fix the post-processing bug in NER (#534)

* fix conll bug

* update DataCollatorForAuto

* adding label_list comments

flaml/automl.py

@@ -1569,7 +1569,10 @@ class AutoML(BaseEstimator):
         ```
 
             **fit_kwargs: Other key word arguments to pass to fit() function of
-                the searched learners, such as sample_weight.
+                the searched learners, such as sample_weight. Include:
+                    period: int | forecast horizon for ts_forecast tasks.
+                    gpu_per_trial: float, default = 0 | A float of the number of gpus per trial,
+                    only used by TransformersEstimator and XGBoostSklearnEstimator.
         """
         task = task or self._settings.get("task")
         eval_method = eval_method or self._settings.get("eval_method")
@@ -2198,6 +2201,8 @@ class AutoML(BaseEstimator):
             }
         }
             fit_kwargs_by_estimator: dict, default=None | The user specified keywords arguments, grouped by estimator name.
+                For TransformersEstimator, available fit_kwargs can be found from
+                [flaml/nlp/training_args.py:TrainingArgumentsForAuto](https://microsoft.github.io/FLAML/docs/reference/nlp/huggingface/training_args).
                 e.g.,
 
         ```python

flaml/ml.py

@@ -165,7 +165,7 @@ def metric_loss_score(
                 import datasets
 
                 datasets_metric_name = huggingface_submetric_to_metric.get(
-                    metric_name, metric_name
+                    metric_name, metric_name.split(":")[0]
                 )
                 metric = datasets.load_metric(datasets_metric_name)
                 metric_mode = huggingface_metric_to_mode[datasets_metric_name]
@@ -174,17 +174,30 @@ def metric_loss_score(
                     score = metric.compute(predictions=y_predict, references=y_true)[
                         metric_name
                     ].mid.fmeasure
-                elif metric_name == "seqeval":
-                    y_true = [
-                        [x for x in each_y_true if x != -100] for each_y_true in y_true
+                elif metric_name.startswith("seqeval"):
+
+                    zip_pred_true = [
+                        [(p, lb) for (p, lb) in zip(prediction, label) if lb != -100]
+                        for (prediction, label) in zip(y_predict, y_true)
                     ]
                     y_pred = [
-                        y_predict[each_idx][: len(y_true[each_idx])]
-                        for each_idx in range(len(y_predict))
+                        [labels[p] for (p, l) in each_list]
+                        for each_list in zip_pred_true
+                    ]  # To compute precision and recall, y_pred and y_true must be converted to string labels
+                    # (B-PER, I-PER, etc.), so that the category-based precision/recall (i.e., PER, LOC, etc.) scores can be computed
+                    y_true = [
+                        [labels[l] for (p, l) in each_list]
+                        for each_list in zip_pred_true
                     ]
+
+                    metric_submetric_names = metric_name.split(":")
+
                     score = metric.compute(predictions=y_pred, references=y_true)[
-                        "overall_accuracy"
+                        metric_submetric_names[1]
+                        if len(metric_submetric_names) > 1
+                        else "overall_accuracy"
                     ]
+
                 else:
                     score = metric.compute(predictions=y_predict, references=y_true)[
                         metric_name
@@ -454,7 +467,9 @@ def evaluate_model_CV(
     if task in CLASSIFICATION:
         labels = np.unique(y_train_all)
     else:
-        labels = None
+        labels = fit_kwargs.get(
+            "label_list"
+        )  # pass the label list on to compute the evaluation metric
     groups = None
     shuffle = False if task in TS_FORECAST else True
     if isinstance(kf, RepeatedStratifiedKFold):
@@ -586,6 +601,9 @@ def compute_estimator(
             groups_val,
             eval_metric,
             task,
+            labels=fit_kwargs.get(
+                "label_list"
+            ),  # pass the label list on to compute the evaluation metric
             budget=budget,
             log_training_metric=log_training_metric,
             fit_kwargs=fit_kwargs,

flaml/model.py

@@ -365,9 +365,9 @@ class TransformersEstimator(BaseEstimator):
         self._TrainingArguments = TrainingArguments
 
     @staticmethod
-    def _join(X_train, y_train):
+    def _join(X_train, y_train, task):
         y_train = DataFrame(y_train, index=X_train.index)
-        y_train.columns = ["label"]
+        y_train.columns = ["label"] if task != TOKENCLASSIFICATION else ["labels"]
         train_df = X_train.join(y_train)
         return train_df
 
@@ -380,7 +380,7 @@ class TransformersEstimator(BaseEstimator):
             },
             "num_train_epochs": {
                 "domain": tune.loguniform(lower=0.1, upper=10.0),
-                "init_value": 1,
+                "init_value": 3.0,  # to be consistent with roberta
             },
             "per_device_train_batch_size": {
                 "domain": tune.choice([4, 8, 16, 32]),
@@ -511,7 +511,7 @@ class TransformersEstimator(BaseEstimator):
             processed_X, processed_y = self._preprocess(X=X, y=y, **self._kwargs)
 
         processed_dataset = Dataset.from_pandas(
-            TransformersEstimator._join(processed_X, processed_y)
+            TransformersEstimator._join(processed_X, processed_y, self._task)
         )
         return processed_dataset, processed_X, processed_y
 
@@ -547,14 +547,14 @@ class TransformersEstimator(BaseEstimator):
 
     @property
     def data_collator(self):
-        from .nlp.huggingface.data_collator import DataCollatorForAuto
+        from .nlp.huggingface.data_collator import task_to_datacollator_class
 
         return (
-            DataCollatorForAuto(
+            task_to_datacollator_class[self._task](
                 tokenizer=self.tokenizer,
-                pad_to_multiple_of=8 if self._training_args.fp16 else None,
+                pad_to_multiple_of=8,  # if self._training_args.fp16 else None,
             )
-            if self._task == MULTICHOICECLASSIFICATION
+            if self._task in (MULTICHOICECLASSIFICATION, TOKENCLASSIFICATION)
             else None
         )
 
@@ -750,7 +750,10 @@ class TransformersEstimator(BaseEstimator):
                 )
             metric_dict = {
                 "automl_metric": metric_loss_score(
-                    metric_name=self._metric, y_predict=predictions, y_true=labels
+                    metric_name=self._metric,
+                    y_predict=predictions,
+                    y_true=labels,
+                    labels=self._training_args.label_list,
                 )
             }
         else:

flaml/nlp/huggingface/data_collator.py

@@ -1,9 +1,15 @@
 from dataclasses import dataclass
-from transformers.data.data_collator import DataCollatorWithPadding
+from transformers.data.data_collator import (
+    DataCollatorWithPadding,
+    DataCollatorForTokenClassification,
+)
+from collections import OrderedDict
+
+from flaml.data import TOKENCLASSIFICATION, MULTICHOICECLASSIFICATION
 
 
 @dataclass
-class DataCollatorForAuto(DataCollatorWithPadding):
+class DataCollatorForMultipleChoiceClassification(DataCollatorWithPadding):
     def __call__(self, features):
         from itertools import chain
         import torch
@@ -22,7 +28,9 @@ class DataCollatorForAuto(DataCollatorWithPadding):
             for feature in features
         ]
         flattened_features = list(chain(*flattened_features))
-        batch = super(DataCollatorForAuto, self).__call__(flattened_features)
+        batch = super(DataCollatorForMultipleChoiceClassification, self).__call__(
+            flattened_features
+        )
         # Un-flatten
         batch = {k: v.view(batch_size, num_choices, -1) for k, v in batch.items()}
         # Add back labels
@@ -31,18 +39,9 @@ class DataCollatorForAuto(DataCollatorWithPadding):
         return batch
 
 
-class DataCollatorForPredict(DataCollatorWithPadding):
-    def __call__(self, features):
-        from itertools import chain
-
-        batch_size = len(features)
-        num_choices = len(features[0]["input_ids"])
-        flattened_features = [
-            [{k: v[i] for k, v in feature.items()} for i in range(num_choices)]
-            for feature in features
-        ]
-        flattened_features = list(chain(*flattened_features))
-        batch = super(DataCollatorForPredict, self).__call__(flattened_features)
-        # Un-flatten
-        batch = {k: v.view(batch_size, num_choices, -1) for k, v in batch.items()}
-        return batch
+task_to_datacollator_class = OrderedDict(
+    [
+        (TOKENCLASSIFICATION, DataCollatorForTokenClassification),
+        (MULTICHOICECLASSIFICATION, DataCollatorForMultipleChoiceClassification),
+    ]
+)

flaml/nlp/huggingface/training_args.py

@@ -17,13 +17,21 @@ class TrainingArgumentsForAuto(TrainingArguments):
     """FLAML custom TrainingArguments.
 
     Args:
+        task (str): the task name for NLP tasks, e.g., seq-classification, token-classification
         output_dir (str): data root directory for outputing the log, etc.
         model_path (str, optional, defaults to "facebook/muppet-roberta-base"): A string,
             the path of the language model file, either a path from huggingface
             model card huggingface.co/models, or a local path for the model.
         fp16 (bool, optional, defaults to "False"): A bool, whether to use FP16.
         max_seq_length (int, optional, defaults to 128): An integer, the max length of the sequence.
+        pad_to_max_length (bool, optional, defaults to "False"):
+            whether to pad all samples to model maximum sentence length.
+            If False, will pad the samples dynamically when batching to the maximum length in the batch.
         ckpt_per_epoch (int, optional, defaults to 1): An integer, the number of checkpoints per epoch.
+        per_device_eval_batch_size (int, optional, defaults to 1): An integer, the per gpu evaluation batch size.
+        label_list (List[str], optional, defaults to None): A list of string, the string list of the label names.
+            When the task is sequence labeling/token classification, need to set the label_list (e.g., B-PER, I-PER, B-LOC)
+            to obtain the correct evaluation metric. See the example in test/nlp/test_autohf_tokenclassification.py.
     """
 
     task: str = field(default="seq-classification")
@@ -37,21 +45,15 @@ class TrainingArgumentsForAuto(TrainingArguments):
         },
     )
 
-    tokenizer_model_path: str = field(
-        default=None,
-        metadata={"help": "tokenizer model path for HPO"},
-    )
-
     fp16: bool = field(default=True, metadata={"help": "whether to use the FP16 mode"})
 
     max_seq_length: int = field(default=128, metadata={"help": "max seq length"})
 
     pad_to_max_length: bool = field(
-        default=True,
+        default=False,
         metadata={
             "help": "Whether to pad all samples to model maximum sentence length. "
-            "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
-            "efficient on GPU but very bad for TPU."
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch. "
         },
     )
 
@@ -62,21 +64,8 @@ class TrainingArgumentsForAuto(TrainingArguments):
         metadata={"help": "per gpu evaluation batch size"},
     )
 
-    report_to: Optional[List[str]] = field(
-        default=None,
-        metadata={
-            "help": "The list of integrations to report the results and logs to."
-        },
-    )
-
-    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
-    do_eval: bool = field(
-        default=False, metadata={"help": "Whether to run eval on the dev set."}
-    )
-
-    metric_for_best_model: Optional[str] = field(
-        default="loss",
-        metadata={"help": "The metric to use to compare two different models."},
+    label_list: Optional[List[str]] = field(
+        default=None, metadata={"help": "The string list of the label names. "}
     )
 
     @staticmethod

flaml/nlp/utils.py

@@ -83,7 +83,9 @@ def tokenize_and_align_labels(
 ):
     tokenized_inputs = tokenizer(
         [list(examples[X_sent_key])],
-        padding="max_length",
+        padding="max_length"
+        if hf_args.pad_to_max_length
+        else False,  # to be consistent with https://github.com/huggingface/transformers/blob/main/examples/pytorch/token-classification/run_ner.py#L394
         truncation=True,
         max_length=hf_args.max_seq_length,
         # We use this argument because the texts in our dataset are lists of words (with a label for each word).
@@ -113,11 +115,11 @@ def tokenize_and_align_labels(
                 # else:
                 #     label_ids.append(b_to_i_label[label_to_id[label[word_idx]]])
             previous_word_idx = word_idx
-        tokenized_inputs["label"] = label_ids
+        tokenized_inputs["labels"] = label_ids
     tmp_column_names = sorted(tokenized_inputs.keys())
     tokenized_input_and_labels = [tokenized_inputs[x] for x in tmp_column_names]
     for key_idx, each_key in enumerate(tmp_column_names):
-        if each_key != "label":
+        if each_key != "labels":
             tokenized_input_and_labels[key_idx] = tokenized_input_and_labels[key_idx][0]
     if return_column_name:
         return tokenized_input_and_labels, tmp_column_names
@@ -151,7 +153,7 @@ def tokenize_text_tokclassification(X, Y, tokenizer, hf_args=None):
             axis=1,
             result_type="expand",
         )
-        label_idx = tokenized_column_names.index("label")
+        label_idx = tokenized_column_names.index("labels")
         other_indices = sorted(
             set(range(len(tokenized_column_names))).difference({label_idx})
         )

test/automl/test_classification.py

@@ -347,3 +347,4 @@ class TestClassification(unittest.TestCase):
 
 if __name__ == "__main__":
     unittest.main()
+    test = TestClassification()

test/nlp/test_autohf_tokenclassification.py

@@ -13,7 +13,18 @@ def test_tokenclassification():
 
     automl_settings = get_automl_settings()
     automl_settings["task"] = "token-classification"
-    automl_settings["metric"] = "seqeval"
+    automl_settings["metric"] = "seqeval:overall_f1"  # evaluating based on the overall_f1 of seqeval
+    automl_settings["fit_kwargs_by_estimator"]["transformer"]["label_list"] = [
+        "O",
+        "B-PER",
+        "I-PER",
+        "B-ORG",
+        "I-ORG",
+        "B-LOC",
+        "I-LOC",
+        "B-MISC",
+        "I-MISC",
+    ]
 
     try:
         automl.fit(