Add POS tagging and Phrase chunking token classification examples (#6457)

* Add more token classification examples

* POS tagging example

* Phrase chunking example

* PR review fixes

* Add conllu to third party list (used in token classification examples)

examples/requirements.txt

@@ -15,3 +15,4 @@ pandas
 nlp
 fire
 pytest
+conllu

examples/token-classification/run.sh

@@ -18,6 +18,7 @@ export SAVE_STEPS=750
 export SEED=1
 
 python3 run_ner.py \
+--task_type NER \
 --data_dir . \
 --labels ./labels.txt \
 --model_name_or_path $BERT_MODEL \

examples/token-classification/run_chunk.sh

@@ -0,0 +1,37 @@
+if ! [ -f ./dev.txt ]; then
+  echo "Downloading CONLL2003 dev dataset...."
+  curl -L -o ./dev.txt 'https://github.com/davidsbatista/NER-datasets/raw/master/CONLL2003/valid.txt'
+fi
+
+if ! [ -f ./test.txt ]; then
+  echo "Downloading CONLL2003 test dataset...."
+  curl -L -o ./test.txt 'https://github.com/davidsbatista/NER-datasets/raw/master/CONLL2003/test.txt'
+fi
+
+if ! [ -f ./train.txt ]; then
+  echo "Downloading CONLL2003 train dataset...."
+  curl -L -o ./train.txt 'https://github.com/davidsbatista/NER-datasets/raw/master/CONLL2003/train.txt'
+fi
+
+export MAX_LENGTH=200
+export BERT_MODEL=bert-base-uncased
+export OUTPUT_DIR=chunker-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+
+python3 run_ner.py \
+--task_type Chunk \
+--data_dir . \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+

examples/token-classification/run_ner.py

@@ -14,16 +14,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ Fine-tuning the library models for named entity recognition on CoNLL-2003. """
-
-
 import logging
 import os
 import sys
 from dataclasses import dataclass, field
+from importlib import import_module
 from typing import Dict, List, Optional, Tuple
 
 import numpy as np
-from seqeval.metrics import f1_score, precision_score, recall_score
+from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
 from torch import nn
 
 from transformers import (
@@ -36,7 +35,7 @@ from transformers import (
     TrainingArguments,
     set_seed,
 )
-from utils_ner import NerDataset, Split, get_labels
+from utils_ner import Split, TokenClassificationDataset, TokenClassificationTask
 
 
 logger = logging.getLogger(__name__)
@@ -54,6 +53,9 @@ class ModelArguments:
     config_name: Optional[str] = field(
         default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
     )
+    task_type: Optional[str] = field(
+        default="NER", metadata={"help": "Task type to fine tune in training (e.g. NER, POS, etc)"}
+    )
     tokenizer_name: Optional[str] = field(
         default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
     )
@@ -113,6 +115,16 @@ def main():
             f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
         )
 
+    module = import_module("tasks")
+    try:
+        token_classification_task_clazz = getattr(module, model_args.task_type)
+        token_classification_task: TokenClassificationTask = token_classification_task_clazz()
+    except AttributeError:
+        raise ValueError(
+            f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
+            f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
+        )
+
     # Setup logging
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
@@ -133,7 +145,7 @@ def main():
     set_seed(training_args.seed)
 
     # Prepare CONLL-2003 task
-    labels = get_labels(data_args.labels)
+    labels = token_classification_task.get_labels(data_args.labels)
     label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
     num_labels = len(labels)
 
@@ -164,7 +176,8 @@ def main():
 
     # Get datasets
     train_dataset = (
-        NerDataset(
+        TokenClassificationDataset(
+            token_classification_task=token_classification_task,
             data_dir=data_args.data_dir,
             tokenizer=tokenizer,
             labels=labels,
@@ -177,7 +190,8 @@ def main():
         else None
     )
     eval_dataset = (
-        NerDataset(
+        TokenClassificationDataset(
+            token_classification_task=token_classification_task,
             data_dir=data_args.data_dir,
             tokenizer=tokenizer,
             labels=labels,
@@ -209,6 +223,7 @@ def main():
     def compute_metrics(p: EvalPrediction) -> Dict:
         preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
         return {
+            "accuracy_score": accuracy_score(out_label_list, preds_list),
             "precision": precision_score(out_label_list, preds_list),
             "recall": recall_score(out_label_list, preds_list),
             "f1": f1_score(out_label_list, preds_list),
@@ -253,7 +268,8 @@ def main():
 
     # Predict
     if training_args.do_predict:
-        test_dataset = NerDataset(
+        test_dataset = TokenClassificationDataset(
+            token_classification_task=token_classification_task,
             data_dir=data_args.data_dir,
             tokenizer=tokenizer,
             labels=labels,
@@ -278,19 +294,7 @@ def main():
         if trainer.is_world_master():
             with open(output_test_predictions_file, "w") as writer:
                 with open(os.path.join(data_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 preds_list[example_id]:
-                                example_id += 1
-                        elif preds_list[example_id]:
-                            output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
-                            writer.write(output_line)
-                        else:
-                            logger.warning(
-                                "Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0]
-                            )
+                    token_classification_task.write_predictions_to_file(writer, f, preds_list)
 
     return results
 

examples/token-classification/run_pl_ner.py

@@ -2,15 +2,17 @@ import argparse
 import glob
 import logging
 import os
+from argparse import Namespace
+from importlib import import_module
 
 import numpy as np
 import torch
-from seqeval.metrics import f1_score, precision_score, recall_score
+from seqeval.metrics import accuracy_score, f1_score, precision_score, recall_score
 from torch.nn import CrossEntropyLoss
 from torch.utils.data import DataLoader, TensorDataset
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
+from utils_ner import TokenClassificationTask
 
 
 logger = logging.getLogger(__name__)
@@ -24,10 +26,20 @@ class NERTransformer(BaseTransformer):
     mode = "token-classification"
 
     def __init__(self, hparams):
-        self.labels = get_labels(hparams.labels)
-        num_labels = len(self.labels)
+        if type(hparams) == dict:
+            hparams = Namespace(**hparams)
+        module = import_module("tasks")
+        try:
+            token_classification_task_clazz = getattr(module, hparams.task_type)
+            self.token_classification_task: TokenClassificationTask = token_classification_task_clazz()
+        except AttributeError:
+            raise ValueError(
+                f"Task {hparams.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
+                f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
+            )
+        self.labels = self.token_classification_task.get_labels(hparams.labels)
         self.pad_token_label_id = CrossEntropyLoss().ignore_index
-        super().__init__(hparams, num_labels, self.mode)
+        super().__init__(hparams, len(self.labels), self.mode)
 
     def forward(self, **inputs):
         return self.model(**inputs)
@@ -42,8 +54,8 @@ class NERTransformer(BaseTransformer):
 
         outputs = self(**inputs)
         loss = outputs[0]
-        tensorboard_logs = {"loss": loss, "rate": self.lr_scheduler.get_last_lr()[-1]}
-        return {"loss": loss, "log": tensorboard_logs}
+        # tensorboard_logs = {"loss": loss, "rate": self.lr_scheduler.get_last_lr()[-1]}
+        return {"loss": loss}
 
     def prepare_data(self):
         "Called to initialize data. Use the call to construct features"
@@ -55,8 +67,8 @@ class NERTransformer(BaseTransformer):
                 features = torch.load(cached_features_file)
             else:
                 logger.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 = self.token_classification_task.read_examples_from_file(args.data_dir, mode)
+                features = self.token_classification_task.convert_examples_to_features(
                     examples,
                     self.labels,
                     args.max_seq_length,
@@ -74,7 +86,7 @@ class NERTransformer(BaseTransformer):
                 logger.info("Saving features into cached file %s", cached_features_file)
                 torch.save(features, cached_features_file)
 
-    def load_dataset(self, mode, batch_size):
+    def get_dataloader(self, mode: int, batch_size: int) -> DataLoader:
         "Load datasets. Called after prepare data."
         cached_features_file = self._feature_file(mode)
         logger.info("Loading features from cached file %s", cached_features_file)
@@ -124,6 +136,7 @@ class NERTransformer(BaseTransformer):
 
         results = {
             "val_loss": val_loss_mean,
+            "accuracy_score": accuracy_score(out_label_list, preds_list),
             "precision": precision_score(out_label_list, preds_list),
             "recall": recall_score(out_label_list, preds_list),
             "f1": f1_score(out_label_list, preds_list),
@@ -154,6 +167,9 @@ class NERTransformer(BaseTransformer):
     def add_model_specific_args(parser, root_dir):
         # Add NER specific options
         BaseTransformer.add_model_specific_args(parser, root_dir)
+        parser.add_argument(
+            "--task_type", default="NER", type=str, help="Task type to fine tune in training (e.g. NER, POS, etc)"
+        )
         parser.add_argument(
             "--max_seq_length",
             default=128,

examples/token-classification/run_pos.sh

@@ -0,0 +1,37 @@
+if ! [ -f ./dev.txt ]; then
+  echo "Download dev dataset...."
+  curl -L -o ./dev.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-dev.conllu'
+fi
+
+if ! [ -f ./test.txt ]; then
+  echo "Download test dataset...."
+  curl -L -o ./test.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-test.conllu'
+fi
+
+if ! [ -f ./train.txt ]; then
+  echo "Download train dataset...."
+  curl -L -o ./train.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-train.conllu'
+fi
+
+export MAX_LENGTH=200
+export BERT_MODEL=bert-base-uncased
+export OUTPUT_DIR=postagger-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+
+python3 run_ner.py \
+--task_type POS \
+--data_dir . \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+

examples/token-classification/run_pos_pl.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+if ! [ -f ./dev.txt ]; then
+  echo "Download dev dataset...."
+  curl -L -o ./dev.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-dev.conllu'
+fi
+
+if ! [ -f ./test.txt ]; then
+  echo "Download test dataset...."
+  curl -L -o ./test.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-test.conllu'
+fi
+
+if ! [ -f ./train.txt ]; then
+  echo "Download train dataset...."
+  curl -L -o ./train.txt 'https://github.com/UniversalDependencies/UD_English-EWT/raw/master/en_ewt-ud-train.conllu'
+fi
+
+export MAX_LENGTH=200
+export BERT_MODEL=bert-base-uncased
+export OUTPUT_DIR=postagger-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+
+
+# Add parent directory to python path to access lightning_base.py
+export PYTHONPATH="../":"${PYTHONPATH}"
+
+python3 run_pl_ner.py --data_dir ./ \
+--task_type POS \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--train_batch_size $BATCH_SIZE \
+--seed $SEED \
+--gpus 1 \
+--do_train \
+--do_predict

examples/token-classification/tasks.py

@@ -0,0 +1,163 @@
+import logging
+import os
+from typing import List, TextIO, Union
+
+from conllu import parse_incr
+
+from utils_ner import InputExample, Split, TokenClassificationTask
+
+
+logger = logging.getLogger(__name__)
+
+
+class NER(TokenClassificationTask):
+    def __init__(self, label_idx=-1):
+        # in NER datasets, the last column is usually reserved for NER label
+        self.label_idx = label_idx
+
+    def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> List[InputExample]:
+        if isinstance(mode, Split):
+            mode = mode.value
+        file_path = os.path.join(data_dir, f"{mode}.txt")
+        guid_index = 1
+        examples = []
+        with open(file_path, encoding="utf-8") as f:
+            words = []
+            labels = []
+            for line in f:
+                if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                    if words:
+                        examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
+                        guid_index += 1
+                        words = []
+                        labels = []
+                else:
+                    splits = line.split(" ")
+                    words.append(splits[0])
+                    if len(splits) > 1:
+                        labels.append(splits[self.label_idx].replace("\n", ""))
+                    else:
+                        # Examples could have no label for mode = "test"
+                        labels.append("O")
+            if words:
+                examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
+        return examples
+
+    def write_predictions_to_file(self, writer: TextIO, test_input_reader: TextIO, preds_list: List):
+        example_id = 0
+        for line in test_input_reader:
+            if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                writer.write(line)
+                if not preds_list[example_id]:
+                    example_id += 1
+            elif preds_list[example_id]:
+                output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
+                writer.write(output_line)
+            else:
+                logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
+
+    def get_labels(self, path: str) -> List[str]:
+        if path:
+            with open(path, "r") as f:
+                labels = f.read().splitlines()
+            if "O" not in labels:
+                labels = ["O"] + labels
+            return labels
+        else:
+            return ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
+
+
+class Chunk(NER):
+    def __init__(self):
+        # in CONLL2003 dataset chunk column is second-to-last
+        super().__init__(label_idx=-2)
+
+    def get_labels(self, path: str) -> List[str]:
+        if path:
+            with open(path, "r") as f:
+                labels = f.read().splitlines()
+            if "O" not in labels:
+                labels = ["O"] + labels
+            return labels
+        else:
+            return [
+                "O",
+                "B-ADVP",
+                "B-INTJ",
+                "B-LST",
+                "B-PRT",
+                "B-NP",
+                "B-SBAR",
+                "B-VP",
+                "B-ADJP",
+                "B-CONJP",
+                "B-PP",
+                "I-ADVP",
+                "I-INTJ",
+                "I-LST",
+                "I-PRT",
+                "I-NP",
+                "I-SBAR",
+                "I-VP",
+                "I-ADJP",
+                "I-CONJP",
+                "I-PP",
+            ]
+
+
+class POS(TokenClassificationTask):
+    def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> List[InputExample]:
+        if isinstance(mode, Split):
+            mode = mode.value
+        file_path = os.path.join(data_dir, f"{mode}.txt")
+        guid_index = 1
+        examples = []
+
+        with open(file_path, encoding="utf-8") as f:
+            for sentence in parse_incr(f):
+                words = []
+                labels = []
+                for token in sentence:
+                    words.append(token["form"])
+                    labels.append(token["upos"])
+                assert len(words) == len(labels)
+                if words:
+                    examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
+                    guid_index += 1
+        return examples
+
+    def write_predictions_to_file(self, writer: TextIO, test_input_reader: TextIO, preds_list: List):
+        example_id = 0
+        for sentence in parse_incr(test_input_reader):
+            s_p = preds_list[example_id]
+            out = ""
+            for token in sentence:
+                out += f'{token["form"]} ({token["upos"]}|{s_p.pop(0)}) '
+            out += "\n"
+            writer.write(out)
+            example_id += 1
+
+    def get_labels(self, path: str) -> List[str]:
+        if path:
+            with open(path, "r") as f:
+                return f.read().splitlines()
+        else:
+            return [
+                "ADJ",
+                "ADP",
+                "ADV",
+                "AUX",
+                "CCONJ",
+                "DET",
+                "INTJ",
+                "NOUN",
+                "NUM",
+                "PART",
+                "PRON",
+                "PROPN",
+                "PUNCT",
+                "SCONJ",
+                "SYM",
+                "VERB",
+                "X",
+            ]

examples/token-classification/utils_ner.py

@@ -66,202 +66,15 @@ class Split(Enum):
     test = "test"
 
 
-if is_torch_available():
-    import torch
-    from torch import nn
-    from torch.utils.data.dataset import Dataset
+class TokenClassificationTask:
+    def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> List[InputExample]:
+        raise NotImplementedError
 
-    class NerDataset(Dataset):
-        """
-        This will be superseded by a framework-agnostic approach
-        soon.
-        """
+    def get_labels(self, path: str) -> List[str]:
+        raise NotImplementedError
 
-        features: List[InputFeatures]
-        pad_token_label_id: int = nn.CrossEntropyLoss().ignore_index
-        # Use cross entropy ignore_index as padding label id so that only
-        # real label ids contribute to the loss later.
-
-        def __init__(
+    def convert_examples_to_features(
         self,
-            data_dir: str,
-            tokenizer: PreTrainedTokenizer,
-            labels: List[str],
-            model_type: str,
-            max_seq_length: Optional[int] = None,
-            overwrite_cache=False,
-            mode: Split = Split.train,
-        ):
-            # Load data features from cache or dataset file
-            cached_features_file = os.path.join(
-                data_dir, "cached_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length)),
-            )
-
-            # Make sure only the first process in distributed training processes the dataset,
-            # and the others will use the cache.
-            lock_path = cached_features_file + ".lock"
-            with FileLock(lock_path):
-
-                if os.path.exists(cached_features_file) and not overwrite_cache:
-                    logger.info(f"Loading features from cached file {cached_features_file}")
-                    self.features = torch.load(cached_features_file)
-                else:
-                    logger.info(f"Creating features from dataset file at {data_dir}")
-                    examples = read_examples_from_file(data_dir, mode)
-                    # TODO clean up all this to leverage built-in features of tokenizers
-                    self.features = convert_examples_to_features(
-                        examples,
-                        labels,
-                        max_seq_length,
-                        tokenizer,
-                        cls_token_at_end=bool(model_type in ["xlnet"]),
-                        # xlnet has a cls token at the end
-                        cls_token=tokenizer.cls_token,
-                        cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
-                        sep_token=tokenizer.sep_token,
-                        sep_token_extra=False,
-                        # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
-                        pad_on_left=bool(tokenizer.padding_side == "left"),
-                        pad_token=tokenizer.pad_token_id,
-                        pad_token_segment_id=tokenizer.pad_token_type_id,
-                        pad_token_label_id=self.pad_token_label_id,
-                    )
-                    logger.info(f"Saving features into cached file {cached_features_file}")
-                    torch.save(self.features, cached_features_file)
-
-        def __len__(self):
-            return len(self.features)
-
-        def __getitem__(self, i) -> InputFeatures:
-            return self.features[i]
-
-
-if is_tf_available():
-    import tensorflow as tf
-
-    class TFNerDataset:
-        """
-        This will be superseded by a framework-agnostic approach
-        soon.
-        """
-
-        features: List[InputFeatures]
-        pad_token_label_id: int = -100
-        # Use cross entropy ignore_index as padding label id so that only
-        # real label ids contribute to the loss later.
-
-        def __init__(
-            self,
-            data_dir: str,
-            tokenizer: PreTrainedTokenizer,
-            labels: List[str],
-            model_type: str,
-            max_seq_length: Optional[int] = None,
-            overwrite_cache=False,
-            mode: Split = Split.train,
-        ):
-            examples = read_examples_from_file(data_dir, mode)
-            # TODO clean up all this to leverage built-in features of tokenizers
-            self.features = convert_examples_to_features(
-                examples,
-                labels,
-                max_seq_length,
-                tokenizer,
-                cls_token_at_end=bool(model_type in ["xlnet"]),
-                # xlnet has a cls token at the end
-                cls_token=tokenizer.cls_token,
-                cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
-                sep_token=tokenizer.sep_token,
-                sep_token_extra=False,
-                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
-                pad_on_left=bool(tokenizer.padding_side == "left"),
-                pad_token=tokenizer.pad_token_id,
-                pad_token_segment_id=tokenizer.pad_token_type_id,
-                pad_token_label_id=self.pad_token_label_id,
-            )
-
-            def gen():
-                for ex in self.features:
-                    if ex.token_type_ids is None:
-                        yield (
-                            {"input_ids": ex.input_ids, "attention_mask": ex.attention_mask},
-                            ex.label_ids,
-                        )
-                    else:
-                        yield (
-                            {
-                                "input_ids": ex.input_ids,
-                                "attention_mask": ex.attention_mask,
-                                "token_type_ids": ex.token_type_ids,
-                            },
-                            ex.label_ids,
-                        )
-
-            if "token_type_ids" not in tokenizer.model_input_names:
-                self.dataset = tf.data.Dataset.from_generator(
-                    gen,
-                    ({"input_ids": tf.int32, "attention_mask": tf.int32}, tf.int64),
-                    (
-                        {"input_ids": tf.TensorShape([None]), "attention_mask": tf.TensorShape([None])},
-                        tf.TensorShape([None]),
-                    ),
-                )
-            else:
-                self.dataset = tf.data.Dataset.from_generator(
-                    gen,
-                    ({"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32}, tf.int64),
-                    (
-                        {
-                            "input_ids": tf.TensorShape([None]),
-                            "attention_mask": tf.TensorShape([None]),
-                            "token_type_ids": tf.TensorShape([None]),
-                        },
-                        tf.TensorShape([None]),
-                    ),
-                )
-
-        def get_dataset(self):
-            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
-
-            return self.dataset
-
-        def __len__(self):
-            return len(self.features)
-
-        def __getitem__(self, i) -> InputFeatures:
-            return self.features[i]
-
-
-def read_examples_from_file(data_dir, mode: Union[Split, str]) -> List[InputExample]:
-    if isinstance(mode, Split):
-        mode = mode.value
-    file_path = os.path.join(data_dir, f"{mode}.txt")
-    guid_index = 1
-    examples = []
-    with open(file_path, encoding="utf-8") as f:
-        words = []
-        labels = []
-        for line in f:
-            if line.startswith("-DOCSTART-") or line == "" or line == "\n":
-                if words:
-                    examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
-                    guid_index += 1
-                    words = []
-                    labels = []
-            else:
-                splits = line.split(" ")
-                words.append(splits[0])
-                if len(splits) > 1:
-                    labels.append(splits[-1].replace("\n", ""))
-                else:
-                    # Examples could have no label for mode = "test"
-                    labels.append("O")
-        if words:
-            examples.append(InputExample(guid=f"{mode}-{guid_index}", words=words, labels=labels))
-    return examples
-
-
-def convert_examples_to_features(
         examples: List[InputExample],
         label_list: List[str],
         max_seq_length: int,
@@ -277,7 +90,7 @@ def convert_examples_to_features(
         pad_token_label_id=-100,
         sequence_a_segment_id=0,
         mask_padding_with_zero=True,
-) -> List[InputFeatures]:
+    ) -> List[InputFeatures]:
         """ Loads a data file into a list of `InputFeatures`
             `cls_token_at_end` define the location of the CLS token:
                 - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
@@ -389,12 +202,169 @@ def convert_examples_to_features(
         return features
 
 
-def get_labels(path: str) -> List[str]:
-    if path:
-        with open(path, "r") as f:
-            labels = f.read().splitlines()
-        if "O" not in labels:
-            labels = ["O"] + labels
-        return labels
+if is_torch_available():
+    import torch
+    from torch import nn
+    from torch.utils.data.dataset import Dataset
+
+    class TokenClassificationDataset(Dataset):
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
+
+        features: List[InputFeatures]
+        pad_token_label_id: int = nn.CrossEntropyLoss().ignore_index
+        # Use cross entropy ignore_index as padding label id so that only
+        # real label ids contribute to the loss later.
+
+        def __init__(
+            self,
+            token_classification_task: TokenClassificationTask,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            labels: List[str],
+            model_type: str,
+            max_seq_length: Optional[int] = None,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+        ):
+            # Load data features from cache or dataset file
+            cached_features_file = os.path.join(
+                data_dir, "cached_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length)),
+            )
+
+            # Make sure only the first process in distributed training processes the dataset,
+            # and the others will use the cache.
+            lock_path = cached_features_file + ".lock"
+            with FileLock(lock_path):
+
+                if os.path.exists(cached_features_file) and not overwrite_cache:
+                    logger.info(f"Loading features from cached file {cached_features_file}")
+                    self.features = torch.load(cached_features_file)
                 else:
-        return ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
+                    logger.info(f"Creating features from dataset file at {data_dir}")
+                    examples = token_classification_task.read_examples_from_file(data_dir, mode)
+                    # TODO clean up all this to leverage built-in features of tokenizers
+                    self.features = token_classification_task.convert_examples_to_features(
+                        examples,
+                        labels,
+                        max_seq_length,
+                        tokenizer,
+                        cls_token_at_end=bool(model_type in ["xlnet"]),
+                        # xlnet has a cls token at the end
+                        cls_token=tokenizer.cls_token,
+                        cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
+                        sep_token=tokenizer.sep_token,
+                        sep_token_extra=False,
+                        # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                        pad_on_left=bool(tokenizer.padding_side == "left"),
+                        pad_token=tokenizer.pad_token_id,
+                        pad_token_segment_id=tokenizer.pad_token_type_id,
+                        pad_token_label_id=self.pad_token_label_id,
+                    )
+                    logger.info(f"Saving features into cached file {cached_features_file}")
+                    torch.save(self.features, cached_features_file)
+
+        def __len__(self):
+            return len(self.features)
+
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    class TFNerDataset:
+        """
+        This will be superseded by a framework-agnostic approach
+        soon.
+        """
+
+        features: List[InputFeatures]
+        pad_token_label_id: int = -100
+        # Use cross entropy ignore_index as padding label id so that only
+        # real label ids contribute to the loss later.
+
+        def __init__(
+            self,
+            token_classification_task: TokenClassificationTask,
+            data_dir: str,
+            tokenizer: PreTrainedTokenizer,
+            labels: List[str],
+            model_type: str,
+            max_seq_length: Optional[int] = None,
+            overwrite_cache=False,
+            mode: Split = Split.train,
+        ):
+            examples = token_classification_task.read_examples_from_file(data_dir, mode)
+            # TODO clean up all this to leverage built-in features of tokenizers
+            self.features = token_classification_task.convert_examples_to_features(
+                examples,
+                labels,
+                max_seq_length,
+                tokenizer,
+                cls_token_at_end=bool(model_type in ["xlnet"]),
+                # xlnet has a cls token at the end
+                cls_token=tokenizer.cls_token,
+                cls_token_segment_id=2 if model_type in ["xlnet"] else 0,
+                sep_token=tokenizer.sep_token,
+                sep_token_extra=False,
+                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                pad_on_left=bool(tokenizer.padding_side == "left"),
+                pad_token=tokenizer.pad_token_id,
+                pad_token_segment_id=tokenizer.pad_token_type_id,
+                pad_token_label_id=self.pad_token_label_id,
+            )
+
+            def gen():
+                for ex in self.features:
+                    if ex.token_type_ids is None:
+                        yield (
+                            {"input_ids": ex.input_ids, "attention_mask": ex.attention_mask},
+                            ex.label_ids,
+                        )
+                    else:
+                        yield (
+                            {
+                                "input_ids": ex.input_ids,
+                                "attention_mask": ex.attention_mask,
+                                "token_type_ids": ex.token_type_ids,
+                            },
+                            ex.label_ids,
+                        )
+
+            if "token_type_ids" not in tokenizer.model_input_names:
+                self.dataset = tf.data.Dataset.from_generator(
+                    gen,
+                    ({"input_ids": tf.int32, "attention_mask": tf.int32}, tf.int64),
+                    (
+                        {"input_ids": tf.TensorShape([None]), "attention_mask": tf.TensorShape([None])},
+                        tf.TensorShape([None]),
+                    ),
+                )
+            else:
+                self.dataset = tf.data.Dataset.from_generator(
+                    gen,
+                    ({"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32}, tf.int64),
+                    (
+                        {
+                            "input_ids": tf.TensorShape([None]),
+                            "attention_mask": tf.TensorShape([None]),
+                            "token_type_ids": tf.TensorShape([None]),
+                        },
+                        tf.TensorShape([None]),
+                    ),
+                )
+
+        def get_dataset(self):
+            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
+
+            return self.dataset
+
+        def __len__(self):
+            return len(self.features)
+
+        def __getitem__(self, i) -> InputFeatures:
+            return self.features[i]

setup.cfg

@@ -5,6 +5,7 @@ include_trailing_comma = True
 known_first_party = transformers
 known_third_party =
     absl
+    conllu
     elasticsearch
     fairseq
     faiss