Merge pull request #3140 from patrickvonplaten/merge_bart_generate_into_default_generate

Merge bart generate into default generate

examples/summarization/bart/evaluate_cnn.py

@@ -28,7 +28,7 @@ def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
             num_beams=4,
             length_penalty=2.0,
             max_length=140,
-            min_len=55,
+            min_length=55,
             no_repeat_ngram_size=3,
         )
         dec = [tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summaries]

src/transformers/configuration_bart.py

@@ -40,8 +40,9 @@ class BartConfig(PretrainedConfig):
         self,
         activation_dropout=0.0,
         vocab_size=50265,
+        bos_token_id=0,
         pad_token_id=1,
-        eos_token_id=2,
+        eos_token_ids=[2],
         d_model=1024,
         encoder_ffn_dim=4096,
         encoder_layers=12,
@@ -58,7 +59,7 @@ class BartConfig(PretrainedConfig):
         classifier_dropout=0.0,
         output_past=False,
         num_labels=3,
-        bos_token_id=0,
+        is_encoder_decoder=True,
         **common_kwargs
     ):
         r"""
@@ -72,11 +73,12 @@ class BartConfig(PretrainedConfig):
             output_past=output_past,
             pad_token_id=pad_token_id,
             bos_token_id=bos_token_id,
+            eos_token_ids=eos_token_ids,
+            is_encoder_decoder=is_encoder_decoder,
             **common_kwargs,
         )
         self.vocab_size = vocab_size
         self.d_model = d_model  # encoder_embed_dim and decoder_embed_dim
-        self.eos_token_id = eos_token_id
         self.encoder_ffn_dim = encoder_ffn_dim
         self.encoder_layers = self.num_hidden_layers = encoder_layers
         self.encoder_attention_heads = encoder_attention_heads

src/transformers/configuration_t5.py

@@ -75,9 +75,12 @@ class T5Config(PretrainedConfig):
         dropout_rate=0.1,
         layer_norm_epsilon=1e-6,
         initializer_factor=1.0,
+        is_encoder_decoder=True,
         **kwargs
     ):
-        super().__init__(**kwargs)
+        super().__init__(
+            is_encoder_decoder=is_encoder_decoder, **kwargs,
+        )
         self.vocab_size = vocab_size
         self.n_positions = n_positions
         self.d_model = d_model

src/transformers/configuration_utils.py

@@ -65,10 +65,12 @@ class PretrainedConfig(object):
         self.pruned_heads = kwargs.pop("pruned_heads", {})
 
         # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
+        self.is_encoder_decoder = kwargs.pop("is_encoder_decoder", False)
         self.is_decoder = kwargs.pop("is_decoder", False)
 
         # Parameters for sequence generation
         self.max_length = kwargs.pop("max_length", 20)
+        self.min_length = kwargs.pop("min_length", 0)
         self.do_sample = kwargs.pop("do_sample", False)
         self.early_stopping = kwargs.pop("early_stopping", False)
         self.num_beams = kwargs.pop("num_beams", 1)
@@ -80,6 +82,7 @@ class PretrainedConfig(object):
         self.pad_token_id = kwargs.pop("pad_token_id", None)
         self.eos_token_ids = kwargs.pop("eos_token_ids", None)
         self.length_penalty = kwargs.pop("length_penalty", 1.0)
+        self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
         self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
 
         # Fine-tuning task arguments

src/transformers/modeling_bart.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """PyTorch BART model, ported from the fairseq repo."""
 import logging
-import math
 import random
 from typing import Dict, List, Optional, Tuple
 
@@ -24,7 +23,7 @@ from torch import Tensor, nn
 
 from .configuration_bart import BartConfig
 from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from .modeling_utils import BeamHypotheses, PreTrainedModel, create_position_ids_from_input_ids
+from .modeling_utils import PreTrainedModel, create_position_ids_from_input_ids
 
 
 logger = logging.getLogger(__name__)
@@ -942,21 +941,31 @@ class BartForConditionalGeneration(PretrainedBartModel):
 
         return outputs
 
-    @staticmethod
-    def prepare_inputs_for_generation(input_ids, past, decoder_input_ids, attention_mask):
+    def prepare_inputs_for_generation(self, decoder_input_ids, past, encoder_inputs, attention_mask):
+        assert attention_mask.shape == encoder_inputs.shape, "attn_mask.shape != encoder_input.shape: {} =! {}".format(
+            attention_mask.shape, encoder_inputs.shape
+        )
         if past is None:  # first step
             encoder_outputs, decoder_cached_states = None, None
         else:
             encoder_outputs, decoder_cached_states = past
+
+        input_ids = encoder_inputs
         return {
             "input_ids": input_ids,  # ignored after first pass
+            "encoder_outputs": encoder_outputs,
             "decoder_cached_states": decoder_cached_states,
             "decoder_input_ids": decoder_input_ids,
-            "encoder_outputs": encoder_outputs,
             "attention_mask": attention_mask,
-            # "decoder_attention_mask": decoder_attention_mask,
         }
 
+    def prepare_scores_for_generation(self, scores, cur_len, max_length):
+        if cur_len == 1:
+            self._force_token_ids_generation(scores, self.config.bos_token_id)
+        if cur_len == max_length - 1 and self.config.eos_token_ids[0] is not None:
+            self._force_token_ids_generation(scores, self.config.eos_token_ids[0])
+        return scores
+
     @staticmethod
     def _reorder_cache(past, beam_idx):
         ((enc_out, enc_mask), decoder_cached_states) = past
@@ -978,272 +987,6 @@ class BartForConditionalGeneration(PretrainedBartModel):
     def get_output_embeddings(self):
         return self.lm_head
 
-    @torch.no_grad()
-    def generate(
-        self,
-        input_ids,
-        attention_mask=None,
-        max_length=20,
-        num_beams=1,
-        repetition_penalty=1.0,
-        length_penalty=1.0,
-        num_return_sequences=1,
-        min_len=0,
-        no_repeat_ngram_size=0,
-    ):
-        r""" Generates summaries using the lm-head and greedy beam search
-
-        Adapted in part from Facebook's `XLM beam search code`_ and `Fairseq beam search code`_.
-
-        .. _`XLM beam search code`:
-           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529
-        .. _`Fairseq beam search code`:
-           https://github.com/pytorch/fairseq/blob/master/fairseq/sequence_generator.py
-
-
-        Parameters:
-
-            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
-                The sequence used as a prompt for the generation. If `None` the method initializes
-                it as an empty `torch.LongTensor` of shape `(1,)`.
-
-            max_length: (`optional`) int
-                The max length of the sequence to be generated. Does not include tokens in input_ids.
-
-            num_beams: (`optional`) int
-                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.
-
-            repetition_penalty: (`optional`) float
-                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.
-
-            length_penalty: (`optional`) float
-                Exponential penalty to the length. Default to 1.
-
-            num_return_sequences: (`optional`) int
-                The number of independently computed returned sequences for each element in the batch. Default to 1.
-
-            min_len: (`optional`) int
-
-        Returns:
-            `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
-                sequence_length is <= max_length (examples can finish early)
-
-        Examples::
-            from transformers import BartTokenizer, BartForConditionalGeneration, BartConfig
-            # see ``examples/summarization/bart/evaluate_cnn.py`` for a longer example
-            model = BartForConditionalGeneration.from_pretrained('bart-large-cnn')
-            tokenizer = BartTokenizer.from_pretrained('bart-large-cnn')
-            ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
-            inputs = tokenizer.batch_encode_plus([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
-            # Generate Summary
-            summary_ids = model.generate(inputs['input_ids'], attention_mask=inputs['attention_mask'], num_beams=4, max_length=5)
-            print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids])
-        """
-        bos_token_id = self.config.bos_token_id
-        pad_token_id = self.config.pad_token_id
-        eos_token_id = self.config.eos_token_id
-        batch_size, cur_len = input_ids.shape
-        assert input_ids is not None
-        assert self.config.output_past, "Generating with bart requires instantiating a config with output_past=True"
-        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
-        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
-        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
-        assert isinstance(pad_token_id, int)
-        assert bos_token_id == 0, "configurable bos_token_id not yet supported"
-        assert length_penalty > 0, "`length_penalty` should be strictly positive."
-        assert (
-            isinstance(num_return_sequences, int) and num_return_sequences > 0
-        ), "`num_return_sequences` should be a positive integer."
-
-        # current position and vocab size
-        cur_len = input_ids.shape[1]
-        vocab_size = self.config.vocab_size
-
-        if num_return_sequences != 1:
-            # Expand input to num return sequences
-            input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
-            input_ids = input_ids.contiguous().view(
-                batch_size * num_return_sequences, cur_len
-            )  # shape: (batch_size * num_return_sequences, cur_len)
-            batch_size *= num_return_sequences
-
-        # Below here somewhat similar to PretrainedModel._generate_beam_search
-        # Expand input to num beams
-        input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)
-
-        input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len)  # (batch_size * num_beams, cur_len)
-        if attention_mask is not None:
-            attention_mask = (
-                attention_mask.unsqueeze(1)
-                .expand(batch_size, num_beams, cur_len)
-                .contiguous()
-                .view(batch_size * num_beams, cur_len)
-            )  # RESHAPE
-
-        # generated hypotheses
-        finalized_hyps = [  # they end in EOS and we wont work on them more!
-            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=True) for _ in range(batch_size)
-        ]
-
-        # scores for each sentence in the beam
-        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
-        beam_scores[:, 1:] = -1e9  # avoid ties in first step
-        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)
-
-        # decoder tokens
-        prev_output_tokens = input_ids.new(batch_size * num_beams, 1).long().fill_(-1)
-        prev_output_tokens[:, 0] = 2  # HARDCODED EOS, which will be removed at the end.
-        decoder_cache = None
-        done = [False for _ in range(batch_size)]  # done sentences
-
-        self.model.decoder.generation_mode = True  # tells decoder not to use causal mask
-        for step in range(max_length + 1):
-            decoder_input_ids = prev_output_tokens.clone()
-            model_inputs = self.prepare_inputs_for_generation(
-                input_ids, decoder_cache, decoder_input_ids, attention_mask,
-            )
-            outputs = self(**model_inputs)
-            lprobs = F.log_softmax(outputs[0][:, -1, :], dim=-1)
-
-            lprobs[lprobs != lprobs] = -math.inf  # block nans
-            lprobs[:, pad_token_id] = -math.inf
-            # TODO(SS): fairseq also takes out <unk> every step, and has unk at slot 3
-
-            if step == 0:  # Force BOS to be chosen
-                lprobs[:, bos_token_id + 1 :] = -math.inf
-            elif step < min_len:  # Prevent EOS from being chosen
-                lprobs[:, eos_token_id] = -math.inf
-            elif step == max_length:  # FORCE EOS to be chosen
-                lprobs[:, :eos_token_id] = -math.inf
-                lprobs[:, eos_token_id + 1 :] = -math.inf
-            assert self._do_output_past(outputs)
-            decoder_cache = outputs[1]
-            if repetition_penalty != 1.0:
-                self.enforce_repetition_penalty_(lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty)
-            num_hypos = batch_size * num_beams
-            if no_repeat_ngram_size > 0:  # copied from fairseq
-                # for each sentence, calculate a list of banned tokens to prevent repetitively generating the same ngrams
-                banned_tokens = self.calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step)
-                # then set their probabilities tof -inf
-                for idx in range(num_hypos):
-                    lprobs[idx, banned_tokens[idx]] = -math.inf
-            assert lprobs.size() == (batch_size * num_beams, vocab_size)
-            _scores = lprobs + beam_scores[:, None].expand_as(lprobs)  # (batch_size * num_beams, vocab_size)
-
-            # re-organize to group the beam together (we are keeping top hypothesis across beams)
-            _scores = _scores.view(batch_size, num_beams * vocab_size)  # (batch_size, num_beams * vocab_size)
-            # Take the best 2 x beam_size predictions for each example, we'll choose the first beam_size of these which don't predict eos to continue with.
-            next_scores, next_words = torch.topk(_scores, 2 * num_beams)
-            assert next_scores.size() == next_words.size() == (batch_size, 2 * num_beams)
-
-            # list of (batch_size * num_beams)
-            next_batch_beam = []  # Tuple(next score, next word, current position in the batch)
-            for batch_idx in range(batch_size):
-                # if we are done with this sentence (because we can't improve)
-                if done[batch_idx]:  # then pad all associated hypotheses
-                    assert (
-                        len(finalized_hyps[batch_idx]) >= num_beams
-                    ), "Example can only be done if at least {} beams have been generated".format(num_beams)
-                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
-                    continue
-
-                # Otherwise generate some next word choices
-                next_sent_beam = []
-                # add next words for this sentence
-                for i, (idx, score) in enumerate(zip(next_words[batch_idx], next_scores[batch_idx])):
-                    beam_id = idx // vocab_size
-                    word_id = idx % vocab_size
-                    assert prev_output_tokens.shape[1] == (step + 1)
-                    if word_id.item() == eos_token_id:
-                        if i >= num_beams:
-                            continue
-                        finalized_hyps[batch_idx].add(
-                            prev_output_tokens[batch_idx * num_beams + beam_id].clone(), score.item(),
-                        )
-                    else:
-                        next_sent_beam.append((score, word_id, batch_idx * num_beams + beam_id))
-
-                    if len(next_sent_beam) == num_beams:  # TODO(SS): can we delete this?
-                        break
-                # Check if were done so that we can save a pad step if all(done)
-                done[batch_idx] = done[batch_idx] or finalized_hyps[batch_idx].is_done(
-                    next_scores[batch_idx].max().item(), cur_len=step + 1,
-                )
-                assert len(next_sent_beam) == num_beams, "Beam should always be full"
-                next_batch_beam.extend(next_sent_beam)
-                assert len(next_batch_beam) == num_beams * (batch_idx + 1)
-
-            if all(done):
-                break
-
-            # sanity check / prepare next batch
-            assert len(next_batch_beam) == batch_size * num_beams
-            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
-            beam_words = input_ids.new([x[1] for x in next_batch_beam])
-            beam_idx = input_ids.new([x[2] for x in next_batch_beam])
-            # re-order decoder inputs to [beam_idx]
-            prev_output_tokens = prev_output_tokens[beam_idx]
-            prev_output_tokens = torch.cat([prev_output_tokens, beam_words.unsqueeze(1)], dim=-1)
-
-            # re-order internal states
-            decoder_cache = self._reorder_cache(decoder_cache, beam_idx)
-
-        for batch_idx in range(batch_size):
-            # Add all open beam hypothesis to generated_hyps
-            if done[batch_idx]:
-                continue
-            offset = batch_idx * num_beams
-            for i in range(num_beams):
-                score = beam_scores[offset + i]
-                final_tokens = prev_output_tokens[offset + i]
-                finalized_hyps[batch_idx].add(final_tokens, score.item())
-
-        # select the best hypotheses
-        sent_lengths = input_ids.new(batch_size)
-        best = []
-        for i, hypotheses in enumerate(finalized_hyps):
-            best_hyp = max(hypotheses.beams, key=lambda x: x[0])[1]
-            sent_lengths[i] = len(best_hyp)
-            best.append(best_hyp)
-
-        # shorter batches are filled with pad_token
-        if sent_lengths.min().item() != sent_lengths.max().item():
-            # TODO(SS): decoded = torch.rnn.utils.pad_sequence(best, batch_first=True, padding_value=pad_token_id)
-            sent_max_len = min(sent_lengths.max().item() + 1, max_length + 1)  # TODO(SS): same as step?
-            decoded = input_ids.new(batch_size, sent_max_len).fill_(pad_token_id)
-            # fill with hypothesis and eos_token_id if necessary
-            for i, hypo in enumerate(best):
-                decoded[i, : sent_lengths[i]] = hypo
-                if sent_lengths[i] < max_length:
-                    decoded[i, sent_lengths[i]] = eos_token_id
-        else:
-            assert (len(hypo) == max_length for hypo in best)
-            decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
-        return decoded[:, 1:]  # get rid of starting EOS
-
-    @staticmethod
-    def calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step):
-        """Copied from fairseq for no_repeat_ngram in beam_search"""
-        # TODO(SS): this can go on parent if there is demand
-        if step + 2 < no_repeat_ngram_size:
-            return [
-                [] for _ in range(num_hypos)
-            ]  # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
-        gen_ngrams = [{} for _ in range(num_hypos)]
-        for idx in range(num_hypos):
-            gen_tokens = prev_output_tokens[idx].tolist()
-            for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
-                k = tuple(ngram[:-1])
-                gen_ngrams[idx][k] = gen_ngrams[idx].get(k, []) + [ngram[-1]]
-
-        def _get_generated_ngrams(hypo_idx):
-            """Before decoding the next token, prevent decoding of ngrams that have already appeared"""
-            ngram_index = tuple(prev_output_tokens[hypo_idx, step + 2 - no_repeat_ngram_size : step + 1].tolist())
-            return gen_ngrams[hypo_idx].get(ngram_index, [])
-
-        banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
-        return banned_tokens
-
 
 @add_start_docstrings(
     """Bart model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. """,
@@ -1313,7 +1056,7 @@ class BartForSequenceClassification(PretrainedBartModel):
             encoder_outputs=encoder_outputs,
         )
         x = outputs[0]  # last hidden state
-        eos_mask = input_ids.eq(self.config.eos_token_id)
+        eos_mask = input_ids.eq(self.config.eos_token_ids[0])
         if len(torch.unique(eos_mask.sum(1))) > 1:
             raise ValueError("All examples must have the same number of <eos> tokens.")
         sentence_representation = x[eos_mask, :].view(x.size(0), -1, x.size(-1))[:, -1, :]

tests/test_modeling_bart.py

@@ -60,7 +60,10 @@ class ModelTester:
         self.hidden_act = "gelu"
         self.hidden_dropout_prob = 0.1
         self.attention_probs_dropout_prob = 0.1
-        self.max_position_embeddings = 12
+        self.max_position_embeddings = 20
+        self.eos_token_id = 2
+        self.pad_token_id = 1
+        self.bos_token_id = 0
         torch.manual_seed(0)
 
     def prepare_config_and_inputs_for_common(self):
@@ -79,6 +82,9 @@ 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],
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
         )
         inputs_dict = prepare_bart_inputs_dict(config, input_ids)
         return config, inputs_dict
@@ -101,6 +107,7 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase):
     all_model_classes = (
         (BartModel, BartForConditionalGeneration, BartForSequenceClassification) if is_torch_available() else ()
     )
+    all_generative_model_classes = (BartForConditionalGeneration,) if is_torch_available() else ()
     is_encoder_decoder = True
     # TODO(SS): fix the below in a separate PR
     test_pruning = False
@@ -207,6 +214,9 @@ class BartHeadTests(unittest.TestCase):
             decoder_ffn_dim=32,
             max_position_embeddings=48,
             output_past=output_past,
+            eos_token_ids=[2],
+            pad_token_id=1,
+            bos_token_id=0,
         )
         return config, input_ids, batch_size
 
@@ -266,14 +276,18 @@ class BartHeadTests(unittest.TestCase):
             decoder_ffn_dim=32,
             max_position_embeddings=48,
             output_past=True,
+            eos_token_ids=[2],
+            pad_token_id=1,
+            bos_token_id=0,
         )
         lm_model = BartForConditionalGeneration(config).to(torch_device)
         lm_model.eval()
 
+        max_length = 5
         new_input_ids = lm_model.generate(
-            input_ids.clone(), num_return_sequences=1, num_beams=2, no_repeat_ngram_size=3, max_length=5
+            input_ids.clone(), num_return_sequences=1, num_beams=2, no_repeat_ngram_size=3, max_length=max_length
         )
-        self.assertEqual(new_input_ids.shape, (input_ids.shape[0], 5))
+        self.assertEqual(new_input_ids.shape, (input_ids.shape[0], max_length - 1))
         # TODO(SS): uneven length batches, empty inputs
 
     def test_shift_tokens_right(self):
@@ -300,9 +314,10 @@ class BartHeadTests(unittest.TestCase):
     @unittest.skipIf(torch_device == "cpu", "Cant do half precision")
     def test_generate_fp16(self):
         config, input_ids, batch_size = self._get_config_and_data(output_past=True)
-        attention_mask = input_ids.ne(1)
+        input_ids = input_ids
+        attention_mask = input_ids.ne(1).to(torch_device)
         lm_model = BartForConditionalGeneration(config).eval().to(torch_device).half()
-        lm_model.generate(input_ids, attention_mask)
+        lm_model.generate(input_ids, attention_mask=attention_mask)
 
     def test_prepare_bart_decoder_inputs(self):
         config, *_ = self._get_config_and_data(output_past=False)
@@ -410,18 +425,24 @@ class BartModelIntegrationTest(unittest.TestCase):
             self.assertIsNotNone(model)
 
     @slow
-    def test_cnn_summarization_same_as_fairseq(self):
+    def test_cnn_summarization_same_as_fairseq_easy(self):
         hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
         tok = BartTokenizer.from_pretrained("bart-large")
         text = " (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian"
         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,)  # repetition_penalty=10.,
+        gen_tokens = hf.generate(
+            tokens, num_beams=4, max_length=extra_len + 2, do_sample=False
+        )  # 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]
         self.assertEqual(expected_result, generated[0])
 
-        # Harder cases with batching
+    @slow
+    def test_cnn_summarization_same_as_fairseq_hard(self):
+        hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
+        tok = BartTokenizer.from_pretrained("bart-large")
+
         FRANCE_ARTICLE = ' Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object.  Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot\'s license. Kumpa emphasized there\'s no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot\'s license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz\'s girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there\'s more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren\'t going to keep doing their job and they\'re upset about that and so they\'re suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person\'s problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN\'s Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN\'s Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report.'  # @noqa
         EXPECTED_SUMMARY_FRANCE = 'French prosecutor says he\'s not aware of any video footage from on board the plane. German daily Bild and French Paris Match claim to have found a cell phone video of the crash. A French Gendarmerie spokesman calls the reports "completely wrong" and "unwarranted" German airline Lufthansa confirms co-pilot Andreas Lubitz had battled depression.'
 
@@ -441,19 +462,27 @@ class BartModelIntegrationTest(unittest.TestCase):
             pad_to_max_length=True,
             return_tensors="pt",
         )
+
+        max_length = 140
+        min_length = 55
+
         self.assertEqual(1024, dct["input_ids"].shape[1])
         hypotheses_batch = hf.generate(
             input_ids=dct["input_ids"].to(torch_device),
             attention_mask=dct["attention_mask"].to(torch_device),
             num_beams=4,
             length_penalty=2.0,
-            max_length=140,
-            min_len=55,
+            max_length=max_length + 2,
+            min_length=min_length + 1,
             no_repeat_ngram_size=3,
+            do_sample=False,
+            early_stopping=True,
         )
+
         decoded = [
             tok.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in hypotheses_batch
         ]
+
         self.assertListEqual(
             [EXPECTED_SUMMARY_FRANCE, EXPECTED_SUMMARY_SHORTER, EXPECTED_SUMMARY_IRAN, EXPECTED_SUMMARY_SUBWAY],
             decoded,

tests/test_modeling_common.py

@@ -628,6 +628,9 @@ class ModelTesterMixin:
             "input_ids", None
         )  # TODO (PVP): ugly workaround to make code work for t5 for the moment - has to changed when t5 is fixed.
 
+        if self.is_encoder_decoder:
+            config.output_past = True  # needed for Bart TODO: might have to update for other encoder-decoder models
+
         for model_class in self.all_generative_model_classes:
             model = model_class(config)
             model.to(torch_device)