf.distribute.experimental.TPUStrategy)
    return False

  def train(self):
    """Trains the model."""
    params = self.params
    flags_obj = self.flags_obj
    # Sets config options.
    keras_utils.set_session_config(enable_xla=flags_obj.enable_xla)

    _ensure_dir(flags_obj.model_dir)
    with distribution_utils.get_strategy_scope(self.distribution_strategy):
      model = transformer.create_model(params, is_train=True)
      opt = self._create_optimizer()

      current_step = 0
      checkpoint = tf.train.Checkpoint(model=model, optimizer=opt)
      latest_checkpoint = tf.train.latest_checkpoint(flags_obj.model_dir)
      if latest_checkpoint:
        checkpoint.restore(latest_checkpoint)
        logging.info("Loaded checkpoint %s", latest_checkpoint)
        current_step = opt.iterations.numpy()

      if params["use_ctl"]:
        train_loss_metric = tf.keras.metrics.Mean(
            "training_loss", dtype=tf.float32)
        if params["enable_tensorboard"]:
          summary_writer = tf.compat.v2.summary.create_file_writer(
              flags_obj.model_dir)
        else:
          summary_writer = tf.compat.v2.summary.create_noop_writer()
        train_metrics = [train_loss_metric]
        if params["enable_metrics_in_training"]:
          train_metrics = train_metrics + model.metrics
      else:
        model.compile(opt)

    model.summary()

    if self.use_tpu:
      # Different from experimental_distribute_dataset,
      # experimental_distribute_datasets_from_function requires
      # per-replica/local batch size.
      params["batch_size"] /= self.distribution_strategy.num_replicas_in_sync
      train_ds = (
          self.distribution_strategy
          .experimental_distribute_datasets_from_function(
              lambda ctx: data_pipeline.train_input_fn(params, ctx)))
    else:
      train_ds = data_pipeline.train_input_fn(params)
      map_data_fn = data_pipeline.map_data_for_transformer_fn
      train_ds = train_ds.map(
          map_data_fn, num_parallel_calls=tf.data.experimental.AUTOTUNE)
    if params["use_ctl"]:
      train_ds_iterator = iter(train_ds)

    callbacks = self._create_callbacks(flags_obj.model_dir, 0, params)

    # Only TimeHistory callback is supported for CTL
    if params["use_ctl"]:
      callbacks = [cb for cb in callbacks
                   if isinstance(cb, keras_utils.TimeHistory)]

    # TODO(b/139418525): Refactor the custom training loop logic.
    @tf.function
    def train_steps(iterator, steps):
      """Training steps function for TPU runs.

      Args:
        iterator: The input iterator of the training dataset.
        steps: An integer, the number of training steps.

      Returns:
        A float, the loss value.
      """

      def _step_fn(inputs):
        """Per-replica step function."""
        inputs, targets = inputs
        with tf.GradientTape() as tape:
          logits = model([inputs, targets], training=True)
          loss = metrics.transformer_loss(logits, targets,
                                          params["label_smoothing"],
                                          params["vocab_size"])
          # Scales the loss, which results in using the average loss across all
          # of the replicas for backprop.
          scaled_loss = loss / self.distribution_strategy.num_replicas_in_sync

        # De-dupes variables due to keras tracking issues.
        tvars = list({id(v): v for v in model.trainable_variables}.values())
        grads = tape.gradient(scaled_loss, tvars)
        opt.apply_gradients(zip(grads, tvars))
        # For reporting, the metric takes the mean of losses.
        train_loss_metric.update_state(loss)

      for _ in tf.range(steps):
        train_loss_metric.reset_states()
        self.distribution_strategy.run(
            _step_fn, args=(next(iterator),))

    cased_score, uncased_score = None, None
    cased_score_history, uncased_score_history = [], []
    while current_step < flags_obj.train_steps:
      remaining_steps = flags_obj.train_steps - current_step
      train_steps_per_eval = (
          remaining_steps if remaining_steps < flags_obj.steps_between_evals
          else flags_obj.steps_between_evals)
      current_iteration = current_step // flags_obj.steps_between_evals

      logging.info(
          "Start train iteration at global step:{}".format(current_step))
      history = None
      if params["use_ctl"]:
        if not self.use_tpu:
          raise NotImplementedError(
              "Custom training loop on GPUs is not implemented.")

        # Runs training steps.
        with summary_writer.as_default():
          for cb in callbacks:
            cb.on_epoch_begin(current_iteration)
            cb.on_batch_begin(0)

          train_steps(
              train_ds_iterator,
              tf.convert_to_tensor(train_steps_per_eval, dtype=tf.int32))
          current_step += train_steps_per_eval
          train_loss = train_loss_metric.result().numpy().astype(float)
          logging.info("Train Step: %d/%d / loss = %s", current_step,
                       flags_obj.train_steps, train_loss)

          for cb in callbacks:
            cb.on_batch_end(train_steps_per_eval - 1)
            cb.on_epoch_end(current_iteration)

          if params["enable_tensorboard"]:
            for metric_obj in train_metrics:
              tf.compat.v2.summary.scalar(metric_obj.name, metric_obj.result(),
                                          current_step)
              summary_writer.flush()

        for cb in callbacks:
          cb.on_train_end()

        if flags_obj.enable_checkpointing:
          # avoid check-pointing when running for benchmarking.
          checkpoint_name = checkpoint.save(
              os.path.join(flags_obj.model_dir,
                           "ctl_step_{}.ckpt".format(current_step)))
          logging.info("Saved checkpoint to %s", checkpoint_name)
      else:
        if self.use_tpu:
          raise NotImplementedError(
              "Keras model.fit on TPUs is not implemented.")
        history = model.fit(
            train_ds,
            initial_epoch=current_iteration,
            epochs=current_iteration + 1,
            steps_per_epoch=train_steps_per_eval,
            callbacks=callbacks,
            # If TimeHistory is enabled, progress bar would be messy. Increase
            # the verbose level to get rid of it.
            verbose=(2 if flags_obj.enable_time_history else 1))
        current_step += train_steps_per_eval
        logging.info("Train history: {}".format(history.history))

      logging.info("End train iteration at global step:{}".format(current_step))

      if (flags_obj.bleu_source and flags_obj.bleu_ref):
        uncased_score, cased_score = self.eval()
        cased_score_history.append([current_iteration + 1, cased_score])
        uncased_score_history.append([current_iteration + 1, uncased_score])

    stats = ({
        "loss": train_loss
    } if history is None else {})
    misc.update_stats(history, stats, callbacks)
    if uncased_score and cased_score:
      stats["bleu_uncased"] = uncased_score
      stats["bleu_cased"] = cased_score
      stats["bleu_uncased_history"] = uncased_score_history
      stats["bleu_cased_history"] = cased_score_history
    return stats

  def eval(self):
    """Evaluates the model."""
    distribution_strategy = self.distribution_strategy if self.use_tpu else None

    # We only want to create the model under DS scope for TPU case.
    # When 'distribution_strategy' is None, a no-op DummyContextManager will
    # be used.
    with distribution_utils.get_strategy_scope(distribution_strategy):
      if not self.predict_model:
        self.predict_model = transformer.create_model(self.params, False)
      self._load_weights_if_possible(
          self.predict_model,
          tf.train.latest_checkpoint(self.flags_obj.model_dir))
      self.predict_model.summary()
    return evaluate_and_log_bleu(
        self.predict_model, self.params, self.flags_obj.bleu_source,
        self.flags_obj.bleu_ref, self.flags_obj.vocab_file,
        distribution_strategy)

  def predict(self):
    """Predicts result from the model."""
    params = self.params
    flags_obj = self.flags_obj

    with tf.name_scope("model"):
      model = transformer.create_model(params, is_train=False)
      self._load_weights_if_possible(
          model, tf.train.latest_checkpoint(self.flags_obj.model_dir))
      model.summary()
    subtokenizer = tokenizer.Subtokenizer(flags_obj.vocab_file)

    ds = data_pipeline.eval_input_fn(params)
    ds = ds.map(lambda x, y: x).take(_SINGLE_SAMPLE)
    ret = model.predict(ds)
    val_outputs, _ = ret
    length = len(val_outputs)
    for i in range(length):
      translate.translate_from_input(val_outputs[i], subtokenizer)

  def _create_callbacks(self, cur_log_dir, init_steps, params):
    """Creates a list of callbacks."""
    sfunc = optimizer.LearningRateFn(params["learning_rate"],
                                     params["hidden_size"],
                                     params["learning_rate_warmup_steps"])
    scheduler_callback = optimizer.LearningRateScheduler(sfunc, init_steps)
    callbacks = misc.get_callbacks()
    callbacks.append(scheduler_callback)
    if params["enable_checkpointing"]:
      ckpt_full_path = os.path.join(cur_log_dir, "cp-{epoch:04d}.ckpt")
      callbacks.append(
          tf.keras.callbacks.ModelCheckpoint(
              ckpt_full_path, save_weights_only=True))
    return callbacks

  def _load_weights_if_possible(self, model, init_weight_path=None):
    """Loads model weights when it is provided."""
    if init_weight_path:
      logging.info("Load weights: {}".format(init_weight_path))
      # TODO(b/139414977): Having the same variable restoring method for both
      # TPU and GPU.
      if self.use_tpu:
        checkpoint = tf.train.Checkpoint(
            model=model, optimizer=self._create_optimizer())
        checkpoint.restore(init_weight_path)
      else:
        model.load_weights(init_weight_path)
    else:
      logging.info("Weights not loaded from path:{}".format(init_weight_path))

  def _create_optimizer(self):
    """Creates optimizer."""
    params = self.params
    lr_schedule = optimizer.LearningRateSchedule(
        params["learning_rate"], params["hidden_size"],
        params["learning_rate_warmup_steps"])
    opt = tf.keras.optimizers.Adam(
        lr_schedule if self.use_tpu else params["learning_rate"],
        params["optimizer_adam_beta1"],
        params["optimizer_adam_beta2"],
        epsilon=params["optimizer_adam_epsilon"])

    opt = performance.configure_optimizer(
        opt,
        use_float16=params["dtype"] == tf.float16,
        use_graph_rewrite=self.flags_obj.fp16_implementation == "graph_rewrite",
        loss_scale=flags_core.get_loss_scale(
            self.flags_obj, default_for_fp16="dynamic"))

    return opt


def _ensure_dir(log_dir):
  """Makes log dir if not existed."""
  if not tf.io.gfile.exists(log_dir):
    tf.io.gfile.makedirs(log_dir)


def main(_):
  flags_obj = flags.FLAGS
  if flags_obj.enable_mlir_bridge:
    tf.config.experimental.enable_mlir_bridge()
  task = TransformerTask(flags_obj)

  # Execute flag override logic for better model performance
  if flags_obj.tf_gpu_thread_mode:
    keras_utils.set_gpu_thread_mode_and_count(
        per_gpu_thread_count=flags_obj.per_gpu_thread_count,
        gpu_thread_mode=flags_obj.tf_gpu_thread_mode,
        num_gpus=flags_obj.num_gpus,
        datasets_num_private_threads=flags_obj.datasets_num_private_threads)

  if flags_obj.mode == "train":
    task.train()
  elif flags_obj.mode == "predict":
    task.predict()
  elif flags_obj.mode == "eval":
    task.eval()
  else:
    raise ValueError("Invalid mode {}".format(flags_obj.mode))


if __name__ == "__main__":
  logging.set_verbosity(logging.INFO)
  misc.define_transformer_flags()
  app.run(main)