MultiLabel-MultiClass Model for Joint Sequence Tagging (facebookresea…

…rch#1335)

Summary:
Pull Request resolved: facebookresearch#1335

We need to support multi-class as well as multi-label prediction for joint models in pytext. This diff implements a

1. Joint Multi Label Decoder
2. MultiLabelClassification Output Layer
3. Loss computation for multi-label-multi-class scenarios
4. Label weights per label and per class
5. Softmax options for output layers
6. Custom Metric Reporter, Metric Class and Output for flow

Reviewed By: seayoung1112

Differential Revision: D20210880

fbshipit-source-id: 701ca0a32302f923f13efe012618bba693b2d4db

pytext/data/tensorizers.py

@@ -853,8 +853,8 @@ def numberize(self, row):
                     label_idx_list.append(self.pad_idx)
                 else:
                     raise Exception(
-                        "Found none or empty value in the list,"
-                        + " while pad_missing is disabled"
+                        "Found none or empty value in the list, \
+                        while pad_missing is disabled"
                     )
             else:
                 label_idx_list.append(self.vocab.lookup_all(label))

pytext/metric_reporters/__init__.py

@@ -15,6 +15,7 @@
 from .regression_metric_reporter import RegressionMetricReporter
 from .squad_metric_reporter import SquadMetricReporter
 from .word_tagging_metric_reporter import (
+    MultiLabelSequenceTaggingMetricReporter,
     NERMetricReporter,
     SequenceTaggingMetricReporter,
     WordTaggingMetricReporter,
@@ -34,7 +35,7 @@
     "CompositionalMetricReporter",
     "PairwiseRankingMetricReporter",
     "SequenceTaggingMetricReporter",
-    "PureLossMetricReporter",
+    "MultiLabelSequenceTaggingMetricReporter",
     "NERMetricReporter",
     "DenseRetrievalMetricReporter",
 ]

pytext/metric_reporters/word_tagging_metric_reporter.py

@@ -13,6 +13,7 @@
     LabelPrediction,
     PRF1Metrics,
     compute_classification_metrics,
+    compute_multi_label_multi_class_soft_metrics,
 )
 from pytext.metrics.intent_slot_metrics import (
     Node,
@@ -92,6 +93,73 @@ def get_model_select_metric(self, metrics):
         return metrics.micro_scores.f1
 
 
+class MultiLabelSequenceTaggingMetricReporter(MetricReporter):
+    def __init__(self, label_names, pad_idx, channels, label_vocabs=None):
+        super().__init__(channels)
+        self.label_names = label_names
+        self.pad_idx = pad_idx
+        self.label_vocabs = label_vocabs
+
+    @classmethod
+    def from_config(cls, config, tensorizers):
+        return MultiLabelSequenceTaggingMetricReporter(
+            channels=[ConsoleChannel(), FileChannel((Stage.TEST,), config.output_path)],
+            label_names=tensorizers.keys(),
+            pad_idx=[v.pad_idx for _, v in tensorizers.items()],
+            label_vocabs=[v.vocab._vocab for _, v in tensorizers.items()],
+        )
+
+    def calculate_metric(self):
+        if len(self.all_scores) == 0:
+            return {}
+        list_score_pred_expect = []
+        for label_idx in range(0, len(self.label_names)):
+            list_score_pred_expect.append(
+                list(
+                    itertools.chain.from_iterable(
+                        (
+                            LabelPrediction(s, p, e)
+                            for s, p, e in zip(scores, pred, expect)
+                            if e != self.pad_idx[label_idx]
+                        )
+                        for scores, pred, expect in zip(
+                            self.all_scores[label_idx],
+                            self.all_preds[label_idx],
+                            self.all_targets[label_idx],
+                        )
+                    )
+                )
+            )
+        metrics = compute_multi_label_multi_class_soft_metrics(
+            list_score_pred_expect,
+            self.label_names,
+            self.label_vocabs,
+            self.calculate_loss(),
+        )
+        return metrics
+
+    def batch_context(self, raw_batch, batch):
+        return {}
+
+    @staticmethod
+    def get_model_select_metric(metrics):
+        if isinstance(metrics, dict):
+            # There are multiclass precision/recall labels
+            # Compute average precision
+            avg_precision = 0.0
+            for _, metric in metrics.items():
+                if metric:
+                    avg_precision += sum(
+                        v.average_precision
+                        for k, v in metric.items()
+                        if v.average_precision > 0
+                    ) / (len(metric.keys()) * 1.0)
+            avg_precision = avg_precision / (len(metrics.keys()) * 1.0)
+        else:
+            avg_precision = metrics.accuracy
+        return avg_precision
+
+
 class SequenceTaggingMetricReporter(MetricReporter):
     def __init__(self, label_names, pad_idx, channels):
         super().__init__(channels)

pytext/metrics/__init__.py

@@ -753,6 +753,37 @@ def compute_multi_label_soft_metrics(
     return soft_metrics
 
 
+def compute_multi_label_multi_class_soft_metrics(
+    predictions: Sequence[Sequence[LabelListPrediction]],
+    label_names: Sequence[str],
+    label_vocabs: Sequence[Sequence[str]],
+    recall_at_precision_thresholds: Sequence[float] = RECALL_AT_PRECISION_THRESHOLDS,
+    precision_at_recall_thresholds: Sequence[float] = PRECISION_AT_RECALL_THRESHOLDS,
+) -> Dict[int, SoftClassificationMetrics]:
+    """
+
+    Computes multi-label soft classification metrics with multi-class accommodation
+
+    Args:
+        predictions: multi-label predictions,
+                     including the confidence score for each label.
+        label_names: Indexed label names.
+        recall_at_precision_thresholds: precision thresholds at which to calculate
+            recall
+        precision_at_recall_thresholds: recall thresholds at which to calculate
+            precision
+
+
+    Returns:
+        Dict from label strings to their corresponding soft metrics.
+    """
+    soft_metrics = {}
+    for label_idx, label_vocab in enumerate(label_vocabs):
+        label = list(label_names)[label_idx]
+        soft_metrics[label] = compute_soft_metrics(predictions[label_idx], label_vocab)
+    return soft_metrics
+
+
 def compute_matthews_correlation_coefficients(
     TP: int, FP: int, FN: int, TN: int
 ) -> float:

pytext/models/decoders/multilabel_decoder.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+from typing import Dict, List
+
+import torch
+import torch.nn as nn
+from pytext.utils.usage import log_class_usage
+
+from .decoder_base import DecoderBase
+
+
+class MultiLabelDecoder(DecoderBase):
+    """
+    Implements a 'n-tower' MLP: one for each of the multi labels
+    Used in USM/EA: the user satisfaction modeling, pTSR prediction and
+    Error Attribution are all 3 label sets that need predicting.
+
+    """
+
+    class Config(DecoderBase.Config):
+        # Intermediate hidden dimensions
+        hidden_dims: List[int] = []
+
+    def __init__(
+        self,
+        config: Config,
+        in_dim: int,
+        output_dim: Dict[str, int],
+        label_names: List[str],
+    ) -> None:
+        super().__init__(config)
+        self.label_mlps = nn.ModuleDict({})
+        # Store the ordered list to preserve the ordering of the labels
+        # when generating the output layer
+        self.label_names = label_names
+        aggregate_out_dim = 0
+        for label_, _ in output_dim.items():
+            self.label_mlps[label_] = MultiLabelDecoder.get_mlp(
+                in_dim, output_dim[label_], config.hidden_dims
+            )
+            aggregate_out_dim += output_dim[label_]
+        self.out_dim = (1, aggregate_out_dim)
+        log_class_usage(__class__)
+
+    @staticmethod
+    def get_mlp(in_dim: int, out_dim: int, hidden_dims: List[int]):
+        layers = []
+        current_dim = in_dim
+        for dim in hidden_dims or []:
+            layers.append(nn.Linear(current_dim, dim))
+            layers.append(nn.ReLU())
+            current_dim = dim
+        layers.append(nn.Linear(current_dim, out_dim))
+        return nn.Sequential(*layers)
+
+    def forward(self, *input: torch.Tensor):
+        logits = tuple(
+            self.label_mlps[x](torch.cat(input, 1)) for x in self.label_names
+        )
+        return logits
+
+    def get_decoder(self) -> List[nn.Module]:
+        return self.label_mlps

pytext/models/output_layers/doc_classification_output_layer.py

@@ -128,7 +128,7 @@ def forward(self, logits: torch.Tensor):
 class BinaryClassificationOutputLayer(ClassificationOutputLayer):
     def get_pred(self, logit, *args, **kwargs):
         """See `OutputLayerBase.get_pred()`."""
-        preds = torch.max(logit, 1)[1]
+        preds = torch.max(logit, -1)[1]
         scores = F.logsigmoid(logit)
         return preds, scores
 
@@ -153,7 +153,7 @@ def export_to_caffe2(
 class MulticlassOutputLayer(ClassificationOutputLayer):
     def get_pred(self, logit, *args, **kwargs):
         """See `OutputLayerBase.get_pred()`."""
-        preds = torch.max(logit, 1)[1]
+        preds = torch.max(logit, -1)[1]
         scores = F.log_softmax(logit, 1)
         return preds, scores