Seq2seq trainer (#9241)

* Add label smoothing in Trainer

* Add options for scheduler and Adafactor in Trainer

* Put Seq2SeqTrainer in the main lib

* Apply suggestions from code review

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Address review comments and adapt scripts

* Documentation

* Move test not using script to tests folder

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

tests/test_trainer_seq2seq.py

@@ -0,0 +1,135 @@
+# coding=utf-8
+# Copyright 2020 the HuggingFace Inc. team.
+#
+# 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.
+
+from transformers import BertTokenizer, EncoderDecoderModel, Seq2SeqTrainer, Seq2SeqTrainingArguments
+from transformers.file_utils import is_datasets_available
+from transformers.testing_utils import TestCasePlus, require_datasets, slow
+
+
+if is_datasets_available():
+    import datasets
+
+
+class Seq2seqTrainerTester(TestCasePlus):
+    @slow
+    @require_datasets
+    def test_finetune_bert2bert(self):
+
+        bert2bert = EncoderDecoderModel.from_encoder_decoder_pretrained("prajjwal1/bert-tiny", "prajjwal1/bert-tiny")
+        tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+
+        bert2bert.config.vocab_size = bert2bert.config.encoder.vocab_size
+        bert2bert.config.eos_token_id = tokenizer.sep_token_id
+        bert2bert.config.decoder_start_token_id = tokenizer.cls_token_id
+        bert2bert.config.max_length = 128
+
+        train_dataset = datasets.load_dataset("cnn_dailymail", "3.0.0", split="train[:1%]")
+        val_dataset = datasets.load_dataset("cnn_dailymail", "3.0.0", split="validation[:1%]")
+
+        train_dataset = train_dataset.select(range(32))
+        val_dataset = val_dataset.select(range(16))
+
+        rouge = datasets.load_metric("rouge")
+
+        batch_size = 4
+
+        def _map_to_encoder_decoder_inputs(batch):
+            # Tokenizer will automatically set [BOS] <text> [EOS]
+            inputs = tokenizer(batch["article"], padding="max_length", truncation=True, max_length=512)
+            outputs = tokenizer(batch["highlights"], padding="max_length", truncation=True, max_length=128)
+            batch["input_ids"] = inputs.input_ids
+            batch["attention_mask"] = inputs.attention_mask
+
+            batch["decoder_input_ids"] = outputs.input_ids
+            batch["labels"] = outputs.input_ids.copy()
+            batch["labels"] = [
+                [-100 if token == tokenizer.pad_token_id else token for token in labels] for labels in batch["labels"]
+            ]
+            batch["decoder_attention_mask"] = outputs.attention_mask
+
+            assert all([len(x) == 512 for x in inputs.input_ids])
+            assert all([len(x) == 128 for x in outputs.input_ids])
+
+            return batch
+
+        def _compute_metrics(pred):
+            labels_ids = pred.label_ids
+            pred_ids = pred.predictions
+
+            # all unnecessary tokens are removed
+            pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
+            label_str = tokenizer.batch_decode(labels_ids, skip_special_tokens=True)
+
+            rouge_output = rouge.compute(predictions=pred_str, references=label_str, rouge_types=["rouge2"])[
+                "rouge2"
+            ].mid
+
+            return {
+                "rouge2_precision": round(rouge_output.precision, 4),
+                "rouge2_recall": round(rouge_output.recall, 4),
+                "rouge2_fmeasure": round(rouge_output.fmeasure, 4),
+            }
+
+        # map train dataset
+        train_dataset = train_dataset.map(
+            _map_to_encoder_decoder_inputs,
+            batched=True,
+            batch_size=batch_size,
+            remove_columns=["article", "highlights"],
+        )
+        train_dataset.set_format(
+            type="torch",
+            columns=["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask", "labels"],
+        )
+
+        # same for validation dataset
+        val_dataset = val_dataset.map(
+            _map_to_encoder_decoder_inputs,
+            batched=True,
+            batch_size=batch_size,
+            remove_columns=["article", "highlights"],
+        )
+        val_dataset.set_format(
+            type="torch",
+            columns=["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask", "labels"],
+        )
+
+        output_dir = self.get_auto_remove_tmp_dir()
+
+        training_args = Seq2SeqTrainingArguments(
+            output_dir=output_dir,
+            per_device_train_batch_size=batch_size,
+            per_device_eval_batch_size=batch_size,
+            predict_with_generate=True,
+            evaluation_strategy="steps",
+            do_train=True,
+            do_eval=True,
+            warmup_steps=0,
+            eval_steps=2,
+            logging_steps=2,
+        )
+
+        # instantiate trainer
+        trainer = Seq2SeqTrainer(
+            model=bert2bert,
+            args=training_args,
+            compute_metrics=_compute_metrics,
+            train_dataset=train_dataset,
+            eval_dataset=val_dataset,
+            tokenizer=tokenizer,
+        )
+
+        # start training
+        trainer.train()

tests/test_trainer_utils.py

@@ -22,7 +22,10 @@ from transformers.testing_utils import require_torch
 
 
 if is_torch_available():
-    from transformers.trainer_pt_utils import DistributedTensorGatherer
+    import torch
+
+    from transformers.modeling_outputs import SequenceClassifierOutput
+    from transformers.trainer_pt_utils import DistributedTensorGatherer, LabelSmoother
 
 
 @require_torch
@@ -56,3 +59,31 @@ class TrainerUtilsTest(unittest.TestCase):
         self.assertTrue(np.array_equal(result[0], predictions))
         self.assertTrue(np.array_equal(result[1][0], predictions))
         self.assertTrue(np.array_equal(result[1][1], predictions))
+
+    def test_label_smoothing(self):
+        epsilon = 0.1
+        num_labels = 12
+        random_logits = torch.randn(4, 5, num_labels)
+        random_labels = torch.randint(0, num_labels, (4, 5))
+        loss = torch.nn.functional.cross_entropy(random_logits.view(-1, num_labels), random_labels.view(-1))
+        model_output = SequenceClassifierOutput(loss=loss, logits=random_logits)
+        label_smoothed_loss = LabelSmoother(0.1)(model_output, random_labels)
+        log_probs = -torch.nn.functional.log_softmax(random_logits, dim=-1)
+        expected_loss = (1 - epsilon) * loss + epsilon * log_probs.mean()
+        self.assertTrue(torch.allclose(label_smoothed_loss, expected_loss))
+
+        # With a few -100 labels
+        random_labels[0, 1] = -100
+        random_labels[2, 1] = -100
+        random_labels[2, 3] = -100
+
+        loss = torch.nn.functional.cross_entropy(random_logits.view(-1, num_labels), random_labels.view(-1))
+        model_output = SequenceClassifierOutput(loss=loss, logits=random_logits)
+        label_smoothed_loss = LabelSmoother(0.1)(model_output, random_labels)
+        log_probs = -torch.nn.functional.log_softmax(random_logits, dim=-1)
+        # Mask the log probs with the -100 labels
+        log_probs[0, 1] = 0.0
+        log_probs[2, 1] = 0.0
+        log_probs[2, 3] = 0.0
+        expected_loss = (1 - epsilon) * loss + epsilon * log_probs.sum() / (num_labels * 17)
+        self.assertTrue(torch.allclose(label_smoothed_loss, expected_loss))