Better last batch logic (#12827) (#12840)

* Better last batch logic

changelog/12827.improvement.md

@@ -0,0 +1 @@
+Improved handling of last batch during DIET and TED training. The last batch is discarded if it contains less than half a batch size of data.

rasa/utils/tensorflow/data_generator.py

@@ -1,3 +1,4 @@
+import math
 from typing import List, Union, Text, Optional, Any, Tuple, Dict, cast
 
 import logging
@@ -380,7 +381,11 @@ def __len__(self) -> int:
         # data was rebalanced, so need to recalculate number of examples
         num_examples = self.model_data.number_of_examples(self._data)
         batch_size = self._current_batch_size
-        return num_examples // batch_size + int(num_examples % batch_size > 0)
+        # keep last batch only if it has at least half a batch size of examples
+        last_batch_half_full = num_examples % batch_size >= math.ceil(batch_size / 2)
+        num_batches = num_examples // batch_size + int(last_batch_half_full)
+        # Return at least 1 if there is an example
+        return max(num_batches, int(num_examples > 0))
 
     def __getitem__(self, index: int) -> Tuple[Any, Any]:
         """Gets batch at position `index`.

tests/nlu/classifiers/test_diet_classifier.py

@@ -22,6 +22,7 @@
     PREDICTED_CONFIDENCE_KEY,
     INTENT_NAME_KEY,
 )
+from rasa.utils import train_utils
 from rasa.utils.tensorflow.constants import (
     LOSS_TYPE,
     RANDOM_SEED,
@@ -896,3 +897,76 @@ async def test_sparse_feature_sizes_decreased_incremental_training(
         train_load_and_process_diet(
             finetune_classifier, pipeline=pipeline, training_data=iter2_path
         )
+
+
+@pytest.mark.timeout(120, func_only=True)
+@pytest.mark.parametrize(
+    "batch_size, expected_num_batches",
+    # the training dataset has 48 NLU examples
+    [
+        (1, 48),
+        (8, 6),
+        (15, 3),
+        (16, 3),
+        (18, 3),
+        (20, 2),
+        (32, 2),
+        (64, 1),
+        (128, 1),
+        (256, 1),
+    ],
+)
+async def test_dropping_of_last_partial_batch(
+    batch_size: int,
+    expected_num_batches: int,
+    create_diet: Callable[..., DIETClassifier],
+    train_and_preprocess: Callable[..., Tuple[TrainingData, List[GraphComponent]]],
+):
+    """test that diets data processing produces the right amount of batches.
+
+    We introduced a change to only keep the last incomplete batch if
+    1. it has more than 50% of examples of batch size
+    2. or it is the only batch in the epoch
+    """
+
+    pipeline = [
+        {"component": WhitespaceTokenizer},
+        {"component": CountVectorsFeaturizer},
+    ]
+    diet = create_diet({ENTITY_RECOGNITION: False, RANDOM_SEED: 1, EPOCHS: 1})
+    # This data set has 48 NLU examples
+    training_data, loaded_pipeline = train_and_preprocess(
+        pipeline, training_data="data/test/demo-rasa-no-ents.yml"
+    )
+
+    model_data = diet.preprocess_train_data(training_data)
+    data_generator, _ = train_utils.create_data_generators(model_data, batch_size, 1)
+
+    assert len(data_generator) == expected_num_batches
+
+
+@pytest.mark.timeout(120, func_only=True)
+async def test_dropping_of_last_partial_batch_empty_data(
+    create_diet: Callable[..., DIETClassifier],
+    train_and_preprocess: Callable[..., Tuple[TrainingData, List[GraphComponent]]],
+):
+    """test that diets data processing produces the right amount of batches.
+
+    We introduced a change to only keep the last incomplete batch if
+    1. it has more than 50% of examples of batch size
+    2. or it is the only batch in the epoch
+    """
+
+    pipeline = [
+        {"component": WhitespaceTokenizer},
+        {"component": CountVectorsFeaturizer},
+    ]
+    diet = create_diet({ENTITY_RECOGNITION: False, RANDOM_SEED: 1, EPOCHS: 1})
+    training_data, loaded_pipeline = train_and_preprocess(
+        pipeline, training_data=TrainingData()
+    )
+
+    model_data = diet.preprocess_train_data(training_data)
+    data_generator, _ = train_utils.create_data_generators(model_data, 64, 1)
+
+    assert len(data_generator) == 0