Merge pull request #3225 from patrickvonplaten/finalize_merge_bart_generate_into_default_generate

Complete merge Seq-2-Seq generation into default generation

examples/summarization/bart/evaluate_cnn.py

@@ -20,6 +20,10 @@ def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
     fout = Path(out_file).open("w")
     model = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(device)
     tokenizer = BartTokenizer.from_pretrained("bart-large")
+
+    max_length = 140
+    min_length = 55
+
     for batch in tqdm(list(chunks(lns, batch_size))):
         dct = tokenizer.batch_encode_plus(batch, max_length=1024, return_tensors="pt", pad_to_max_length=True)
         summaries = model.generate(
@@ -27,11 +31,12 @@ def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
             attention_mask=dct["attention_mask"].to(device),
             num_beams=4,
             length_penalty=2.0,
-            max_length=142,  # +2 from original because we start at step=1 and stop before max_length
-            min_length=56,  # +1 from original because we start at step=1
+            max_length=max_length + 2,  # +2 from original because we start at step=1 and stop before max_length
+            min_length=min_length + 1,  # +1 from original because we start at step=1
             no_repeat_ngram_size=3,
             early_stopping=True,
             do_sample=False,
+            decoder_start_token_id=model.config.eos_token_ids[0],
         )
         dec = [tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summaries]
         for hypothesis in dec:

src/transformers/modeling_utils.py

@@ -628,6 +628,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         no_repeat_ngram_size=None,
         num_return_sequences=None,
         attention_mask=None,
+        decoder_start_token_id=None,
     ):
         r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
         and beam-search.
@@ -739,6 +740,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         num_return_sequences = (
             num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
         )
+        # TODO: think about how to make this cleaner
+        decoder_start_token_id = (
+            decoder_start_token_id if decoder_start_token_id is not None else self.config.bos_token_id
+        )
 
         if input_ids is not None:
             batch_size = input_ids.shape[0]  # overriden by the input batch_size
@@ -765,6 +770,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         assert (eos_token_ids is None) or (
             isinstance(eos_token_ids, (list, tuple)) and ((isinstance(e, int) and e >= 0) for e in eos_token_ids)
         ), "`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
+        assert (
+            decoder_start_token_id is not None or self.config.is_encoder_decoder is False
+        ), "`decoder_start_token_id` has to be defined if model is encoder-decoder model"
         assert length_penalty > 0, "`length_penalty` should be strictly positive."
         assert (
             isinstance(no_repeat_ngram_size, int) and no_repeat_ngram_size >= 0
@@ -845,7 +853,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
             encoder_inputs = input_ids
             input_ids = torch.full(
                 (effective_batch_size * num_beams, 1),
-                bos_token_id,
+                decoder_start_token_id,  # TODO: see whether this is the best result
                 dtype=torch.long,
                 device=next(self.parameters()).device,
             )
@@ -1082,7 +1090,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
             scores = F.log_softmax(next_token_logits, dim=-1)  # (batch_size * num_beams, vocab_size)
             if self.config.is_encoder_decoder and do_sample is False:
-                # TODO(PVP) to be refactored later - do we need this boolean flag here? Also Only add for beam_search or also for no_beam_search? The prepare scores fn is ugly here
+                # TODO: maybe give better naming
                 scores = self.prepare_scores_for_generation(scores, cur_len, max_length)
 
             # set eos token prob to zero if min_length is not reached
@@ -1276,7 +1284,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
             decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
 
         if self.config.is_encoder_decoder:
-            # do not return first <EOS> token
             return decoded[:, 1:]
         return decoded
 

tests/test_modeling_bart.py

@@ -61,7 +61,7 @@ class ModelTester:
         self.hidden_dropout_prob = 0.1
         self.attention_probs_dropout_prob = 0.1
         self.max_position_embeddings = 20
-        self.eos_token_id = 2
+        self.eos_token_ids = [2]
         self.pad_token_id = 1
         self.bos_token_id = 0
         torch.manual_seed(0)
@@ -82,7 +82,7 @@ class ModelTester:
             dropout=self.hidden_dropout_prob,
             attention_dropout=self.attention_probs_dropout_prob,
             max_position_embeddings=self.max_position_embeddings,
-            eos_token_ids=[self.eos_token_id],
+            eos_token_ids=[2],
             bos_token_id=self.bos_token_id,
             pad_token_id=self.pad_token_id,
         )
@@ -438,7 +438,11 @@ class BartModelIntegrationTest(unittest.TestCase):
         tokens = tok.encode(text, return_tensors="pt").to(torch_device)
         extra_len = 20
         gen_tokens = hf.generate(
-            tokens, num_beams=4, max_length=extra_len + 2, do_sample=False
+            tokens,
+            num_beams=4,
+            max_length=extra_len + 2,
+            do_sample=False,
+            decoder_start_token_id=hf.config.eos_token_ids[0],
         )  # repetition_penalty=10.,
         expected_result = "<s>The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday."
         generated = [tok.decode(g,) for g in gen_tokens]
@@ -483,6 +487,7 @@ class BartModelIntegrationTest(unittest.TestCase):
             no_repeat_ngram_size=3,
             do_sample=False,
             early_stopping=True,
+            decoder_start_token_id=hf.config.eos_token_ids[0],
         )
 
         decoded = [