Merge pull request #2046 from jplu/tf2-ner-example

Add NER TF2 example.

docs/source/main_classes/optimizer_schedules.rst

@@ -5,6 +5,7 @@ The ``.optimization`` module provides:
 
 - an optimizer with weight decay fixed that can be used to fine-tuned models, and
 - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
+- a gradient accumulation class to accumulate the gradients of multiple batches
 
 ``AdamW``
 ~~~~~~~~~~~~~~~~
@@ -12,6 +13,15 @@ The ``.optimization`` module provides:
 .. autoclass:: transformers.AdamW
     :members:
 
+``AdamWeightDecay``
+~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AdamWeightDecay
+    :members:
+
+.. autofunction:: transformers.create_optimizer
+    :members:
+
 Schedules
 ----------------------------------------------------
 
@@ -49,3 +59,17 @@ Learning Rate Schedules
 .. image:: /imgs/warmup_linear_schedule.png
     :target: /imgs/warmup_linear_schedule.png
     :alt:
+
+``Warmup``
+~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Warmup
+    :members:
+
+Gradient Strategies
+----------------------------------------------------
+
+``GradientAccumulator``
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.GradientAccumulator

examples/README.md

@@ -467,7 +467,8 @@ Training with the previously defined hyper-parameters yields the following resul
 
 ## Named Entity Recognition
 
-Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
+Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
+[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
 This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
 Details and results for the fine-tuning provided by @stefan-it.
 
@@ -512,7 +513,7 @@ The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so
 cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
 ```
 
-### Training
+### Prepare the run
 
 Additional environment variables must be set:
 
@@ -524,6 +525,8 @@ export SAVE_STEPS=750
 export SEED=1
 ```
 
+### Run the Pytorch version
+
 To start training, just run:
 
 ```bash
@@ -544,7 +547,7 @@ python3 run_ner.py --data_dir ./ \
 
 If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
 
-### Evaluation
+#### Evaluation
 
 Evaluation on development dataset outputs the following for our example:
 
@@ -566,7 +569,7 @@ On the test dataset the following results could be achieved:
 10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
 ```
 
-### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
+#### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
 
 Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
 
@@ -576,6 +579,72 @@ Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) a
 | `roberta-large`                  | 95.96 | 91.87
 | `distilbert-base-uncased` | 94.34 | 90.32
 
+### Run the Tensorflow 2 version
+
+To start training, just run:
+
+```bash
+python3 run_tf_ner.py --data_dir ./ \
+--model_type bert \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_device_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+```
+
+Such as the Pytorch version, if your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
+
+#### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
+```bash
+           precision    recall  f1-score   support
+
+ LOCderiv     0.7619    0.6154    0.6809        52
+  PERpart     0.8724    0.8997    0.8858      4057
+  OTHpart     0.9360    0.9466    0.9413       711
+  ORGpart     0.7015    0.6989    0.7002       269
+  LOCpart     0.7668    0.8488    0.8057       496
+      LOC     0.8745    0.9191    0.8963       235
+ ORGderiv     0.7723    0.8571    0.8125        91
+ OTHderiv     0.4800    0.6667    0.5581        18
+      OTH     0.5789    0.6875    0.6286        16
+ PERderiv     0.5385    0.3889    0.4516        18
+      PER     0.5000    0.5000    0.5000         2
+      ORG     0.0000    0.0000    0.0000         3
+
+micro avg     0.8574    0.8862    0.8715      5968
+macro avg     0.8575    0.8862    0.8713      5968
+```
+
+On the test dataset the following results could be achieved:
+```bash
+           precision    recall  f1-score   support
+
+  PERpart     0.8847    0.8944    0.8896      9397
+  OTHpart     0.9376    0.9353    0.9365      1639
+  ORGpart     0.7307    0.7044    0.7173       697
+      LOC     0.9133    0.9394    0.9262       561
+  LOCpart     0.8058    0.8157    0.8107      1150
+      ORG     0.0000    0.0000    0.0000         8
+ OTHderiv     0.5882    0.4762    0.5263        42
+ PERderiv     0.6571    0.5227    0.5823        44
+      OTH     0.4906    0.6667    0.5652        39
+ ORGderiv     0.7016    0.7791    0.7383       172
+ LOCderiv     0.8256    0.6514    0.7282       109
+      PER     0.0000    0.0000    0.0000        11
+
+micro avg     0.8722    0.8774    0.8748     13869
+macro avg     0.8712    0.8774    0.8740     13869
+```
+
 ## Abstractive summarization
 
 Based on the script

examples/run_tf_ner.py

@@ -0,0 +1,615 @@
+# coding=utf-8
+import datetime
+import os
+import math
+import glob
+import re
+import tensorflow as tf
+import collections
+import numpy as np
+from seqeval import metrics
+import _pickle as pickle
+from absl import logging
+from transformers import TF2_WEIGHTS_NAME, BertConfig, BertTokenizer, TFBertForTokenClassification
+from transformers import RobertaConfig, RobertaTokenizer, TFRobertaForTokenClassification
+from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForTokenClassification
+from transformers import create_optimizer, GradientAccumulator
+from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
+from fastprogress import master_bar, progress_bar
+from absl import flags
+from absl import app
+
+
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
+    ())
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, TFBertForTokenClassification, BertTokenizer),
+    "roberta": (RobertaConfig, TFRobertaForTokenClassification, RobertaTokenizer),
+    "distilbert": (DistilBertConfig, TFDistilBertForTokenClassification, DistilBertTokenizer)
+}
+
+
+flags.DEFINE_string(
+    "data_dir", None,
+    "The input data dir. Should contain the .conll files (or other data files) "
+    "for the task.")
+
+flags.DEFINE_string(
+    "model_type", None,
+    "Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+
+flags.DEFINE_string(
+    "model_name_or_path", None,
+    "Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+
+flags.DEFINE_string(
+    "output_dir", None,
+    "The output directory where the model checkpoints will be written.")
+
+flags.DEFINE_string(
+    "labels", "",
+    "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
+
+flags.DEFINE_string(
+    "config_name", "",
+    "Pretrained config name or path if not the same as model_name")
+
+flags.DEFINE_string(
+    "tokenizer_name", "",
+    "Pretrained tokenizer name or path if not the same as model_name")
+
+flags.DEFINE_string(
+    "cache_dir", "",
+    "Where do you want to store the pre-trained models downloaded from s3")
+
+flags.DEFINE_integer(
+    "max_seq_length", 128,
+    "The maximum total input sentence length after tokenization. "
+    "Sequences longer than this will be truncated, sequences shorter "
+    "will be padded.")
+
+flags.DEFINE_string(
+    "tpu", None,
+    "The Cloud TPU to use for training. This should be either the name "
+    "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 "
+    "url.")
+
+flags.DEFINE_integer(
+    "num_tpu_cores", 8,
+    "Total number of TPU cores to use.")
+
+flags.DEFINE_boolean(
+    "do_train", False,
+    "Whether to run training.")
+
+flags.DEFINE_boolean(
+    "do_eval", False,
+    "Whether to run eval on the dev set.")
+
+flags.DEFINE_boolean(
+    "do_predict", False,
+    "Whether to run predictions on the test set.")
+
+flags.DEFINE_boolean(
+    "evaluate_during_training", False,
+    "Whether to run evaluation during training at each logging step.")
+
+flags.DEFINE_boolean(
+    "do_lower_case", False,
+    "Set this flag if you are using an uncased model.")
+
+flags.DEFINE_integer(
+    "per_device_train_batch_size", 8,
+    "Batch size per GPU/CPU/TPU for training.")
+
+flags.DEFINE_integer(
+    "per_device_eval_batch_size", 8,
+    "Batch size per GPU/CPU/TPU for evaluation.")
+
+flags.DEFINE_integer(
+    "gradient_accumulation_steps", 1,
+    "Number of updates steps to accumulate before performing a backward/update pass.")
+
+flags.DEFINE_float(
+    "learning_rate", 5e-5,
+    "The initial learning rate for Adam.")
+
+flags.DEFINE_float(
+    "weight_decay", 0.0,
+    "Weight decay if we apply some.")
+
+flags.DEFINE_float(
+    "adam_epsilon", 1e-8,
+    "Epsilon for Adam optimizer.")
+
+flags.DEFINE_float(
+    "max_grad_norm", 1.0,
+    "Max gradient norm.")
+
+flags.DEFINE_integer(
+    "num_train_epochs", 3,
+    "Total number of training epochs to perform.")
+
+flags.DEFINE_integer(
+    "max_steps", -1,
+    "If > 0: set total number of training steps to perform. Override num_train_epochs.")
+
+flags.DEFINE_integer(
+    "warmup_steps", 0,
+    "Linear warmup over warmup_steps.")
+
+flags.DEFINE_integer(
+    "logging_steps", 50,
+    "Log every X updates steps.")
+
+flags.DEFINE_integer(
+    "save_steps", 50,
+    "Save checkpoint every X updates steps.")
+
+flags.DEFINE_boolean(
+    "eval_all_checkpoints", False,
+    "Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+
+flags.DEFINE_boolean(
+    "no_cuda", False,
+    "Avoid using CUDA when available")
+
+flags.DEFINE_boolean(
+    "overwrite_output_dir", False,
+    "Overwrite the content of the output directory")
+
+flags.DEFINE_boolean(
+    "overwrite_cache", False,
+    "Overwrite the cached training and evaluation sets")
+
+flags.DEFINE_integer(
+    "seed", 42,
+    "random seed for initialization")
+
+flags.DEFINE_boolean(
+    "fp16", False,
+    "Whether to use 16-bit (mixed) precision instead of 32-bit")
+
+flags.DEFINE_string(
+    "gpus", "0",
+    "Comma separated list of gpus devices. If only one, switch to single "
+    "gpu strategy, if None takes all the gpus available.")
+
+
+def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id):
+    if args['max_steps'] > 0:
+        num_train_steps = args['max_steps'] * args['gradient_accumulation_steps']
+        args['num_train_epochs'] = 1
+    else:
+        num_train_steps = math.ceil(num_train_examples / train_batch_size) // args['gradient_accumulation_steps'] * args['num_train_epochs']
+
+    writer = tf.summary.create_file_writer("/tmp/mylogs")
+
+    with strategy.scope():
+        loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
+        optimizer = create_optimizer(args['learning_rate'], num_train_steps, args['warmup_steps'])
+
+        if args['fp16']:
+            optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(optimizer, 'dynamic')
+
+        loss_metric = tf.keras.metrics.Mean(name='loss', dtype=tf.float32)
+        gradient_accumulator = GradientAccumulator()
+        
+    logging.info("***** Running training *****")
+    logging.info("  Num examples = %d", num_train_examples)
+    logging.info("  Num Epochs = %d", args['num_train_epochs'])
+    logging.info("  Instantaneous batch size per device = %d", args['per_device_train_batch_size'])
+    logging.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                train_batch_size * args['gradient_accumulation_steps'])
+    logging.info("  Gradient Accumulation steps = %d", args['gradient_accumulation_steps'])
+    logging.info("  Total training steps = %d", num_train_steps)
+
+    model.summary()
+
+    @tf.function
+    def apply_gradients():
+        grads_and_vars = []
+
+        for gradient, variable in zip(gradient_accumulator.gradients, model.trainable_variables):
+            if gradient is not None:
+                scaled_gradient = gradient / (args['n_device'] * args['gradient_accumulation_steps'])
+                grads_and_vars.append((scaled_gradient, variable))
+            else:
+                grads_and_vars.append((gradient, variable))
+
+        optimizer.apply_gradients(grads_and_vars, args['max_grad_norm'])
+        gradient_accumulator.reset()
+
+    @tf.function
+    def train_step(train_features, train_labels):
+        def step_fn(train_features, train_labels):
+            inputs = {'attention_mask': train_features['input_mask'], 'training': True}
+
+            if args['model_type'] != "distilbert":
+                inputs["token_type_ids"] = train_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
+
+            with tf.GradientTape() as tape:
+                logits = model(train_features['input_ids'], **inputs)[0]
+                logits = tf.reshape(logits, (-1, len(labels) + 1))
+                active_loss = tf.reshape(train_features['input_mask'], (-1,))
+                active_logits = tf.boolean_mask(logits, active_loss)
+                train_labels = tf.reshape(train_labels, (-1,))
+                active_labels = tf.boolean_mask(train_labels, active_loss)
+                cross_entropy = loss_fct(active_labels, active_logits)
+                loss = tf.reduce_sum(cross_entropy) * (1.0 / train_batch_size)
+                grads = tape.gradient(loss, model.trainable_variables)
+
+                gradient_accumulator(grads)
+
+            return cross_entropy
+
+        per_example_losses = strategy.experimental_run_v2(step_fn, args=(train_features, train_labels))
+        mean_loss = strategy.reduce(tf.distribute.ReduceOp.MEAN, per_example_losses, axis=0)
+
+        return mean_loss
+
+    current_time = datetime.datetime.now()
+    train_iterator = master_bar(range(args['num_train_epochs']))
+    global_step = 0
+    logging_loss = 0.0
+
+    for epoch in train_iterator:
+        epoch_iterator = progress_bar(train_dataset, total=num_train_steps, parent=train_iterator, display=args['n_device'] > 1)
+        step = 1
+
+        with strategy.scope():
+            for train_features, train_labels in epoch_iterator:
+                loss = train_step(train_features, train_labels)
+
+                if step % args['gradient_accumulation_steps'] == 0:
+                    strategy.experimental_run_v2(apply_gradients)
+
+                    loss_metric(loss)
+
+                    global_step += 1
+
+                    if args['logging_steps'] > 0 and global_step % args['logging_steps'] == 0:
+                        # Log metrics
+                        if args['n_device'] == 1 and args['evaluate_during_training']:  # Only evaluate when single GPU otherwise metrics may not average well
+                            y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
+                            report = metrics.classification_report(y_true, y_pred, digits=4)
+                            
+                            logging.info("Eval at step " + str(global_step) + "\n" + report)
+                            logging.info("eval_loss: " + str(eval_loss))
+                            
+                            precision = metrics.precision_score(y_true, y_pred)
+                            recall = metrics.recall_score(y_true, y_pred)
+                            f1 = metrics.f1_score(y_true, y_pred)
+
+                            with writer.as_default():
+                                tf.summary.scalar("eval_loss", eval_loss, global_step)
+                                tf.summary.scalar("precision", precision, global_step)
+                                tf.summary.scalar("recall", recall, global_step)
+                                tf.summary.scalar("f1", f1, global_step)
+                        
+                        lr = optimizer.learning_rate
+                        learning_rate = lr(step)
+
+                        with writer.as_default():
+                            tf.summary.scalar("lr", learning_rate, global_step)
+                            tf.summary.scalar("loss", (loss_metric.result() - logging_loss) / args['logging_steps'], global_step)
+                        
+                        logging_loss = loss_metric.result()
+
+                    with writer.as_default():
+                        tf.summary.scalar("loss", loss_metric.result(), step=step)
+
+                    if args['save_steps'] > 0 and global_step % args['save_steps'] == 0:
+                        # Save model checkpoint
+                        output_dir = os.path.join(args['output_dir'], "checkpoint-{}".format(global_step))
+
+                        if not os.path.exists(output_dir):
+                            os.makedirs(output_dir)
+                        
+                        model.save_pretrained(output_dir)
+                        logging.info("Saving model checkpoint to %s", output_dir)
+                
+                train_iterator.child.comment = f'loss : {loss_metric.result()}'
+                step += 1
+
+        train_iterator.write(f'loss epoch {epoch + 1}: {loss_metric.result()}')
+
+        loss_metric.reset_states()
+
+    logging.info("  Training took time = {}".format(datetime.datetime.now() - current_time))
+
+
+def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode):
+    eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
+    eval_dataset, size = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode=mode)
+    eval_dataset = strategy.experimental_distribute_dataset(eval_dataset)
+    preds = None
+    num_eval_steps = math.ceil(size / eval_batch_size)
+    master = master_bar(range(1))
+    eval_iterator = progress_bar(eval_dataset, total=num_eval_steps, parent=master, display=args['n_device'] > 1)
+    loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
+    loss = 0.0
+
+    logging.info("***** Running evaluation *****")
+    logging.info("  Num examples = %d", size)
+    logging.info("  Batch size = %d", eval_batch_size)
+
+    for eval_features, eval_labels in eval_iterator:
+        inputs = {'attention_mask': eval_features['input_mask'], 'training': False}
+
+        if args['model_type'] != "distilbert":
+            inputs["token_type_ids"] = eval_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
+
+        with strategy.scope():
+            logits = model(eval_features['input_ids'], **inputs)[0]
+            tmp_logits = tf.reshape(logits, (-1, len(labels) + 1))
+            active_loss = tf.reshape(eval_features['input_mask'], (-1,))
+            active_logits = tf.boolean_mask(tmp_logits, active_loss)
+            tmp_eval_labels = tf.reshape(eval_labels, (-1,))
+            active_labels = tf.boolean_mask(tmp_eval_labels, active_loss)
+            cross_entropy = loss_fct(active_labels, active_logits)
+            loss += tf.reduce_sum(cross_entropy) * (1.0 / eval_batch_size)
+
+        if preds is None:
+            preds = logits.numpy()
+            label_ids = eval_labels.numpy()
+        else:
+            preds = np.append(preds, logits.numpy(), axis=0)
+            label_ids = np.append(label_ids, eval_labels.numpy(), axis=0)
+
+    preds = np.argmax(preds, axis=2)
+    y_pred = [[] for _ in range(label_ids.shape[0])]
+    y_true = [[] for _ in range(label_ids.shape[0])]
+    loss = loss / num_eval_steps
+
+    for i in range(label_ids.shape[0]):
+        for j in range(label_ids.shape[1]):
+            if label_ids[i, j] != pad_token_label_id:
+                y_pred[i].append(labels[preds[i, j] - 1])
+                y_true[i].append(labels[label_ids[i, j] - 1])
+
+    return y_true, y_pred, loss.numpy()
+
+
+def load_cache(cached_file, max_seq_length):
+    name_to_features = {
+        "input_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "input_mask": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "segment_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+        "label_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
+    }
+
+    def _decode_record(record):
+        example = tf.io.parse_single_example(record, name_to_features)
+        features = {}
+        features['input_ids'] = example['input_ids']
+        features['input_mask'] = example['input_mask']
+        features['segment_ids'] = example['segment_ids']
+
+        return features, example['label_ids']
+
+    d = tf.data.TFRecordDataset(cached_file)
+    d = d.map(_decode_record, num_parallel_calls=4)
+    count = d.reduce(0, lambda x, _: x + 1)
+
+    return d, count.numpy()
+
+
+def save_cache(features, cached_features_file):
+    writer = tf.io.TFRecordWriter(cached_features_file)
+
+    for (ex_index, feature) in enumerate(features):
+        if ex_index % 5000 == 0:
+            logging.info("Writing example %d of %d" % (ex_index, len(features)))
+
+        def create_int_feature(values):
+            f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))
+            return f
+
+        record_feature = collections.OrderedDict()
+        record_feature["input_ids"] = create_int_feature(feature.input_ids)
+        record_feature["input_mask"] = create_int_feature(feature.input_mask)
+        record_feature["segment_ids"] = create_int_feature(feature.segment_ids)
+        record_feature["label_ids"] = create_int_feature(feature.label_ids)
+
+        tf_example = tf.train.Example(features=tf.train.Features(feature=record_feature))
+
+        writer.write(tf_example.SerializeToString())
+
+    writer.close()
+
+
+def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, batch_size, mode):
+    drop_remainder = True if args['tpu'] or mode == 'train' else False
+
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(args['data_dir'], "cached_{}_{}_{}.tf_record".format(mode,
+        list(filter(None, args['model_name_or_path'].split("/"))).pop(),
+        str(args['max_seq_length'])))
+    if os.path.exists(cached_features_file) and not args['overwrite_cache']:
+        logging.info("Loading features from cached file %s", cached_features_file)
+        dataset, size = load_cache(cached_features_file, args['max_seq_length'])
+    else:
+        logging.info("Creating features from dataset file at %s", args['data_dir'])
+        examples = read_examples_from_file(args['data_dir'], mode)
+        features = convert_examples_to_features(examples, labels, args['max_seq_length'], tokenizer,
+                                                cls_token_at_end=bool(args['model_type'] in ["xlnet"]),
+                                                # xlnet has a cls token at the end
+                                                cls_token=tokenizer.cls_token,
+                                                cls_token_segment_id=2 if args['model_type'] in ["xlnet"] else 0,
+                                                sep_token=tokenizer.sep_token,
+                                                sep_token_extra=bool(args['model_type'] in ["roberta"]),
+                                                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                                                pad_on_left=bool(args['model_type'] in ["xlnet"]),
+                                                # pad on the left for xlnet
+                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
+                                                pad_token_segment_id=4 if args['model_type'] in ["xlnet"] else 0,
+                                                pad_token_label_id=pad_token_label_id
+                                                )
+        logging.info("Saving features into cached file %s", cached_features_file)
+        save_cache(features, cached_features_file)
+        dataset, size = load_cache(cached_features_file, args['max_seq_length'])
+
+    if mode == 'train':
+        dataset = dataset.repeat()
+        dataset = dataset.shuffle(buffer_size=8192, seed=args['seed'])
+
+    dataset = dataset.batch(batch_size, drop_remainder)
+    dataset = dataset.prefetch(buffer_size=batch_size)
+
+    return dataset, size
+
+
+def main(_):
+    logging.set_verbosity(logging.INFO)
+    args = flags.FLAGS.flag_values_dict()
+
+    if os.path.exists(args['output_dir']) and os.listdir(
+            args['output_dir']) and args['do_train'] and not args['overwrite_output_dir']:
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args['output_dir']))
+
+    if args['fp16']:
+        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
+
+    if args['tpu']:
+        resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=args['tpu'])
+        tf.config.experimental_connect_to_cluster(resolver)
+        tf.tpu.experimental.initialize_tpu_system(resolver)
+        strategy = tf.distribute.experimental.TPUStrategy(resolver)
+        args['n_device'] = args['num_tpu_cores']
+    elif len(args['gpus'].split(',')) > 1:
+        args['n_device'] = len([f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
+        strategy = tf.distribute.MirroredStrategy(devices=[f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
+    elif args['no_cuda']:
+        args['n_device'] = 1
+        strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
+    else:
+        args['n_device'] = len(args['gpus'].split(','))
+        strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" + args['gpus'].split(',')[0])
+
+    logging.warning("n_device: %s, distributed training: %s, 16-bits training: %s",
+                   args['n_device'], bool(args['n_device'] > 1), args['fp16'])
+
+    labels = get_labels(args['labels'])
+    num_labels = len(labels) + 1
+    pad_token_label_id = 0
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args['model_type']]
+    config = config_class.from_pretrained(args['config_name'] if args['config_name'] else args['model_name_or_path'],
+                                          num_labels=num_labels,
+                                          cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+
+    logging.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args['do_train']:
+        tokenizer = tokenizer_class.from_pretrained(args['tokenizer_name'] if args['tokenizer_name'] else args['model_name_or_path'],
+                                                    do_lower_case=args['do_lower_case'],
+                                                    cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+
+        with strategy.scope():
+            model = model_class.from_pretrained(args['model_name_or_path'],
+                                                from_pt=bool(".bin" in args['model_name_or_path']),
+                                                config=config,
+                                                cache_dir=args['cache_dir'] if args['cache_dir'] else None)
+            model.layers[-1].activation = tf.keras.activations.softmax
+
+        train_batch_size = args['per_device_train_batch_size'] * args['n_device']
+        train_dataset, num_train_examples = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, train_batch_size, mode="train")
+        train_dataset = strategy.experimental_distribute_dataset(train_dataset)
+        train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id)
+
+        if not os.path.exists(args['output_dir']):
+            os.makedirs(args['output_dir'])
+
+        logging.info("Saving model to %s", args['output_dir'])
+
+        model.save_pretrained(args['output_dir'])
+        tokenizer.save_pretrained(args['output_dir'])
+
+    # Evaluation
+    if args['do_eval']:
+        tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
+        checkpoints = []
+        results = []
+
+        if args['eval_all_checkpoints']:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args['output_dir'] + "/**/" + TF2_WEIGHTS_NAME, recursive=True), key=lambda f: int(''.join(filter(str.isdigit, f)) or -1)))
+        
+        logging.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        if len(checkpoints) == 0:
+            checkpoints.append(args['output_dir'])
+        
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if re.match(".*checkpoint-[0-9]", checkpoint) else "final"
+
+            with strategy.scope():
+                model = model_class.from_pretrained(checkpoint)
+
+            y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
+            report = metrics.classification_report(y_true, y_pred, digits=4)
+
+            if global_step:
+                results.append({global_step + "_report": report, global_step + "_loss": eval_loss})
+
+        output_eval_file = os.path.join(args['output_dir'], "eval_results.txt")
+        
+        with tf.io.gfile.GFile(output_eval_file, "w") as writer:
+            for res in results:
+                for key, val in res.items():
+                    if "loss" in key:
+                        logging.info(key + " = " + str(val))
+                        writer.write(key + " = " + str(val))
+                        writer.write("\n")
+                    else:
+                        logging.info(key)
+                        logging.info("\n" + report)
+                        writer.write(key + "\n")
+                        writer.write(report)
+                        writer.write("\n")
+
+    if args['do_predict']:
+        tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
+        model = model_class.from_pretrained(args['output_dir'])
+        eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
+        predict_dataset, _ = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode="test")
+        y_true, y_pred, pred_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="test")
+        output_test_results_file = os.path.join(args['output_dir'], "test_results.txt")
+        output_test_predictions_file = os.path.join(args['output_dir'], "test_predictions.txt")
+        report = metrics.classification_report(y_true, y_pred, digits=4)
+
+        with tf.io.gfile.GFile(output_test_results_file, "w") as writer:
+            report = metrics.classification_report(y_true, y_pred, digits=4)
+            
+            logging.info("\n" + report)
+            
+            writer.write(report)
+            writer.write("\n\nloss = " + str(pred_loss))
+
+        with tf.io.gfile.GFile(output_test_predictions_file, "w") as writer:
+            with tf.io.gfile.GFile(os.path.join(args['data_dir'], "test.txt"), "r") as f:
+                example_id = 0
+
+                for line in f:
+                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                        writer.write(line)
+
+                        if not y_pred[example_id]:
+                            example_id += 1
+                    elif y_pred[example_id]:
+                        output_line = line.split()[0] + " " + y_pred[example_id].pop(0) + "\n"
+                        writer.write(output_line)
+                    else:
+                        logging.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
+
+
+if __name__ == "__main__":
+    flags.mark_flag_as_required("data_dir")
+    flags.mark_flag_as_required("output_dir")
+    flags.mark_flag_as_required("model_name_or_path")
+    flags.mark_flag_as_required("model_type")
+    app.run(main)

transformers/__init__.py

@@ -164,6 +164,7 @@ if is_tf_available():
     from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
                                          TFDistilBertModel, TFDistilBertForMaskedLM,
                                          TFDistilBertForSequenceClassification,
+                                         TFDistilBertForTokenClassification,
                                          TFDistilBertForQuestionAnswering,
                                          TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
@@ -174,6 +175,8 @@ if is_tf_available():
     from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
                                      TFAlbertForSequenceClassification,
                                     TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+    # Optimization
+    from .optimization_tf import (WarmUp, create_optimizer, AdamWeightDecay, GradientAccumulator)
 
 # TF 2.0 <=> PyTorch conversion utilities
 from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,

transformers/modeling_tf_distilbert.py

@@ -704,6 +704,53 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel):
         return outputs  # logits, (hidden_states), (attentions)
 
 
+@add_start_docstrings("""DistilBert Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+    Examples::
+        import tensorflow as tf
+        from transformers import DistilBertTokenizer, TFDistilBertForTokenClassification
+        tokenizer = DistilBertTokenizer.from_pretrained('bert-base-uncased')
+        model = TFDistilBertForTokenClassification.from_pretrained('bert-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        scores = outputs[0]
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFDistilBertForTokenClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.distilbert = TFDistilBertMainLayer(config, name='distilbert')
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.distilbert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # scores, (hidden_states), (attentions)
+
+
 @add_start_docstrings("""DistilBert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
                          the hidden-states output to compute `span start logits` and `span end logits`). """,
                       DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)

transformers/optimization_tf.py

@@ -0,0 +1,254 @@
+# Copyright 2019 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.
+# ==============================================================================
+"""Functions and classes related to optimization (weight updates)."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import re
+
+import tensorflow as tf
+
+
+class WarmUp(tf.keras.optimizers.schedules.LearningRateSchedule):
+  """Applys a warmup schedule on a given learning rate decay schedule."""
+
+  def __init__(
+      self,
+      initial_learning_rate,
+      decay_schedule_fn,
+      warmup_steps,
+      power=1.0,
+      name=None):
+    super(WarmUp, self).__init__()
+    self.initial_learning_rate = initial_learning_rate
+    self.warmup_steps = warmup_steps
+    self.power = power
+    self.decay_schedule_fn = decay_schedule_fn
+    self.name = name
+
+  def __call__(self, step):
+    with tf.name_scope(self.name or 'WarmUp') as name:
+      # Implements polynomial warmup. i.e., if global_step < warmup_steps, the
+      # learning rate will be `global_step/num_warmup_steps * init_lr`.
+      global_step_float = tf.cast(step, tf.float32)
+      warmup_steps_float = tf.cast(self.warmup_steps, tf.float32)
+      warmup_percent_done = global_step_float / warmup_steps_float
+      warmup_learning_rate = (
+          self.initial_learning_rate *
+          tf.math.pow(warmup_percent_done, self.power))
+      return tf.cond(global_step_float < warmup_steps_float,
+                     lambda: warmup_learning_rate,
+                     lambda: self.decay_schedule_fn(step),
+                     name=name)
+
+  def get_config(self):
+    return {
+        'initial_learning_rate': self.initial_learning_rate,
+        'decay_schedule_fn': self.decay_schedule_fn,
+        'warmup_steps': self.warmup_steps,
+        'power': self.power,
+        'name': self.name
+    }
+
+
+def create_optimizer(init_lr, num_train_steps, num_warmup_steps):
+  """Creates an optimizer with learning rate schedule."""
+  # Implements linear decay of the learning rate.
+  learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay(
+      initial_learning_rate=init_lr,
+      decay_steps=num_train_steps,
+      end_learning_rate=0.0)
+  if num_warmup_steps:
+    learning_rate_fn = WarmUp(initial_learning_rate=init_lr,
+                              decay_schedule_fn=learning_rate_fn,
+                              warmup_steps=num_warmup_steps)
+  optimizer = AdamWeightDecay(
+      learning_rate=learning_rate_fn,
+      weight_decay_rate=0.01,
+      beta_1=0.9,
+      beta_2=0.999,
+      epsilon=1e-6,
+      exclude_from_weight_decay=['layer_norm', 'bias'])
+  return optimizer
+
+
+class AdamWeightDecay(tf.keras.optimizers.Adam):
+  """Adam enables L2 weight decay and clip_by_global_norm on gradients.
+
+  Just adding the square of the weights to the loss function is *not* the
+  correct way of using L2 regularization/weight decay with Adam, since that will
+  interact with the m and v parameters in strange ways.
+
+  Instead we want ot decay the weights in a manner that doesn't interact with
+  the m/v parameters. This is equivalent to adding the square of the weights to
+  the loss with plain (non-momentum) SGD.
+  """
+
+  def __init__(self,
+               learning_rate=0.001,
+               beta_1=0.9,
+               beta_2=0.999,
+               epsilon=1e-7,
+               amsgrad=False,
+               weight_decay_rate=0.0,
+               include_in_weight_decay=None,
+               exclude_from_weight_decay=None,
+               name='AdamWeightDecay',
+               **kwargs):
+    super(AdamWeightDecay, self).__init__(
+        learning_rate, beta_1, beta_2, epsilon, amsgrad, name, **kwargs)
+    self.weight_decay_rate = weight_decay_rate
+    self._include_in_weight_decay = include_in_weight_decay
+    self._exclude_from_weight_decay = exclude_from_weight_decay
+
+  @classmethod
+  def from_config(cls, config):
+    """Creates an optimizer from its config with WarmUp custom object."""
+    custom_objects = {'WarmUp': WarmUp}
+    return super(AdamWeightDecay, cls).from_config(
+        config, custom_objects=custom_objects)
+
+  def _prepare_local(self, var_device, var_dtype, apply_state):
+    super(AdamWeightDecay, self)._prepare_local(var_device, var_dtype,
+                                                apply_state)
+    apply_state['weight_decay_rate'] = tf.constant(
+        self.weight_decay_rate, name='adam_weight_decay_rate')
+
+  def _decay_weights_op(self, var, learning_rate, apply_state):
+    do_decay = self._do_use_weight_decay(var.name)
+    if do_decay:
+      return var.assign_sub(
+          learning_rate * var *
+          apply_state['weight_decay_rate'],
+          use_locking=self._use_locking)
+    return tf.no_op()
+
+  def apply_gradients(self, grads_and_vars, clip_norm, name=None):
+    grads, tvars = list(zip(*grads_and_vars))
+    (grads, _) = tf.clip_by_global_norm(grads, clip_norm=clip_norm)
+    return super(AdamWeightDecay, self).apply_gradients(zip(grads, tvars))
+
+  def _get_lr(self, var_device, var_dtype, apply_state):
+    """Retrieves the learning rate with the given state."""
+    if apply_state is None:
+      return self._decayed_lr_t[var_dtype], {}
+
+    apply_state = apply_state or {}
+    coefficients = apply_state.get((var_device, var_dtype))
+    if coefficients is None:
+      coefficients = self._fallback_apply_state(var_device, var_dtype)
+      apply_state[(var_device, var_dtype)] = coefficients
+
+    return coefficients['lr_t'], dict(apply_state=apply_state)
+
+  def _resource_apply_dense(self, grad, var, apply_state=None):
+    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
+    decay = self._decay_weights_op(var, lr_t, apply_state)
+    with tf.control_dependencies([decay]):
+      return super(AdamWeightDecay, self)._resource_apply_dense(
+          grad, var, **kwargs)
+
+  def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
+    lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
+    decay = self._decay_weights_op(var, lr_t, apply_state)
+    with tf.control_dependencies([decay]):
+      return super(AdamWeightDecay, self)._resource_apply_sparse(
+          grad, var, indices, **kwargs)
+
+  def get_config(self):
+    config = super(AdamWeightDecay, self).get_config()
+    config.update({
+        'weight_decay_rate': self.weight_decay_rate,
+    })
+    return config
+
+  def _do_use_weight_decay(self, param_name):
+    """Whether to use L2 weight decay for `param_name`."""
+    if self.weight_decay_rate == 0:
+      return False
+
+    if self._include_in_weight_decay:
+      for r in self._include_in_weight_decay:
+        if re.search(r, param_name) is not None:
+          return True
+
+    if self._exclude_from_weight_decay:
+      for r in self._exclude_from_weight_decay:
+        if re.search(r, param_name) is not None:
+          return False
+    return True
+
+
+## Inspired from https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/optimizers/utils.py
+class GradientAccumulator(object):
+    """Distribution strategies-aware gradient accumulation utility."""
+
+    def __init__(self):
+        """Initializes the accumulator."""
+        self._gradients = []
+        self._accum_steps = tf.Variable(
+            initial_value=0,
+            dtype=tf.int64,
+            trainable=False,
+            aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA)
+
+    @property
+    def step(self):
+        """Number of accumulated steps."""
+        return self._accum_steps.value()
+
+    @property
+    def gradients(self):
+        """The accumulated gradients."""
+        return list(gradient.value() if gradient is not None else gradient for gradient in self._get_replica_gradients())
+
+    def __call__(self, gradients):
+        """Accumulates :obj:`gradients`."""
+        if not self._gradients:
+            self._gradients.extend([tf.Variable(tf.zeros_like(gradient), trainable=False) if gradient is not None else gradient for gradient in gradients])
+
+        if len(gradients) != len(self._gradients):
+            raise ValueError("Expected %s gradients, but got %d" % (len(self._gradients), len(gradients)))
+
+        for accum_gradient, gradient in zip(self._get_replica_gradients(), gradients):
+            if accum_gradient is not None:
+                accum_gradient.assign_add(gradient)
+
+        self._accum_steps.assign_add(1)
+
+    def reset(self):
+        """Resets the accumulated gradients."""
+        if self._gradients:
+            self._accum_steps.assign(0)
+
+        for gradient in self._get_replica_gradients():
+            if gradient is not None:
+                gradient.assign(tf.zeros_like(gradient))
+
+    def _get_replica_gradients(self):
+        if tf.distribute.has_strategy():
+            # In a replica context, we want to accumulate gradients on each replica
+            # without synchronization, so we directly assign the value of the
+            # current replica.
+            replica_context = tf.distribute.get_replica_context()
+
+            if replica_context is None or tf.distribute.get_strategy().num_replicas_in_sync == 1:
+                return self._gradients
+
+            return (gradient.device_map.select_for_current_replica(gradient.values, replica_context) for gradient in self._gradients)
+        else:
+            return self._gradients

transformers/tests/optimization_tf_test.py

@@ -0,0 +1,89 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import pytest
+
+from transformers import is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+    from tensorflow.python.eager import context
+    from tensorflow.python.framework import ops
+    from transformers import (create_optimizer, GradientAccumulator)
+else:
+    pytestmark = pytest.mark.skip("Require TensorFlow")
+
+class OptimizationFTest(unittest.TestCase):
+    def assertListAlmostEqual(self, list1, list2, tol):
+        self.assertEqual(len(list1), len(list2))
+        for a, b in zip(list1, list2):
+            self.assertAlmostEqual(a, b, delta=tol)
+    
+    def testGradientAccumulator(self):
+        accumulator = GradientAccumulator()
+        accumulator([tf.constant([1.0, 2.0])])
+        accumulator([tf.constant([-2.0, 1.0])])
+        accumulator([tf.constant([-1.0, 2.0])])
+        with self.assertRaises(ValueError):
+            accumulator([tf.constant([1.0, 1.0]), tf.constant([2.0, 2.0])])
+        self.assertEqual(accumulator.step, 3)
+        self.assertEqual(len(accumulator.gradients), 1)
+        self.assertListAlmostEqual(accumulator.gradients[0].numpy().tolist(), [-2.0, 5.0], tol=1e-2)
+        accumulator.reset()
+        self.assertEqual(accumulator.step, 0)
+        self.assertListAlmostEqual(accumulator.gradients[0].numpy().tolist(), [0.0, 0.0], tol=1e-2)
+
+    def testGradientAccumulatorDistributionStrategy(self):
+        context._context = None
+        ops.enable_eager_execution_internal()
+        physical_devices = tf.config.experimental.list_physical_devices("CPU")
+        tf.config.experimental.set_virtual_device_configuration(
+            physical_devices[0],
+            [tf.config.experimental.VirtualDeviceConfiguration(),
+            tf.config.experimental.VirtualDeviceConfiguration()])
+
+        devices = tf.config.experimental.list_logical_devices(device_type="CPU")
+        strategy = tf.distribute.MirroredStrategy(devices=[device.name for device in devices])
+
+        with strategy.scope():
+            accumulator = GradientAccumulator()
+            variable = tf.Variable([4.0, 3.0])
+            optimizer = create_optimizer(5e-5, 10, 5)
+            gradient_placeholder = tf.Variable([0.0, 0.0], trainable=False)
+
+        def accumulate_on_replica(gradient):
+            accumulator([gradient])
+
+        def apply_on_replica():
+            optimizer.apply_gradients(list(zip(accumulator.gradients, [variable])), 1.0)
+
+        @tf.function
+        def accumulate(grad1, grad2):
+            with strategy.scope():
+                gradient_placeholder.values[0].assign(grad1)
+                gradient_placeholder.values[1].assign(grad2)
+                strategy.experimental_run_v2(accumulate_on_replica, args=(gradient_placeholder,))
+
+        @tf.function
+        def apply_grad():
+            with strategy.scope():
+                strategy.experimental_run_v2(apply_on_replica)
+
+        accumulate([1.0, 2.0], [-1.0, 1.0])
+        accumulate([3.0, -1.0], [-1.0, -1.0])
+        accumulate([-2.0, 2.0], [3.0, -2.0])
+        self.assertEqual(accumulator.step, 3)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[0].value().numpy().tolist(), [2.0, 3.0], tol=1e-2)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[1].value().numpy().tolist(), [1.0, -2.0], tol=1e-2)
+        apply_grad()
+        self.assertListAlmostEqual(variable.value().numpy().tolist(), [4.0, 3.0], tol=1e-2)
+        accumulator.reset()
+        self.assertEqual(accumulator.step, 0)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[0].value().numpy().tolist(), [0.0, 0.0], tol=1e-2)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[1].value().numpy().tolist(), [0.0, 0.0], tol=1e-2)
+
+
+if __name__ == "__main__":
+    unittest.main()