Exemplo n.º 1
0
def main(argv):
  del argv  # unused arg
  tf.io.gfile.makedirs(FLAGS.output_dir)
  logging.info('Saving checkpoints at %s', FLAGS.output_dir)
  tf.random.set_seed(FLAGS.seed)

  if FLAGS.use_gpu:
    logging.info('Use GPU')
    strategy = tf.distribute.MirroredStrategy()
  else:
    logging.info('Use TPU at %s',
                 FLAGS.tpu if FLAGS.tpu is not None else 'local')
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=FLAGS.tpu)
    tf.config.experimental_connect_to_cluster(resolver)
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.TPUStrategy(resolver)

  ds_info = tfds.builder(FLAGS.dataset).info
  train_batch_size = FLAGS.per_core_batch_size * FLAGS.num_cores // FLAGS.batch_repetitions
  test_batch_size = FLAGS.per_core_batch_size * FLAGS.num_cores
  train_dataset_size = ds_info.splits['train'].num_examples
  steps_per_epoch = train_dataset_size // train_batch_size
  steps_per_eval = ds_info.splits['test'].num_examples // test_batch_size
  num_classes = ds_info.features['label'].num_classes

  train_dataset = utils.load_dataset(
      split=tfds.Split.TRAIN,
      name=FLAGS.dataset,
      batch_size=train_batch_size,
      use_bfloat16=FLAGS.use_bfloat16)
  clean_test_dataset = utils.load_dataset(
      split=tfds.Split.TEST,
      name=FLAGS.dataset,
      batch_size=test_batch_size,
      use_bfloat16=FLAGS.use_bfloat16)
  train_dataset = strategy.experimental_distribute_dataset(train_dataset)
  test_datasets = {
      'clean': strategy.experimental_distribute_dataset(clean_test_dataset),
  }
  if FLAGS.corruptions_interval > 0:
    if FLAGS.dataset == 'cifar10':
      load_c_dataset = utils.load_cifar10_c
    else:
      load_c_dataset = functools.partial(utils.load_cifar100_c,
                                         path=FLAGS.cifar100_c_path)
    corruption_types, max_intensity = utils.load_corrupted_test_info(
        FLAGS.dataset)
    for corruption in corruption_types:
      for intensity in range(1, max_intensity + 1):
        dataset = load_c_dataset(
            corruption_name=corruption,
            corruption_intensity=intensity,
            batch_size=test_batch_size,
            use_bfloat16=FLAGS.use_bfloat16)
        test_datasets['{0}_{1}'.format(corruption, intensity)] = (
            strategy.experimental_distribute_dataset(dataset))

  if FLAGS.use_bfloat16:
    policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
    tf.keras.mixed_precision.experimental.set_policy(policy)

  summary_writer = tf.summary.create_file_writer(
      os.path.join(FLAGS.output_dir, 'summaries'))

  with strategy.scope():
    logging.info('Building Keras model')
    model = cifar_model.wide_resnet(
        input_shape=[FLAGS.ensemble_size] +
        list(ds_info.features['image'].shape),
        depth=28,
        width_multiplier=FLAGS.width_multiplier,
        num_classes=num_classes,
        ensemble_size=FLAGS.ensemble_size)
    logging.info('Model input shape: %s', model.input_shape)
    logging.info('Model output shape: %s', model.output_shape)
    logging.info('Model number of weights: %s', model.count_params())
    # Linearly scale learning rate and the decay epochs by vanilla settings.
    base_lr = FLAGS.base_learning_rate * train_batch_size / 128
    lr_decay_epochs = [(int(start_epoch_str) * FLAGS.train_epochs) // 200
                       for start_epoch_str in FLAGS.lr_decay_epochs]
    lr_schedule = utils.LearningRateSchedule(steps_per_epoch, base_lr,
                                             FLAGS.lr_decay_ratio,
                                             lr_decay_epochs,
                                             FLAGS.lr_warmup_epochs)
    optimizer = tf.keras.optimizers.SGD(
        lr_schedule, momentum=0.9, nesterov=True)
    metrics = {
        'train/negative_log_likelihood': tf.keras.metrics.Mean(),
        'train/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
        'train/loss': tf.keras.metrics.Mean(),
        'train/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
        'test/negative_log_likelihood': tf.keras.metrics.Mean(),
        'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
        'test/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
    }
    if FLAGS.corruptions_interval > 0:
      corrupt_metrics = {}
      for intensity in range(1, max_intensity + 1):
        for corruption in corruption_types:
          dataset_name = '{0}_{1}'.format(corruption, intensity)
          corrupt_metrics['test/nll_{}'.format(dataset_name)] = (
              tf.keras.metrics.Mean())
          corrupt_metrics['test/accuracy_{}'.format(dataset_name)] = (
              tf.keras.metrics.SparseCategoricalAccuracy())
          corrupt_metrics['test/ece_{}'.format(dataset_name)] = (
              um.ExpectedCalibrationError(num_bins=FLAGS.num_bins))

    for i in range(FLAGS.ensemble_size):
      metrics['test/nll_member_{}'.format(i)] = tf.keras.metrics.Mean()
      metrics['test/accuracy_member_{}'.format(i)] = (
          tf.keras.metrics.SparseCategoricalAccuracy())
    test_diversity = {
        'test/disagreement': tf.keras.metrics.Mean(),
        'test/average_kl': tf.keras.metrics.Mean(),
        'test/cosine_similarity': tf.keras.metrics.Mean(),
    }

    checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
    latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
    initial_epoch = 0
    if latest_checkpoint:
      # checkpoint.restore must be within a strategy.scope() so that optimizer
      # slot variables are mirrored.
      checkpoint.restore(latest_checkpoint)
      logging.info('Loaded checkpoint %s', latest_checkpoint)
      initial_epoch = optimizer.iterations.numpy() // steps_per_epoch

  @tf.function
  def train_step(iterator):
    """Training StepFn."""

    def step_fn(inputs):
      """Per-Replica StepFn."""
      images, labels = inputs
      batch_size = tf.shape(images)[0]

      main_shuffle = tf.random.shuffle(tf.tile(
          tf.range(batch_size), [FLAGS.batch_repetitions]))
      to_shuffle = tf.cast(tf.cast(tf.shape(main_shuffle)[0], tf.float32)
                           * (1. - FLAGS.input_repetition_probability),
                           tf.int32)
      shuffle_indices = [
          tf.concat([tf.random.shuffle(main_shuffle[:to_shuffle]),
                     main_shuffle[to_shuffle:]], axis=0)
          for _ in range(FLAGS.ensemble_size)]
      images = tf.stack([tf.gather(images, indices, axis=0)
                         for indices in shuffle_indices], axis=1)
      labels = tf.stack([tf.gather(labels, indices, axis=0)
                         for indices in shuffle_indices], axis=1)

      with tf.GradientTape() as tape:
        logits = model(images, training=True)
        if FLAGS.use_bfloat16:
          logits = tf.cast(logits, tf.float32)

        negative_log_likelihood = tf.reduce_mean(tf.reduce_sum(
            tf.keras.losses.sparse_categorical_crossentropy(
                labels, logits, from_logits=True), axis=1))
        filtered_variables = []
        for var in model.trainable_variables:
          # Apply l2 on the BN parameters and bias terms.
          if ('kernel' in var.name or 'batch_norm' in var.name or
              'bias' in var.name):
            filtered_variables.append(tf.reshape(var, (-1,)))

        l2_loss = FLAGS.l2 * 2 * tf.nn.l2_loss(
            tf.concat(filtered_variables, axis=0))

        # Scale the loss given the TPUStrategy will reduce sum all gradients.
        loss = negative_log_likelihood + l2_loss
        scaled_loss = loss / strategy.num_replicas_in_sync

      grads = tape.gradient(scaled_loss, model.trainable_variables)
      optimizer.apply_gradients(zip(grads, model.trainable_variables))

      probs = tf.nn.softmax(tf.reshape(logits, [-1, num_classes]))
      flat_labels = tf.reshape(labels, [-1])
      metrics['train/ece'].update_state(flat_labels, probs)
      metrics['train/loss'].update_state(loss)
      metrics['train/negative_log_likelihood'].update_state(
          negative_log_likelihood)
      metrics['train/accuracy'].update_state(flat_labels, probs)

    strategy.run(step_fn, args=(next(iterator),))

  @tf.function
  def test_step(iterator, dataset_name):
    """Evaluation StepFn."""

    def step_fn(inputs):
      """Per-Replica StepFn."""
      images, labels = inputs
      images = tf.tile(
          tf.expand_dims(images, 1), [1, FLAGS.ensemble_size, 1, 1, 1])
      logits = model(images, training=False)
      if FLAGS.use_bfloat16:
        logits = tf.cast(logits, tf.float32)
      probs = tf.nn.softmax(logits)

      if dataset_name == 'clean':
        per_probs = tf.transpose(probs, perm=[1, 0, 2])
        diversity_results = um.average_pairwise_diversity(
            per_probs, FLAGS.ensemble_size)
        for k, v in diversity_results.items():
          test_diversity['test/' + k].update_state(v)

      for i in range(FLAGS.ensemble_size):
        member_probs = probs[:, i]
        member_loss = tf.keras.losses.sparse_categorical_crossentropy(
            labels, member_probs)
        metrics['test/nll_member_{}'.format(i)].update_state(member_loss)
        metrics['test/accuracy_member_{}'.format(i)].update_state(
            labels, member_probs)

      # Negative log marginal likelihood computed in a numerically-stable way.
      labels_tiled = tf.tile(
          tf.expand_dims(labels, 1), [1, FLAGS.ensemble_size])
      log_likelihoods = -tf.keras.losses.sparse_categorical_crossentropy(
          labels_tiled, logits, from_logits=True)
      negative_log_likelihood = tf.reduce_mean(
          -tf.reduce_logsumexp(log_likelihoods, axis=[1]) +
          tf.math.log(float(FLAGS.ensemble_size)))
      probs = tf.math.reduce_mean(probs, axis=1)  # marginalize

      if dataset_name == 'clean':
        metrics['test/negative_log_likelihood'].update_state(
            negative_log_likelihood)
        metrics['test/accuracy'].update_state(labels, probs)
        metrics['test/ece'].update_state(labels, probs)
      else:
        corrupt_metrics['test/nll_{}'.format(dataset_name)].update_state(
            negative_log_likelihood)
        corrupt_metrics['test/accuracy_{}'.format(dataset_name)].update_state(
            labels, probs)
        corrupt_metrics['test/ece_{}'.format(dataset_name)].update_state(
            labels, probs)

    strategy.run(step_fn, args=(next(iterator),))

  metrics.update({'test/ms_per_example': tf.keras.metrics.Mean()})

  train_iterator = iter(train_dataset)
  start_time = time.time()
  for epoch in range(initial_epoch, FLAGS.train_epochs):
    logging.info('Starting to run epoch: %s', epoch)

    for step in range(steps_per_epoch):
      train_step(train_iterator)

      current_step = epoch * steps_per_epoch + (step + 1)
      max_steps = steps_per_epoch * (FLAGS.train_epochs)
      time_elapsed = time.time() - start_time
      steps_per_sec = float(current_step) / time_elapsed
      eta_seconds = (max_steps - current_step) / steps_per_sec
      message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
                 'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
                     current_step / max_steps, epoch + 1, FLAGS.train_epochs,
                     steps_per_sec, eta_seconds / 60, time_elapsed / 60))
      if step % 20 == 0:
        logging.info(message)

    datasets_to_evaluate = {'clean': test_datasets['clean']}
    if (FLAGS.corruptions_interval > 0 and
        (epoch + 1) % FLAGS.corruptions_interval == 0):
      datasets_to_evaluate = test_datasets
    for dataset_name, test_dataset in datasets_to_evaluate.items():
      test_iterator = iter(test_dataset)
      logging.info('Testing on dataset %s', dataset_name)
      for step in range(steps_per_eval):
        if step % 20 == 0:
          logging.info('Starting to run eval step %s of epoch: %s', step, epoch)
        test_start_time = time.time()
        test_step(test_iterator, dataset_name)
        ms_per_example = (time.time() - test_start_time) * 1e6 / test_batch_size
        metrics['test/ms_per_example'].update_state(ms_per_example)
      logging.info('Done with testing on %s', dataset_name)

    corrupt_results = {}
    if (FLAGS.corruptions_interval > 0 and
        (epoch + 1) % FLAGS.corruptions_interval == 0):
      corrupt_results = utils.aggregate_corrupt_metrics(corrupt_metrics,
                                                        corruption_types,
                                                        max_intensity)

    logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
                 metrics['train/loss'].result(),
                 metrics['train/accuracy'].result() * 100)
    logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
                 metrics['test/negative_log_likelihood'].result(),
                 metrics['test/accuracy'].result() * 100)
    for i in range(FLAGS.ensemble_size):
      logging.info(
          'Member %d Test Loss: %.4f, Accuracy: %.2f%%', i,
          metrics['test/nll_member_{}'.format(i)].result(),
          metrics['test/accuracy_member_{}'.format(i)].result() * 100)

    metrics.update(test_diversity)
    total_results = {name: metric.result() for name, metric in metrics.items()}
    total_results.update(corrupt_results)
    with summary_writer.as_default():
      for name, result in total_results.items():
        tf.summary.scalar(name, result, step=epoch + 1)

    for metric in metrics.values():
      metric.reset_states()

    if (FLAGS.checkpoint_interval > 0 and
        (epoch + 1) % FLAGS.checkpoint_interval == 0):
      checkpoint_name = checkpoint.save(
          os.path.join(FLAGS.output_dir, 'checkpoint'))
      logging.info('Saved checkpoint to %s', checkpoint_name)

  final_checkpoint_name = checkpoint.save(
      os.path.join(FLAGS.output_dir, 'checkpoint'))
  logging.info('Saved last checkpoint to %s', final_checkpoint_name)
Exemplo n.º 2
0
def main(argv):
  del argv  # Unused arg.

  tf.random.set_seed(FLAGS.seed)

  if FLAGS.version2:
    per_core_bs_train = FLAGS.per_core_batch_size // (FLAGS.ensemble_size *
                                                      FLAGS.num_train_samples)
    per_core_bs_eval = FLAGS.per_core_batch_size // (FLAGS.ensemble_size *
                                                     FLAGS.num_eval_samples)
  else:
    per_core_bs_train = FLAGS.per_core_batch_size // FLAGS.num_train_samples
    per_core_bs_eval = FLAGS.per_core_batch_size // FLAGS.num_eval_samples
  batch_size_train = per_core_bs_train * FLAGS.num_cores
  batch_size_eval = per_core_bs_eval * FLAGS.num_cores

  logging.info('Saving checkpoints at %s', FLAGS.output_dir)

  if FLAGS.use_gpu:
    logging.info('Use GPU')
    strategy = tf.distribute.MirroredStrategy()
  else:
    logging.info('Use TPU at %s',
                 FLAGS.tpu if FLAGS.tpu is not None else 'local')
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=FLAGS.tpu)
    tf.config.experimental_connect_to_cluster(resolver)
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.TPUStrategy(resolver)

  train_dataset = utils.load_dataset(
      split=tfds.Split.TRAIN,
      name=FLAGS.dataset,
      batch_size=batch_size_train,
      use_bfloat16=FLAGS.use_bfloat16,
      normalize=False)
  clean_test_dataset = utils.load_dataset(
      split=tfds.Split.TEST,
      name=FLAGS.dataset,
      batch_size=batch_size_eval,
      use_bfloat16=FLAGS.use_bfloat16,
      normalize=False)
  train_dataset = strategy.experimental_distribute_dataset(train_dataset)
  test_datasets = {
      'clean': strategy.experimental_distribute_dataset(clean_test_dataset),
  }
  if FLAGS.corruptions_interval > 0:
    if FLAGS.dataset == 'cifar10':
      load_c_dataset = utils.load_cifar10_c
    else:
      load_c_dataset = functools.partial(utils.load_cifar100_c,
                                         path=FLAGS.cifar100_c_path)
    corruption_types, max_intensity = utils.load_corrupted_test_info(
        FLAGS.dataset)
    for corruption in corruption_types:
      for intensity in range(1, max_intensity + 1):
        dataset = load_c_dataset(
            corruption_name=corruption,
            corruption_intensity=intensity,
            batch_size=batch_size_eval,
            use_bfloat16=FLAGS.use_bfloat16,
            normalize=False)
        test_datasets['{0}_{1}'.format(corruption, intensity)] = (
            strategy.experimental_distribute_dataset(dataset))

  ds_info = tfds.builder(FLAGS.dataset).info
  train_dataset_size = ds_info.splits['train'].num_examples
  test_dataset_size = ds_info.splits['test'].num_examples
  num_classes = ds_info.features['label'].num_classes

  steps_per_epoch = train_dataset_size // batch_size_train
  steps_per_eval = test_dataset_size // batch_size_eval

  if FLAGS.use_bfloat16:
    policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
    tf.keras.mixed_precision.experimental.set_policy(policy)

  summary_writer = tf.summary.create_file_writer(
      os.path.join(FLAGS.output_dir, 'summaries'))

  with strategy.scope():
    logging.info('Building Keras ResNet-32 model')
    model = resnet_cifar_model.rank1_resnet_v1(
        input_shape=ds_info.features['image'].shape,
        depth=32,
        num_classes=num_classes,
        width_multiplier=4,
        alpha_initializer=FLAGS.alpha_initializer,
        gamma_initializer=FLAGS.gamma_initializer,
        alpha_regularizer=FLAGS.alpha_regularizer,
        gamma_regularizer=FLAGS.gamma_regularizer,
        use_additive_perturbation=FLAGS.use_additive_perturbation,
        ensemble_size=FLAGS.ensemble_size,
        random_sign_init=FLAGS.random_sign_init,
        dropout_rate=FLAGS.dropout_rate)
    logging.info(model.summary())
    base_lr = FLAGS.base_learning_rate * batch_size_train / 128
    lr_decay_epochs = [(int(start_epoch_str) * FLAGS.train_epochs) // 200
                       for start_epoch_str in FLAGS.lr_decay_epochs]
    lr_schedule = utils.LearningRateSchedule(
        steps_per_epoch,
        base_lr,
        decay_ratio=FLAGS.lr_decay_ratio,
        decay_epochs=lr_decay_epochs,
        warmup_epochs=FLAGS.lr_warmup_epochs)
    optimizer = tf.keras.optimizers.SGD(
        lr_schedule, momentum=0.9, nesterov=True)
    metrics = {
        'train/negative_log_likelihood': tf.keras.metrics.Mean(),
        'train/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
        'train/loss': tf.keras.metrics.Mean(),
        'train/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
        'test/negative_log_likelihood': tf.keras.metrics.Mean(),
        'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
        'test/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
        'test/loss': tf.keras.metrics.Mean(),
    }
    if FLAGS.corruptions_interval > 0:
      corrupt_metrics = {}
      for intensity in range(1, max_intensity + 1):
        for corruption in corruption_types:
          dataset_name = '{0}_{1}'.format(corruption, intensity)
          corrupt_metrics['test/nll_{}'.format(dataset_name)] = (
              tf.keras.metrics.Mean())
          corrupt_metrics['test/accuracy_{}'.format(dataset_name)] = (
              tf.keras.metrics.SparseCategoricalAccuracy())
          corrupt_metrics['test/ece_{}'.format(dataset_name)] = (
              um.ExpectedCalibrationError(num_bins=FLAGS.num_bins))

    test_diversity = {}
    training_diversity = {}
    if FLAGS.ensemble_size > 1:
      for i in range(FLAGS.ensemble_size):
        metrics['test/nll_member_{}'.format(i)] = tf.keras.metrics.Mean()
        metrics['test/accuracy_member_{}'.format(i)] = (
            tf.keras.metrics.SparseCategoricalAccuracy())
      test_diversity = {
          'test/disagreement': tf.keras.metrics.Mean(),
          'test/average_kl': tf.keras.metrics.Mean(),
          'test/cosine_similarity': tf.keras.metrics.Mean(),
      }
      training_diversity = {
          'train/disagreement': tf.keras.metrics.Mean(),
          'train/average_kl': tf.keras.metrics.Mean(),
          'train/cosine_similarity': tf.keras.metrics.Mean(),
      }

    checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
    latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
    initial_epoch = 0
    if latest_checkpoint:
      # checkpoint.restore must be within a strategy.scope() so that optimizer
      # slot variables are mirrored.
      checkpoint.restore(latest_checkpoint)
      logging.info('Loaded checkpoint %s', latest_checkpoint)
      initial_epoch = optimizer.iterations.numpy() // steps_per_epoch

  @tf.function
  def train_step(iterator):
    """Training StepFn."""
    def step_fn(inputs):
      """Per-Replica StepFn."""
      images, labels = inputs
      if FLAGS.version2 and FLAGS.ensemble_size > 1:
        images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
        if not (FLAGS.member_sampling or FLAGS.expected_probs):
          labels = tf.tile(labels, [FLAGS.ensemble_size])

      if FLAGS.num_train_samples > 1:
        images = tf.tile(images, [FLAGS.num_train_samples, 1, 1, 1])

      with tf.GradientTape() as tape:
        logits = model(images, training=True)
        probs = tf.nn.softmax(logits)
        # Diversity evaluation.
        if FLAGS.version2 and FLAGS.ensemble_size > 1:
          per_probs = tf.reshape(
              probs, tf.concat([[FLAGS.ensemble_size, -1], probs.shape[1:]], 0))

          diversity_results = um.average_pairwise_diversity(
              per_probs, FLAGS.ensemble_size)

        if FLAGS.num_train_samples > 1:
          probs = tf.reshape(probs,
                             tf.concat([[FLAGS.num_train_samples, -1],
                                        probs.shape[1:]], 0))
          probs = tf.reduce_mean(probs, 0)

        if FLAGS.member_sampling and FLAGS.version2 and FLAGS.ensemble_size > 1:
          idx = tf.random.uniform([], maxval=FLAGS.ensemble_size,
                                  dtype=tf.int64)
          idx_one_hot = tf.expand_dims(tf.one_hot(idx, FLAGS.ensemble_size,
                                                  dtype=probs.dtype), 0)
          probs_shape = probs.shape
          probs = tf.reshape(probs, [FLAGS.ensemble_size, -1])
          probs = tf.matmul(idx_one_hot, probs)
          probs = tf.reshape(probs, tf.concat([[-1], probs_shape[1:]], 0))

        elif FLAGS.expected_probs and FLAGS.version2 and FLAGS.ensemble_size > 1:
          probs = tf.reshape(probs,
                             tf.concat([[FLAGS.ensemble_size, -1],
                                        probs.shape[1:]], 0))
          probs = tf.reduce_mean(probs, 0)

        negative_log_likelihood = tf.reduce_mean(
            tf.keras.losses.sparse_categorical_crossentropy(labels, probs))

        filtered_variables = []
        for var in model.trainable_variables:
          # Apply l2 on the slow weights and bias terms. This excludes BN
          # parameters and fast weight approximate posterior/prior parameters,
          # but pay caution to their naming scheme.
          if 'kernel' in var.name or 'bias' in var.name:
            filtered_variables.append(tf.reshape(var, (-1,)))

        l2_loss = FLAGS.l2 * 2 * tf.nn.l2_loss(
            tf.concat(filtered_variables, axis=0))
        kl = sum(model.losses) / train_dataset_size
        kl_scale = tf.cast(optimizer.iterations + 1, kl.dtype)
        kl_scale /= FLAGS.kl_annealing_steps
        kl_scale = tf.minimum(1., kl_scale)
        kl_loss = kl_scale * kl

        # Scale the loss given the TPUStrategy will reduce sum all gradients.
        loss = negative_log_likelihood + l2_loss + kl_loss
        scaled_loss = loss / strategy.num_replicas_in_sync

      grads = tape.gradient(scaled_loss, model.trainable_variables)

      # Separate learning rate implementation.
      grad_list = []
      if FLAGS.fast_weight_lr_multiplier != 1.0:
        grads_and_vars = list(zip(grads, model.trainable_variables))
        for vec, var in grads_and_vars:
          # Apply different learning rate on the fast weight approximate
          # posterior/prior parameters. This is excludes BN and slow weights,
          # but pay caution to the naming scheme.
          if ('batch_norm' not in var.name and 'kernel' not in var.name):
            grad_list.append((vec * FLAGS.fast_weight_lr_multiplier, var))
          else:
            grad_list.append((vec, var))
        optimizer.apply_gradients(grad_list)
      else:
        optimizer.apply_gradients(zip(grads, model.trainable_variables))

      metrics['train/ece'].update_state(labels, probs)
      metrics['train/loss'].update_state(loss)
      metrics['train/negative_log_likelihood'].update_state(
          negative_log_likelihood)
      metrics['train/accuracy'].update_state(labels, probs)
      if FLAGS.version2 and FLAGS.ensemble_size > 1:
        for k, v in diversity_results.items():
          training_diversity['train/' + k].update_state(v)

    strategy.run(step_fn, args=(next(iterator),))

  @tf.function
  def test_step(iterator, dataset_name):
    """Evaluation StepFn."""
    def step_fn(inputs):
      """Per-Replica StepFn."""
      images, labels = inputs
      if FLAGS.ensemble_size > 1:
        images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
      if FLAGS.num_eval_samples > 1:
        images = tf.tile(images, [FLAGS.num_eval_samples, 1, 1, 1])
      logits = model(images, training=False)
      probs = tf.nn.softmax(logits)

      if FLAGS.num_eval_samples > 1:
        probs = tf.reshape(probs,
                           tf.concat([[FLAGS.num_eval_samples, -1],
                                      probs.shape[1:]], 0))
        probs = tf.reduce_mean(probs, 0)

      if FLAGS.ensemble_size > 1:
        per_probs = tf.split(probs,
                             num_or_size_splits=FLAGS.ensemble_size,
                             axis=0)
        if dataset_name == 'clean':
          per_probs_tensor = tf.reshape(
              probs, tf.concat([[FLAGS.ensemble_size, -1], probs.shape[1:]], 0))
          diversity_results = um.average_pairwise_diversity(
              per_probs_tensor, FLAGS.ensemble_size)

          for k, v in diversity_results.items():
            test_diversity['test/' + k].update_state(v)

          for i in range(FLAGS.ensemble_size):
            member_probs = per_probs[i]
            member_nll = tf.keras.losses.sparse_categorical_crossentropy(
                labels, member_probs)
            metrics['test/nll_member_{}'.format(i)].update_state(member_nll)
            metrics['test/accuracy_member_{}'.format(i)].update_state(
                labels, member_probs)

        probs = tf.reduce_mean(per_probs, axis=0)

      negative_log_likelihood = tf.reduce_mean(
          tf.keras.losses.sparse_categorical_crossentropy(labels, probs))
      filtered_variables = []
      for var in model.trainable_variables:
        if 'kernel' in var.name or 'bias' in var.name:
          filtered_variables.append(tf.reshape(var, (-1,)))

      kl = sum(model.losses) / test_dataset_size
      l2_loss = kl + FLAGS.l2 * 2 * tf.nn.l2_loss(
          tf.concat(filtered_variables, axis=0))
      loss = negative_log_likelihood + l2_loss
      if dataset_name == 'clean':
        metrics['test/negative_log_likelihood'].update_state(
            negative_log_likelihood)
        metrics['test/accuracy'].update_state(labels, probs)
        metrics['test/ece'].update_state(labels, probs)
        metrics['test/loss'].update_state(loss)
      else:
        corrupt_metrics['test/nll_{}'.format(dataset_name)].update_state(
            negative_log_likelihood)
        corrupt_metrics['test/accuracy_{}'.format(dataset_name)].update_state(
            labels, probs)
        corrupt_metrics['test/ece_{}'.format(dataset_name)].update_state(
            labels, probs)

    strategy.run(step_fn, args=(next(iterator),))

  train_iterator = iter(train_dataset)
  start_time = time.time()
  for epoch in range(initial_epoch, FLAGS.train_epochs):
    logging.info('Starting to run epoch: %s', epoch)
    for step in range(steps_per_epoch):
      train_step(train_iterator)

      current_step = epoch * steps_per_epoch + (step + 1)
      max_steps = steps_per_epoch * FLAGS.train_epochs
      time_elapsed = time.time() - start_time
      steps_per_sec = float(current_step) / time_elapsed
      eta_seconds = (max_steps - current_step) / steps_per_sec
      message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
                 'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
                     current_step / max_steps,
                     epoch + 1,
                     FLAGS.train_epochs,
                     steps_per_sec,
                     eta_seconds / 60,
                     time_elapsed / 60))
      work_unit.set_notes(message)
      if step % 20 == 0:
        logging.info(message)

    datasets_to_evaluate = {'clean': test_datasets['clean']}
    if (FLAGS.corruptions_interval > 0 and
        (epoch + 1) % FLAGS.corruptions_interval == 0):
      datasets_to_evaluate = test_datasets
    for dataset_name, test_dataset in datasets_to_evaluate.items():
      test_iterator = iter(test_dataset)
      logging.info('Testing on dataset %s', dataset_name)
      for step in range(steps_per_eval):
        if step % 20 == 0:
          logging.info('Starting to run eval step %s of epoch: %s', step,
                       epoch)
        test_step(test_iterator, dataset_name)
      logging.info('Done with testing on %s', dataset_name)

    corrupt_results = {}
    if (FLAGS.corruptions_interval > 0 and
        (epoch + 1) % FLAGS.corruptions_interval == 0):
      corrupt_results = utils.aggregate_corrupt_metrics(corrupt_metrics,
                                                        corruption_types,
                                                        max_intensity)

    logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
                 metrics['train/loss'].result(),
                 metrics['train/accuracy'].result() * 100)
    logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
                 metrics['test/negative_log_likelihood'].result(),
                 metrics['test/accuracy'].result() * 100)
    for i in range(FLAGS.ensemble_size):
      logging.info('Member %d Test Loss: %.4f, Accuracy: %.2f%%',
                   i, metrics['test/nll_member_{}'.format(i)].result(),
                   metrics['test/accuracy_member_{}'.format(i)].result() * 100)
    total_metrics = itertools.chain(metrics.items(),
                                    training_diversity.items(),
                                    test_diversity.items())
    total_results = {name: metric.result() for name, metric in total_metrics}
    total_results.update(corrupt_results)
    with summary_writer.as_default():
      for name, result in total_results.items():
        tf.summary.scalar(name, result, step=epoch + 1)

    for name, result in total_results.items():
      name = name.replace('/', '_')
      if 'negative_log_likelihood' in name:
        # Plots sort WIDs from high-to-low so look at maximization objectives.
        name = name.replace('negative_log_likelihood', 'log_likelihood')
        result = -result
      objective = work_unit.get_measurement_series(name)
      objective.create_measurement(result, epoch + 1)

    for _, metric in total_metrics:
      metric.reset_states()
    summary_writer.flush()

    if (FLAGS.checkpoint_interval > 0 and
        (epoch + 1) % FLAGS.checkpoint_interval == 0):
      checkpoint_name = checkpoint.save(
          os.path.join(FLAGS.output_dir, 'checkpoint'))
      logging.info('Saved checkpoint to %s', checkpoint_name)
Exemplo n.º 3
0
def main(argv):
    del argv  # unused arg
    tf.io.gfile.makedirs(FLAGS.output_dir)
    logging.info('Saving checkpoints at %s', FLAGS.output_dir)
    tf.random.set_seed(FLAGS.seed)

    if FLAGS.use_gpu:
        logging.info('Use GPU')
        strategy = tf.distribute.MirroredStrategy()
    else:
        logging.info('Use TPU at %s',
                     FLAGS.tpu if FLAGS.tpu is not None else 'local')
        resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
            tpu=FLAGS.tpu)
        tf.config.experimental_connect_to_cluster(resolver)
        tf.tpu.experimental.initialize_tpu_system(resolver)
        strategy = tf.distribute.TPUStrategy(resolver)

    ds_info = tfds.builder(FLAGS.dataset).info
    batch_size = ((FLAGS.per_core_batch_size // FLAGS.ensemble_size) *
                  FLAGS.num_cores)
    train_dataset_size = ds_info.splits['train'].num_examples
    steps_per_epoch = train_dataset_size // batch_size
    test_dataset_size = ds_info.splits['test'].num_examples
    steps_per_eval = test_dataset_size // batch_size
    num_classes = ds_info.features['label'].num_classes

    train_dataset = utils.load_dataset(split=tfds.Split.TRAIN,
                                       name=FLAGS.dataset,
                                       batch_size=batch_size,
                                       use_bfloat16=FLAGS.use_bfloat16)
    clean_test_dataset = utils.load_dataset(split=tfds.Split.TEST,
                                            name=FLAGS.dataset,
                                            batch_size=batch_size,
                                            use_bfloat16=FLAGS.use_bfloat16)
    train_dataset = strategy.experimental_distribute_dataset(train_dataset)
    test_datasets = {
        'clean': strategy.experimental_distribute_dataset(clean_test_dataset),
    }
    if FLAGS.corruptions_interval > 0:
        if FLAGS.dataset == 'cifar10':
            load_c_dataset = utils.load_cifar10_c
        else:
            load_c_dataset = functools.partial(utils.load_cifar100_c,
                                               path=FLAGS.cifar100_c_path)
        corruption_types, max_intensity = utils.load_corrupted_test_info(
            FLAGS.dataset)
        for corruption in corruption_types:
            for intensity in range(1, max_intensity + 1):
                dataset = load_c_dataset(corruption_name=corruption,
                                         corruption_intensity=intensity,
                                         batch_size=batch_size,
                                         use_bfloat16=FLAGS.use_bfloat16)
                test_datasets['{0}_{1}'.format(corruption, intensity)] = (
                    strategy.experimental_distribute_dataset(dataset))

    if FLAGS.use_bfloat16:
        policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
        tf.keras.mixed_precision.experimental.set_policy(policy)

    summary_writer = tf.summary.create_file_writer(
        os.path.join(FLAGS.output_dir, 'summaries'))

    with strategy.scope():
        logging.info('Building Keras model')
        model = cifar_model.wide_resnet(
            input_shape=ds_info.features['image'].shape,
            depth=28,
            width_multiplier=10,
            num_classes=num_classes,
            alpha_initializer=FLAGS.alpha_initializer,
            gamma_initializer=FLAGS.gamma_initializer,
            alpha_regularizer=FLAGS.alpha_regularizer,
            gamma_regularizer=FLAGS.gamma_regularizer,
            use_additive_perturbation=FLAGS.use_additive_perturbation,
            ensemble_size=FLAGS.ensemble_size,
            random_sign_init=FLAGS.random_sign_init,
            dropout_rate=FLAGS.dropout_rate,
            prior_mean=FLAGS.prior_mean,
            prior_stddev=FLAGS.prior_stddev)
        logging.info('Model input shape: %s', model.input_shape)
        logging.info('Model output shape: %s', model.output_shape)
        logging.info('Model number of weights: %s', model.count_params())
        # Linearly scale learning rate and the decay epochs by vanilla settings.
        base_lr = FLAGS.base_learning_rate * batch_size / 128
        lr_decay_epochs = [(int(start_epoch_str) * FLAGS.train_epochs) // 200
                           for start_epoch_str in FLAGS.lr_decay_epochs]
        lr_schedule = refining.LearningRateScheduleWithRefining(
            steps_per_epoch,
            base_lr,
            decay_ratio=FLAGS.lr_decay_ratio,
            decay_epochs=lr_decay_epochs,
            warmup_epochs=FLAGS.lr_warmup_epochs,
            train_epochs=FLAGS.train_epochs,
            refining_learning_rate=FLAGS.refining_learning_rate)
        optimizer = tf.keras.optimizers.SGD(lr_schedule,
                                            momentum=0.9,
                                            nesterov=True)
        metrics = {
            'train/negative_log_likelihood': tf.keras.metrics.Mean(),
            'train/kl': tf.keras.metrics.Mean(),
            'train/kl_scale': tf.keras.metrics.Mean(),
            'train/elbo': tf.keras.metrics.Mean(),
            'train/loss': tf.keras.metrics.Mean(),
            'train/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
            'train/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
            'test/negative_log_likelihood': tf.keras.metrics.Mean(),
            'test/kl': tf.keras.metrics.Mean(),
            'test/elbo': tf.keras.metrics.Mean(),
            'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
            'test/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
        }
        if FLAGS.ensemble_size > 1:
            for i in range(FLAGS.ensemble_size):
                metrics['test/nll_member_{}'.format(
                    i)] = tf.keras.metrics.Mean()
                metrics['test/accuracy_member_{}'.format(i)] = (
                    tf.keras.metrics.SparseCategoricalAccuracy())
        if FLAGS.corruptions_interval > 0:
            corrupt_metrics = {}
            for intensity in range(1, max_intensity + 1):
                for corruption in corruption_types:
                    dataset_name = '{0}_{1}'.format(corruption, intensity)
                    corrupt_metrics['test/nll_{}'.format(dataset_name)] = (
                        tf.keras.metrics.Mean())
                    corrupt_metrics['test/kl_{}'.format(dataset_name)] = (
                        tf.keras.metrics.Mean())
                    corrupt_metrics['test/elbo_{}'.format(dataset_name)] = (
                        tf.keras.metrics.Mean())
                    corrupt_metrics['test/accuracy_{}'.format(
                        dataset_name)] = (
                            tf.keras.metrics.SparseCategoricalAccuracy())
                    corrupt_metrics['test/ece_{}'.format(dataset_name)] = (
                        um.ExpectedCalibrationError(num_bins=FLAGS.num_bins))

        checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
        latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
        initial_epoch = 0
        if latest_checkpoint:
            # checkpoint.restore must be within a strategy.scope() so that optimizer
            # slot variables are mirrored.
            checkpoint.restore(latest_checkpoint)
            logging.info('Loaded checkpoint %s', latest_checkpoint)
            initial_epoch = optimizer.iterations.numpy() // steps_per_epoch

    def compute_l2_loss(model):
        filtered_variables = []
        for var in model.trainable_variables:
            # Apply l2 on the BN parameters and bias terms. This
            # excludes only fast weight approximate posterior/prior parameters,
            # but pay caution to their naming scheme.
            if ('kernel' in var.name or 'batch_norm' in var.name
                    or 'bias' in var.name):
                filtered_variables.append(tf.reshape(var, (-1, )))
        l2_loss = FLAGS.l2 * 2 * tf.nn.l2_loss(
            tf.concat(filtered_variables, axis=0))
        return l2_loss

    @tf.function
    def train_step(iterator):
        """Training StepFn."""
        def step_fn(inputs):
            """Per-Replica StepFn."""
            images, labels = inputs
            if FLAGS.ensemble_size > 1:
                images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
                labels = tf.tile(labels, [FLAGS.ensemble_size])

            with tf.GradientTape() as tape:
                logits = model(images, training=True)
                if FLAGS.use_bfloat16:
                    logits = tf.cast(logits, tf.float32)
                negative_log_likelihood = tf.reduce_mean(
                    tf.keras.losses.sparse_categorical_crossentropy(
                        labels, logits, from_logits=True))
                l2_loss = compute_l2_loss(model)
                kl = sum(model.losses) / train_dataset_size
                kl_scale = tf.cast(optimizer.iterations + 1, kl.dtype)
                kl_scale /= steps_per_epoch * FLAGS.kl_annealing_epochs
                kl_scale = tf.minimum(1., kl_scale)
                kl_loss = kl_scale * kl

                # Scale the loss given the TPUStrategy will reduce sum all gradients.
                loss = negative_log_likelihood + l2_loss + kl_loss
                scaled_loss = loss / strategy.num_replicas_in_sync
                elbo = -(negative_log_likelihood + l2_loss + kl)

            grads = tape.gradient(scaled_loss, model.trainable_variables)

            # Separate learning rate implementation.
            if FLAGS.fast_weight_lr_multiplier != 1.0:
                grads_and_vars = []
                for grad, var in zip(grads, model.trainable_variables):
                    # Apply different learning rate on the fast weight approximate
                    # posterior/prior parameters. This is excludes BN and slow weights,
                    # but pay caution to the naming scheme.
                    if ('kernel' not in var.name
                            and 'batch_norm' not in var.name
                            and 'bias' not in var.name):
                        grads_and_vars.append(
                            (grad * FLAGS.fast_weight_lr_multiplier, var))
                    else:
                        grads_and_vars.append((grad, var))
                optimizer.apply_gradients(grads_and_vars)
            else:
                optimizer.apply_gradients(zip(grads,
                                              model.trainable_variables))

            probs = tf.nn.softmax(logits)
            metrics['train/negative_log_likelihood'].update_state(
                negative_log_likelihood)
            metrics['train/kl'].update_state(kl)
            metrics['train/kl_scale'].update_state(kl_scale)
            metrics['train/elbo'].update_state(elbo)
            metrics['train/loss'].update_state(loss)
            metrics['train/accuracy'].update_state(labels, probs)
            metrics['train/ece'].update_state(labels, probs)

        strategy.run(step_fn, args=(next(iterator), ))

    @tf.function
    def test_step(iterator, dataset_name):
        """Evaluation StepFn."""
        def step_fn(inputs):
            """Per-Replica StepFn."""
            images, labels = inputs
            if FLAGS.ensemble_size > 1:
                images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
            logits = tf.reshape([
                model(images, training=False)
                for _ in range(FLAGS.num_eval_samples)
            ], [FLAGS.num_eval_samples, FLAGS.ensemble_size, -1, num_classes])
            if FLAGS.use_bfloat16:
                logits = tf.cast(logits, tf.float32)
            probs = tf.nn.softmax(logits)

            if FLAGS.ensemble_size > 1:
                per_probs = tf.reduce_mean(probs,
                                           axis=0)  # marginalize samples
                for i in range(FLAGS.ensemble_size):
                    member_probs = per_probs[i]
                    member_loss = tf.keras.losses.sparse_categorical_crossentropy(
                        labels, member_probs)
                    metrics['test/nll_member_{}'.format(i)].update_state(
                        member_loss)
                    metrics['test/accuracy_member_{}'.format(i)].update_state(
                        labels, member_probs)

            # Negative log marginal likelihood computed in a numerically-stable way.
            labels_broadcasted = tf.broadcast_to(
                labels,
                [FLAGS.num_eval_samples, FLAGS.ensemble_size, labels.shape[0]])
            log_likelihoods = -tf.keras.losses.sparse_categorical_crossentropy(
                labels_broadcasted, logits, from_logits=True)
            negative_log_likelihood = tf.reduce_mean(
                -tf.reduce_logsumexp(log_likelihoods, axis=[0, 1]) +
                tf.math.log(float(FLAGS.num_eval_samples *
                                  FLAGS.ensemble_size)))
            probs = tf.math.reduce_mean(probs, axis=[0, 1])  # marginalize

            l2_loss = compute_l2_loss(model)
            kl = sum(model.losses) / test_dataset_size
            elbo = -(negative_log_likelihood + l2_loss + kl)

            if dataset_name == 'clean':
                metrics['test/negative_log_likelihood'].update_state(
                    negative_log_likelihood)
                metrics['test/kl'].update_state(kl)
                metrics['test/elbo'].update_state(elbo)
                metrics['test/accuracy'].update_state(labels, probs)
                metrics['test/ece'].update_state(labels, probs)
            else:
                corrupt_metrics['test/nll_{}'.format(
                    dataset_name)].update_state(negative_log_likelihood)
                corrupt_metrics['test/kl_{}'.format(
                    dataset_name)].update_state(kl)
                corrupt_metrics['test/elbo_{}'.format(
                    dataset_name)].update_state(elbo)
                corrupt_metrics['test/accuracy_{}'.format(
                    dataset_name)].update_state(labels, probs)
                corrupt_metrics['test/ece_{}'.format(
                    dataset_name)].update_state(labels, probs)

        strategy.run(step_fn, args=(next(iterator), ))

    train_iterator = iter(train_dataset)
    start_time = time.time()
    for epoch in range(initial_epoch,
                       FLAGS.train_epochs + FLAGS.refining_epochs):
        logging.info('Starting to run epoch: %s', epoch)
        if epoch in np.linspace(FLAGS.train_epochs,
                                FLAGS.train_epochs + FLAGS.refining_epochs,
                                FLAGS.num_auxiliary_variables,
                                dtype=int):
            logging.info('Sampling auxiliary variables with ratio %f',
                         FLAGS.auxiliary_variance_ratio)
            refining.sample_rank1_auxiliaries(model,
                                              FLAGS.auxiliary_variance_ratio)
            if FLAGS.freeze_weights_during_refining:
                refining.freeze_rank1_weights(model)

        for step in range(steps_per_epoch):
            train_step(train_iterator)

            current_step = epoch * steps_per_epoch + (step + 1)
            max_steps = steps_per_epoch * (FLAGS.train_epochs +
                                           FLAGS.refining_epochs)
            time_elapsed = time.time() - start_time
            steps_per_sec = float(current_step) / time_elapsed
            eta_seconds = (max_steps - current_step) / steps_per_sec
            message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
                       'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
                           current_step / max_steps, epoch + 1,
                           FLAGS.train_epochs + FLAGS.refining_epochs,
                           steps_per_sec, eta_seconds / 60, time_elapsed / 60))
            if step % 20 == 0:
                logging.info(message)

        datasets_to_evaluate = {'clean': test_datasets['clean']}
        if (FLAGS.corruptions_interval > 0
                and (epoch + 1) % FLAGS.corruptions_interval == 0):
            datasets_to_evaluate = test_datasets
        for dataset_name, test_dataset in datasets_to_evaluate.items():
            test_iterator = iter(test_dataset)
            logging.info('Testing on dataset %s', dataset_name)
            for step in range(steps_per_eval):
                if step % 20 == 0:
                    logging.info('Starting to run eval step %s of epoch: %s',
                                 step, epoch)
                test_step(test_iterator, dataset_name)
            logging.info('Done with testing on %s', dataset_name)

        corrupt_results = {}
        if (FLAGS.corruptions_interval > 0
                and (epoch + 1) % FLAGS.corruptions_interval == 0):
            corrupt_results = utils.aggregate_corrupt_metrics(
                corrupt_metrics, corruption_types, max_intensity)

        logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
                     metrics['train/loss'].result(),
                     metrics['train/accuracy'].result() * 100)
        logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
                     metrics['test/negative_log_likelihood'].result(),
                     metrics['test/accuracy'].result() * 100)
        if FLAGS.ensemble_size > 1:
            for i in range(FLAGS.ensemble_size):
                logging.info(
                    'Member %d Test Loss: %.4f, Accuracy: %.2f%%', i,
                    metrics['test/nll_member_{}'.format(i)].result(),
                    metrics['test/accuracy_member_{}'.format(i)].result() *
                    100)
        total_results = {
            name: metric.result()
            for name, metric in metrics.items()
        }
        total_results.update(corrupt_results)
        with summary_writer.as_default():
            for name, result in total_results.items():
                tf.summary.scalar(name, result, step=epoch + 1)

        for metric in metrics.values():
            metric.reset_states()

        if (FLAGS.checkpoint_interval > 0
                and (epoch + 1) % FLAGS.checkpoint_interval == 0):
            checkpoint_name = checkpoint.save(
                os.path.join(FLAGS.output_dir, 'checkpoint'))
            logging.info('Saved checkpoint to %s', checkpoint_name)

    final_checkpoint_name = checkpoint.save(
        os.path.join(FLAGS.output_dir, 'checkpoint'))
    logging.info('Saved last checkpoint to %s', final_checkpoint_name)
Exemplo n.º 4
0
def main(argv):
    del argv  # unused arg
    tf.io.gfile.makedirs(FLAGS.output_dir)
    logging.info('Saving checkpoints at %s', FLAGS.output_dir)
    tf.random.set_seed(FLAGS.seed)

    if FLAGS.use_gpu:
        logging.info('Use GPU')
        strategy = tf.distribute.MirroredStrategy()
    else:
        logging.info('Use TPU at %s',
                     FLAGS.tpu if FLAGS.tpu is not None else 'local')
        resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
            tpu=FLAGS.tpu)
        tf.config.experimental_connect_to_cluster(resolver)
        tf.tpu.experimental.initialize_tpu_system(resolver)
        strategy = tf.distribute.TPUStrategy(resolver)

    aug_params = {
        'augmix': FLAGS.augmix,
        'aug_count': FLAGS.aug_count,
        'augmix_depth': FLAGS.augmix_depth,
        'augmix_prob_coeff': FLAGS.augmix_prob_coeff,
        'augmix_width': FLAGS.augmix_width,
        'ensemble_size': 1,
        'mixup_alpha': FLAGS.mixup_alpha,
        'adaptive_mixup': FLAGS.adaptive_mixup,
        'random_augment': FLAGS.random_augment,
        'forget_mixup': FLAGS.forget_mixup,
        'num_cores': FLAGS.num_cores,
        'threshold': FLAGS.forget_threshold,
    }
    batch_size = (FLAGS.per_core_batch_size * FLAGS.num_cores //
                  FLAGS.num_dropout_samples_training)
    train_input_fn = data_utils.load_input_fn(
        split=tfds.Split.TRAIN,
        name=FLAGS.dataset,
        batch_size=batch_size,
        use_bfloat16=FLAGS.use_bfloat16,
        proportion=FLAGS.train_proportion,
        validation_set=FLAGS.validation,
        aug_params=aug_params)
    if FLAGS.validation:
        validation_input_fn = data_utils.load_input_fn(
            split=tfds.Split.VALIDATION,
            name=FLAGS.dataset,
            batch_size=FLAGS.per_core_batch_size,
            use_bfloat16=FLAGS.use_bfloat16,
            validation_set=True)
        val_dataset = strategy.experimental_distribute_datasets_from_function(
            validation_input_fn)
    clean_test_dataset = utils.load_dataset(
        split=tfds.Split.TEST,
        name=FLAGS.dataset,
        batch_size=FLAGS.per_core_batch_size * FLAGS.num_cores,
        use_bfloat16=FLAGS.use_bfloat16)
    train_dataset = strategy.experimental_distribute_dataset(train_input_fn())
    test_datasets = {
        'clean': strategy.experimental_distribute_dataset(clean_test_dataset),
    }
    if FLAGS.corruptions_interval > 0:
        if FLAGS.dataset == 'cifar10':
            load_c_dataset = utils.load_cifar10_c
        else:
            load_c_dataset = functools.partial(utils.load_cifar100_c,
                                               path=FLAGS.cifar100_c_path)
        corruption_types, max_intensity = utils.load_corrupted_test_info(
            FLAGS.dataset)
        for corruption in corruption_types:
            for intensity in range(1, max_intensity + 1):
                dataset = load_c_dataset(corruption_name=corruption,
                                         corruption_intensity=intensity,
                                         batch_size=FLAGS.per_core_batch_size *
                                         FLAGS.num_cores,
                                         use_bfloat16=FLAGS.use_bfloat16)
                test_datasets['{0}_{1}'.format(corruption, intensity)] = (
                    strategy.experimental_distribute_dataset(dataset))

    ds_info = tfds.builder(FLAGS.dataset).info
    batch_size = (FLAGS.per_core_batch_size * FLAGS.num_cores //
                  FLAGS.num_dropout_samples_training)
    num_train_examples = ds_info.splits['train'].num_examples
    # Train_proportion is a float so need to convert steps_per_epoch to int.
    if FLAGS.validation:
        # TODO(ywenxu): Remove hard-coding validation images.
        steps_per_epoch = int(
            (num_train_examples * FLAGS.train_proportion - 2500) // batch_size)
        steps_per_val = 2500 // (FLAGS.per_core_batch_size * FLAGS.num_cores)
    else:
        steps_per_epoch = int(
            num_train_examples * FLAGS.train_proportion) // batch_size
    steps_per_eval = ds_info.splits['test'].num_examples // batch_size
    num_classes = ds_info.features['label'].num_classes

    if FLAGS.use_bfloat16:
        policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
        tf.keras.mixed_precision.experimental.set_policy(policy)

    summary_writer = tf.summary.create_file_writer(
        os.path.join(FLAGS.output_dir, 'summaries'))

    with strategy.scope():
        logging.info('Building ResNet model')
        if FLAGS.use_spec_norm:
            logging.info('Use Spectral Normalization with norm bound %.2f',
                         FLAGS.spec_norm_bound)
        if FLAGS.use_gp_layer:
            logging.info('Use GP layer with hidden units %d',
                         FLAGS.gp_hidden_dim)

        model = ub.models.wide_resnet_sngp(
            input_shape=ds_info.features['image'].shape,
            batch_size=batch_size,
            depth=28,
            width_multiplier=10,
            num_classes=num_classes,
            l2=FLAGS.l2,
            use_mc_dropout=FLAGS.use_mc_dropout,
            dropout_rate=FLAGS.dropout_rate,
            use_gp_layer=FLAGS.use_gp_layer,
            gp_input_dim=FLAGS.gp_input_dim,
            gp_hidden_dim=FLAGS.gp_hidden_dim,
            gp_scale=FLAGS.gp_scale,
            gp_bias=FLAGS.gp_bias,
            gp_input_normalization=FLAGS.gp_input_normalization,
            gp_cov_discount_factor=FLAGS.gp_cov_discount_factor,
            gp_cov_ridge_penalty=FLAGS.gp_cov_ridge_penalty,
            use_spec_norm=FLAGS.use_spec_norm,
            spec_norm_iteration=FLAGS.spec_norm_iteration,
            spec_norm_bound=FLAGS.spec_norm_bound)
        logging.info('Model input shape: %s', model.input_shape)
        logging.info('Model output shape: %s', model.output_shape)
        logging.info('Model number of weights: %s', model.count_params())
        # Linearly scale learning rate and the decay epochs by vanilla settings.
        base_lr = FLAGS.base_learning_rate * batch_size / 128
        lr_decay_epochs = [(int(start_epoch_str) * FLAGS.train_epochs) // 200
                           for start_epoch_str in FLAGS.lr_decay_epochs]
        lr_schedule = utils.LearningRateSchedule(
            steps_per_epoch,
            base_lr,
            decay_ratio=FLAGS.lr_decay_ratio,
            decay_epochs=lr_decay_epochs,
            warmup_epochs=FLAGS.lr_warmup_epochs)
        optimizer = tf.keras.optimizers.SGD(lr_schedule,
                                            momentum=0.9,
                                            nesterov=True)
        metrics = {
            'train/negative_log_likelihood': tf.keras.metrics.Mean(),
            'train/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
            'train/loss': tf.keras.metrics.Mean(),
            'train/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
            'test/negative_log_likelihood': tf.keras.metrics.Mean(),
            'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
            'test/ece': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
            'test/stddev': tf.keras.metrics.Mean(),
        }
        if FLAGS.corruptions_interval > 0:
            corrupt_metrics = {}
            for intensity in range(1, max_intensity + 1):
                for corruption in corruption_types:
                    dataset_name = '{0}_{1}'.format(corruption, intensity)
                    corrupt_metrics['test/nll_{}'.format(dataset_name)] = (
                        tf.keras.metrics.Mean())
                    corrupt_metrics['test/accuracy_{}'.format(
                        dataset_name)] = (
                            tf.keras.metrics.SparseCategoricalAccuracy())
                    corrupt_metrics['test/ece_{}'.format(dataset_name)] = (
                        um.ExpectedCalibrationError(num_bins=FLAGS.num_bins))
                    corrupt_metrics['test/stddev_{}'.format(dataset_name)] = (
                        tf.keras.metrics.Mean())

        checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
        latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
        initial_epoch = 0
        if latest_checkpoint:
            # checkpoint.restore must be within a strategy.scope() so that optimizer
            # slot variables are mirrored.
            checkpoint.restore(latest_checkpoint)
            logging.info('Loaded checkpoint %s', latest_checkpoint)
            initial_epoch = optimizer.iterations.numpy() // steps_per_epoch

    @tf.function
    def train_step(iterator):
        """Training StepFn."""
        def step_fn(inputs):
            """Per-Replica StepFn."""
            if FLAGS.forget_mixup:
                images, labels, idx = inputs
            else:
                images, labels = inputs
            if FLAGS.augmix and FLAGS.aug_count >= 1:
                # Index 0 at augmix preprocessing is the unperturbed image.
                images = images[:, 1, ...]
                # This is for the case of combining AugMix and Mixup.
                if FLAGS.mixup_alpha > 0:
                    labels = tf.split(labels, FLAGS.aug_count + 1, axis=0)[1]

            images = tf.tile(images,
                             [FLAGS.num_dropout_samples_training, 1, 1, 1])
            if FLAGS.mixup_alpha > 0:
                labels = tf.tile(labels,
                                 [FLAGS.num_dropout_samples_training, 1])
            else:
                labels = tf.tile(labels, [FLAGS.num_dropout_samples_training])

            with tf.GradientTape() as tape:
                logits, _ = model(images, training=True)
                if FLAGS.use_bfloat16:
                    logits = tf.cast(logits, tf.float32)
                if FLAGS.mixup_alpha > 0:
                    negative_log_likelihood = tf.reduce_mean(
                        tf.keras.losses.categorical_crossentropy(
                            labels, logits, from_logits=True))
                else:
                    negative_log_likelihood = tf.reduce_mean(
                        tf.keras.losses.sparse_categorical_crossentropy(
                            labels, logits, from_logits=True))

                l2_loss = sum(model.losses)
                loss = negative_log_likelihood + l2_loss
                # Scale the loss given the TPUStrategy will reduce sum all gradients.
                scaled_loss = loss / strategy.num_replicas_in_sync

            grads = tape.gradient(scaled_loss, model.trainable_variables)
            optimizer.apply_gradients(zip(grads, model.trainable_variables))

            probs = tf.nn.softmax(logits)
            if FLAGS.mixup_alpha > 0:
                labels = tf.argmax(labels, axis=-1)
            metrics['train/ece'].update_state(labels, probs)
            metrics['train/loss'].update_state(loss)
            metrics['train/negative_log_likelihood'].update_state(
                negative_log_likelihood)
            metrics['train/accuracy'].update_state(labels, logits)
            if FLAGS.forget_mixup:
                train_predictions = tf.argmax(probs, -1)
                labels = tf.cast(labels, train_predictions.dtype)
                # For each ensemble member (1 here), we accumulate the accuracy counts.
                accuracy_counts = tf.cast(
                    tf.reshape((train_predictions == labels), [1, -1]),
                    tf.float32)
                return accuracy_counts, idx

        if FLAGS.forget_mixup:
            return strategy.run(step_fn, args=(next(iterator), ))
        else:
            strategy.run(step_fn, args=(next(iterator), ))

    @tf.function
    def test_step(iterator, dataset_name):
        """Evaluation StepFn."""
        def step_fn(inputs):
            """Per-Replica StepFn."""
            images, labels = inputs

            logits_list = []
            stddev_list = []
            for _ in range(FLAGS.num_dropout_samples):
                logits, covmat = model(images, training=False)
                if FLAGS.use_bfloat16:
                    logits = tf.cast(logits, tf.float32)
                logits = ed.layers.utils.mean_field_logits(
                    logits,
                    covmat,
                    mean_field_factor=FLAGS.gp_mean_field_factor)
                stddev = tf.sqrt(tf.linalg.diag_part(covmat))

                stddev_list.append(stddev)
                logits_list.append(logits)

            # Logits dimension is (num_samples, batch_size, num_classes).
            logits_list = tf.stack(logits_list, axis=0)
            stddev_list = tf.stack(stddev_list, axis=0)

            stddev = tf.reduce_mean(stddev_list, axis=0)
            probs_list = tf.nn.softmax(logits_list)
            probs = tf.reduce_mean(probs_list, axis=0)

            labels_broadcasted = tf.broadcast_to(
                labels, [FLAGS.num_dropout_samples, labels.shape[0]])
            log_likelihoods = -tf.keras.losses.sparse_categorical_crossentropy(
                labels_broadcasted, logits_list, from_logits=True)
            negative_log_likelihood = tf.reduce_mean(
                -tf.reduce_logsumexp(log_likelihoods, axis=[0]) +
                tf.math.log(float(FLAGS.num_dropout_samples)))

            if dataset_name == 'clean':
                metrics['test/negative_log_likelihood'].update_state(
                    negative_log_likelihood)
                metrics['test/accuracy'].update_state(labels, probs)
                metrics['test/ece'].update_state(labels, probs)
                metrics['test/stddev'].update_state(stddev)
            elif dataset_name != 'validation':
                corrupt_metrics['test/nll_{}'.format(
                    dataset_name)].update_state(negative_log_likelihood)
                corrupt_metrics['test/accuracy_{}'.format(
                    dataset_name)].update_state(labels, probs)
                corrupt_metrics['test/ece_{}'.format(
                    dataset_name)].update_state(labels, probs)
                corrupt_metrics['test/stddev_{}'.format(
                    dataset_name)].update_state(stddev)

            if dataset_name == 'validation':
                return tf.reshape(probs, [1, -1, num_classes]), labels

        if dataset_name == 'validation':
            return strategy.run(step_fn, args=(next(iterator), ))
        else:
            strategy.run(step_fn, args=(next(iterator), ))

    metrics.update({'test/ms_per_example': tf.keras.metrics.Mean()})

    train_iterator = iter(train_dataset)
    forget_counts_history = []
    start_time = time.time()
    for epoch in range(initial_epoch, FLAGS.train_epochs):
        logging.info('Starting to run epoch: %s', epoch)
        acc_counts_list = []
        idx_list = []
        for step in range(steps_per_epoch):
            if FLAGS.forget_mixup:
                temp_accuracy_counts, temp_idx = train_step(train_iterator)
                acc_counts_list.append(temp_accuracy_counts)
                idx_list.append(temp_idx)
            else:
                train_step(train_iterator)

            current_step = epoch * steps_per_epoch + (step + 1)
            max_steps = steps_per_epoch * FLAGS.train_epochs
            time_elapsed = time.time() - start_time
            steps_per_sec = float(current_step) / time_elapsed
            eta_seconds = (max_steps - current_step) / steps_per_sec
            message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
                       'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
                           current_step / max_steps, epoch + 1,
                           FLAGS.train_epochs, steps_per_sec, eta_seconds / 60,
                           time_elapsed / 60))
            if step % 20 == 0:
                logging.info(message)

        # Only one of the forget_mixup and adaptive_mixup can be true.
        if FLAGS.forget_mixup:
            current_acc = [
                tf.concat(list(acc_counts_list[i].values), axis=1)
                for i in range(len(acc_counts_list))
            ]
            total_idx = [
                tf.concat(list(idx_list[i].values), axis=0)
                for i in range(len(idx_list))
            ]
            current_acc = tf.cast(tf.concat(current_acc, axis=1), tf.int32)
            total_idx = tf.concat(total_idx, axis=0)

            current_forget_path = os.path.join(FLAGS.output_dir,
                                               'forget_counts.npy')
            last_acc_path = os.path.join(FLAGS.output_dir, 'last_acc.npy')
            if epoch == 0:
                forget_counts = tf.zeros([1, num_train_examples],
                                         dtype=tf.int32)
                last_acc = tf.zeros([1, num_train_examples], dtype=tf.int32)
            else:
                if 'last_acc' not in locals():
                    with tf.io.gfile.GFile(last_acc_path, 'rb') as f:
                        last_acc = np.load(f)
                    last_acc = tf.cast(tf.convert_to_tensor(last_acc),
                                       tf.int32)
                if 'forget_counts' not in locals():
                    with tf.io.gfile.GFile(current_forget_path, 'rb') as f:
                        forget_counts = np.load(f)
                    forget_counts = tf.cast(
                        tf.convert_to_tensor(forget_counts), tf.int32)

            selected_last_acc = tf.gather(last_acc, total_idx, axis=1)
            forget_this_epoch = tf.cast(current_acc < selected_last_acc,
                                        tf.int32)
            forget_this_epoch = tf.transpose(forget_this_epoch)
            target_shape = tf.constant([num_train_examples, 1])
            current_forget_counts = tf.scatter_nd(
                tf.reshape(total_idx, [-1, 1]), forget_this_epoch,
                target_shape)
            current_forget_counts = tf.transpose(current_forget_counts)
            acc_this_epoch = tf.transpose(current_acc)
            last_acc = tf.scatter_nd(tf.reshape(total_idx, [-1, 1]),
                                     acc_this_epoch, target_shape)
            # This is lower bound of true acc.
            last_acc = tf.transpose(last_acc)

            # TODO(ywenxu): We count the dropped examples as forget. Fix this later.
            forget_counts += current_forget_counts
            forget_counts_history.append(forget_counts)
            logging.info('forgetting counts')
            logging.info(tf.stack(forget_counts_history, 0))
            with tf.io.gfile.GFile(
                    os.path.join(FLAGS.output_dir,
                                 'forget_counts_history.npy'), 'wb') as f:
                np.save(f, tf.stack(forget_counts_history, 0).numpy())
            with tf.io.gfile.GFile(current_forget_path, 'wb') as f:
                np.save(f, forget_counts.numpy())
            with tf.io.gfile.GFile(last_acc_path, 'wb') as f:
                np.save(f, last_acc.numpy())
            aug_params['forget_counts_dir'] = current_forget_path

            train_input_fn = data_utils.load_input_fn(
                split=tfds.Split.TRAIN,
                name=FLAGS.dataset,
                batch_size=batch_size,
                use_bfloat16=FLAGS.use_bfloat16,
                validation_set=FLAGS.validation,
                aug_params=aug_params)
            train_dataset = strategy.experimental_distribute_dataset(
                train_input_fn())
            train_iterator = iter(train_dataset)

        elif FLAGS.adaptive_mixup:
            val_iterator = iter(val_dataset)
            logging.info('Testing on validation dataset')
            predictions_list = []
            labels_list = []
            for step in range(steps_per_val):
                temp_predictions, temp_labels = test_step(
                    val_iterator, 'validation')
                predictions_list.append(temp_predictions)
                labels_list.append(temp_labels)
            predictions = [
                tf.concat(list(predictions_list[i].values), axis=1)
                for i in range(len(predictions_list))
            ]
            labels = [
                tf.concat(list(labels_list[i].values), axis=0)
                for i in range(len(labels_list))
            ]
            predictions = tf.concat(predictions, axis=1)
            labels = tf.cast(tf.concat(labels, axis=0), tf.int64)

            def compute_acc_conf(preds, label, focus_class):
                class_preds = tf.boolean_mask(preds,
                                              label == focus_class,
                                              axis=1)
                class_pred_labels = tf.argmax(class_preds, axis=-1)
                confidence = tf.reduce_mean(
                    tf.reduce_max(class_preds, axis=-1), -1)
                accuracy = tf.reduce_mean(tf.cast(
                    class_pred_labels == focus_class, tf.float32),
                                          axis=-1)
                return accuracy - confidence

            calibration_per_class = [
                compute_acc_conf(predictions, labels, i)
                for i in range(num_classes)
            ]
            calibration_per_class = tf.stack(calibration_per_class, axis=1)
            logging.info('calibration per class')
            logging.info(calibration_per_class)
            mixup_coeff = tf.where(calibration_per_class > 0, 1.0,
                                   FLAGS.mixup_alpha)
            mixup_coeff = tf.clip_by_value(mixup_coeff, 0, 1)
            logging.info('mixup coeff')
            logging.info(mixup_coeff)
            aug_params['mixup_coeff'] = mixup_coeff
            train_input_fn = data_utils.load_input_fn(
                split=tfds.Split.TRAIN,
                name=FLAGS.dataset,
                batch_size=batch_size,
                use_bfloat16=FLAGS.use_bfloat16,
                validation_set=True,
                aug_params=aug_params)
            train_dataset = strategy.experimental_distribute_dataset(
                train_input_fn())
            train_iterator = iter(train_dataset)

        datasets_to_evaluate = {'clean': test_datasets['clean']}
        if (FLAGS.corruptions_interval > 0
                and (epoch + 1) % FLAGS.corruptions_interval == 0):
            datasets_to_evaluate = test_datasets
        for dataset_name, test_dataset in datasets_to_evaluate.items():
            test_iterator = iter(test_dataset)
            logging.info('Testing on dataset %s', dataset_name)
            for step in range(steps_per_eval):
                if step % 20 == 0:
                    logging.info('Starting to run eval step %s of epoch: %s',
                                 step, epoch)
                test_start_time = time.time()
                test_step(test_iterator, dataset_name)
                ms_per_example = (time.time() -
                                  test_start_time) * 1e6 / batch_size
                metrics['test/ms_per_example'].update_state(ms_per_example)

            logging.info('Done with testing on %s', dataset_name)

        corrupt_results = {}
        if (FLAGS.corruptions_interval > 0
                and (epoch + 1) % FLAGS.corruptions_interval == 0):
            corrupt_results = utils.aggregate_corrupt_metrics(
                corrupt_metrics, corruption_types, max_intensity)

        logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
                     metrics['train/loss'].result(),
                     metrics['train/accuracy'].result() * 100)
        logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
                     metrics['test/negative_log_likelihood'].result(),
                     metrics['test/accuracy'].result() * 100)
        total_results = {
            name: metric.result()
            for name, metric in metrics.items()
        }
        total_results.update(corrupt_results)
        with summary_writer.as_default():
            for name, result in total_results.items():
                tf.summary.scalar(name, result, step=epoch + 1)

        for metric in metrics.values():
            metric.reset_states()

        if (FLAGS.checkpoint_interval > 0
                and (epoch + 1) % FLAGS.checkpoint_interval == 0):
            checkpoint_name = checkpoint.save(
                os.path.join(FLAGS.output_dir, 'checkpoint'))
            logging.info('Saved checkpoint to %s', checkpoint_name)

    final_checkpoint_name = checkpoint.save(
        os.path.join(FLAGS.output_dir, 'checkpoint'))
    logging.info('Saved last checkpoint to %s', final_checkpoint_name)