Skip to content

GitLab

Unverified Commit b8462b5b authored by Patrick von Platen's avatar Patrick von Platen Committed by GitHub
Browse files

[GenerationOutputs] Fix GenerationOutputs Tests (#9443) 

* fix generation models

* fix led

* fix docs

* add is_decoder

* fix last docstrings

* make style

* fix t5 cross attentions

* correct t5
parent 0c96262f
Hide whitespace changes
Inline Side-by-side
Showing with 89 additions and 53 deletions
src/transformers/generation_utils.py
View file @ b8462b5b
...@@ -113,21 +113,20 @@ class SampleDecoderOnlyOutput(ModelOutput): ...@@ -113,21 +113,20 @@ class SampleDecoderOnlyOutput(ModelOutput):
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_return_sequences, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
at each generation step. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of at each generation step. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of
shape :obj:`(batch_size * num_return_sequences, config.vocab_size)`). shape :obj:`(batch_size*num_return_sequences, config.vocab_size)`).
attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(num_return_sequences * batch_size, num_heads, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(num_return_sequences*batch_size, num_heads, generated_length,
sequence_length)`. sequence_length)`.
hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(num_return_sequences * batch_size, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(num_return_sequences*batch_size, generated_length, hidden_size)`.
hidden_size)`.
""" """
sequences: torch.LongTensor = None sequences: torch.LongTensor = None
...@@ -145,27 +144,26 @@ class SampleEncoderDecoderOutput(ModelOutput): ...@@ -145,27 +144,26 @@ class SampleEncoderDecoderOutput(ModelOutput):
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_return_sequences, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
at each generation step. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of at each generation step. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of
shape :obj:`(batch_size * num_return_sequences, config.vocab_size)`). shape :obj:`(batch_size*num_return_sequences, config.vocab_size)`).
encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape :obj:`(batch_size * Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape
num_return_sequences, num_heads, sequence_length, sequence_length)`. :obj:`(batch_size*num_return_sequences, num_heads, sequence_length, sequence_length)`.
encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
of shape :obj:`(batch_size * num_return_sequences, sequence_length, hidden_size)`. of shape :obj:`(batch_size*num_return_sequences, sequence_length, hidden_size)`.
decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequences, num_heads, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_return_sequences, num_heads, generated_length,
sequence_length)`. sequence_length)`.
decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequences, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_return_sequences, generated_length, hidden_size)`.
hidden_size)`.
""" """
sequences: torch.LongTensor = None sequences: torch.LongTensor = None
...@@ -182,23 +180,23 @@ class BeamSearchDecoderOnlyOutput(ModelOutput): ...@@ -182,23 +180,23 @@ class BeamSearchDecoderOnlyOutput(ModelOutput):
Base class for outputs of decoder-only generation models using beam search. Base class for outputs of decoder-only generation models using beam search.
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_return_sequences, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequences)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_return_sequences)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Final beam scores of the generated ``sequences``. Final beam scores of the generated ``sequences``.
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log
softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam
. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape :obj:`(batch_size . :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape
* num_beams * num_return_sequences, config.vocab_size)`). :obj:`(batch_size*num_beams*num_return_sequences, config.vocab_size)`).
attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams, num_heads, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams, num_heads, generated_length,
sequence_length)`. sequence_length)`.
hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams * num_return_sequences, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams*num_return_sequences, generated_length,
hidden_size)`. hidden_size)`.
""" """
...@@ -217,30 +215,30 @@ class BeamSearchEncoderDecoderOutput(ModelOutput): ...@@ -217,30 +215,30 @@ class BeamSearchEncoderDecoderOutput(ModelOutput):
attributes (respectively the decoder_attentions and the decoder_hidden_states attributes) attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_return_sequences, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequences)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_return_sequences)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Final beam scores of the generated ``sequences``. Final beam scores of the generated ``sequences``.
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log
softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam
. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape :obj:`(batch_size . :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape
* num_beams, config.vocab_size)`). :obj:`(batch_size*num_beams, config.vocab_size)`).
attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape :obj:`(batch_size, Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape :obj:`(batch_size,
num_heads, sequence_length, sequence_length)`. num_heads, sequence_length, sequence_length)`.
encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
of shape :obj:`(batch_size * num_beams * num_return_sequences, sequence_length, hidden_size)`. of shape :obj:`(batch_size*num_beams*num_return_sequences, sequence_length, hidden_size)`.
decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams * num_return_sequences, num_heads, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams*num_return_sequences, num_heads,
generated_length, sequence_length)`. generated_length, sequence_length)`.
decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams * num_return_sequences, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams*num_return_sequences, generated_length,
hidden_size)`. hidden_size)`.
""" """
...@@ -259,7 +257,7 @@ class BeamSampleDecoderOnlyOutput(ModelOutput): ...@@ -259,7 +257,7 @@ class BeamSampleDecoderOnlyOutput(ModelOutput):
Base class for outputs of decoder-only generation models using beam sample. Base class for outputs of decoder-only generation models using beam sample.
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_return_sequences, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequence)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequence)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
...@@ -267,15 +265,15 @@ class BeamSampleDecoderOnlyOutput(ModelOutput): ...@@ -267,15 +265,15 @@ class BeamSampleDecoderOnlyOutput(ModelOutput):
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log
softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam
. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape :obj:`(batch_size . :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape
* num_beams * num_return_sequences, config.vocab_size)`). :obj:`(batch_size*num_beams*num_return_sequences, config.vocab_size)`).
attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams, num_heads, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams, num_heads, generated_length,
sequence_length)`. sequence_length)`.
hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams, generated_length, hidden_size)`. :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams, generated_length, hidden_size)`.
""" """
sequences: torch.LongTensor = None sequences: torch.LongTensor = None
...@@ -293,7 +291,7 @@ class BeamSampleEncoderDecoderOutput(ModelOutput): ...@@ -293,7 +291,7 @@ class BeamSampleEncoderDecoderOutput(ModelOutput):
encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes) encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)
Args: Args:
sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`): sequences (:obj:`torch.LongTensor` of shape :obj:`(batch_size*num_beams, sequence_length)`):
The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
shorter if all batches finished early due to the :obj:`eos_token_id`. shorter if all batches finished early due to the :obj:`eos_token_id`.
sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequence)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): sequences_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_return_sequence)`, `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
...@@ -301,21 +299,21 @@ class BeamSampleEncoderDecoderOutput(ModelOutput): ...@@ -301,21 +299,21 @@ class BeamSampleEncoderDecoderOutput(ModelOutput):
scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``): scores (:obj:`tuple(torch.FloatTensor)` `optional`, returned when ``output_scores=True`` is passed or when ``config.output_scores=True``):
Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log Processed beam scores for each vocabulary token at each generation step. Beam scores consisting of log
softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam softmax scores for each vocabulary token and sum of log softmax of previously generated tokens in this beam
. :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape :obj:`(batch_size . :obj:`(max_length,)`-shaped tuple of :obj:`torch.FloatTensor` with each tensor of shape
* num_beams, config.vocab_size)`). :obj:`(batch_size*num_beams, config.vocab_size)`).
encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape :obj:`(batch_size, Tuple of :obj:`torch.FloatTensor` (one for each layer of the decoder) of shape :obj:`(batch_size,
num_heads, sequence_length, sequence_length)`. num_heads, sequence_length, sequence_length)`.
encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
of shape :obj:`(batch_size * num_beams, sequence_length, hidden_size)`. of shape :obj:`(batch_size*num_beams, sequence_length, hidden_size)`.
decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``): decoder_attentions (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams, num_heads, generated_length, :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams, num_heads, generated_length,
sequence_length)`. sequence_length)`.
decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``): decoder_hidden_states (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams, generated_length, hidden_size)`. :obj:`torch.FloatTensor` of shape :obj:`(batch_size*num_beams, generated_length, hidden_size)`.
""" """
sequences: torch.LongTensor = None sequences: torch.LongTensor = None
......
src/transformers/models/bert/modeling_bert.py
View file @ b8462b5b
...@@ -1227,7 +1227,7 @@ class BertLMHeadModel(BertPreTrainedModel): ...@@ -1227,7 +1227,7 @@ class BertLMHeadModel(BertPreTrainedModel):
if past is not None: if past is not None:
input_ids = input_ids[:, -1:] input_ids = input_ids[:, -1:]
return {"input_ids": input_ids, "attention_mask": attention_mask} return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
def _reorder_cache(self, past, beam_idx): def _reorder_cache(self, past, beam_idx):
reordered_past = () reordered_past = ()
......
src/transformers/models/bert_generation/modeling_bert_generation.py
View file @ b8462b5b
...@@ -570,7 +570,7 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel): ...@@ -570,7 +570,7 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
if past is not None: if past is not None:
input_ids = input_ids[:, -1:] input_ids = input_ids[:, -1:]
return {"input_ids": input_ids, "attention_mask": attention_mask} return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
def _reorder_cache(self, past, beam_idx): def _reorder_cache(self, past, beam_idx):
reordered_past = () reordered_past = ()
......
src/transformers/models/encoder_decoder/modeling_encoder_decoder.py
View file @ b8462b5b
...@@ -455,7 +455,7 @@ class EncoderDecoderModel(PreTrainedModel): ...@@ -455,7 +455,7 @@ class EncoderDecoderModel(PreTrainedModel):
"decoder_attention_mask": decoder_attention_mask, "decoder_attention_mask": decoder_attention_mask,
"decoder_input_ids": decoder_inputs["input_ids"], "decoder_input_ids": decoder_inputs["input_ids"],
"encoder_outputs": encoder_outputs, "encoder_outputs": encoder_outputs,
"past_key_values": past, "past_key_values": decoder_inputs["past_key_values"],
"use_cache": use_cache, "use_cache": use_cache,
} }
return input_dict return input_dict
......
src/transformers/models/roberta/modeling_roberta.py
View file @ b8462b5b
...@@ -962,7 +962,7 @@ class RobertaForCausalLM(RobertaPreTrainedModel): ...@@ -962,7 +962,7 @@ class RobertaForCausalLM(RobertaPreTrainedModel):
if past is not None: if past is not None:
input_ids = input_ids[:, -1:] input_ids = input_ids[:, -1:]
return {"input_ids": input_ids, "attention_mask": attention_mask} return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
def _reorder_cache(self, past, beam_idx): def _reorder_cache(self, past, beam_idx):
reordered_past = () reordered_past = ()
......
src/transformers/models/t5/modeling_t5.py
View file @ b8462b5b
...@@ -1148,6 +1148,7 @@ class T5Model(T5PreTrainedModel): ...@@ -1148,6 +1148,7 @@ class T5Model(T5PreTrainedModel):
self.shared = nn.Embedding(config.vocab_size, config.d_model) self.shared = nn.Embedding(config.vocab_size, config.d_model)
encoder_config = copy.deepcopy(config) encoder_config = copy.deepcopy(config)
encoder_config.is_decoder = False
encoder_config.use_cache = False encoder_config.use_cache = False
encoder_config.is_encoder_decoder = False encoder_config.is_encoder_decoder = False
self.encoder = T5Stack(encoder_config, self.shared) self.encoder = T5Stack(encoder_config, self.shared)
...@@ -1325,6 +1326,7 @@ class T5ForConditionalGeneration(T5PreTrainedModel): ...@@ -1325,6 +1326,7 @@ class T5ForConditionalGeneration(T5PreTrainedModel):
self.shared = nn.Embedding(config.vocab_size, config.d_model) self.shared = nn.Embedding(config.vocab_size, config.d_model)
encoder_config = copy.deepcopy(config) encoder_config = copy.deepcopy(config)
encoder_config.is_decoder = False
encoder_config.use_cache = False encoder_config.use_cache = False
encoder_config.is_encoder_decoder = False encoder_config.is_encoder_decoder = False
self.encoder = T5Stack(encoder_config, self.shared) self.encoder = T5Stack(encoder_config, self.shared)
......
templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py
View file @ b8462b5b
...@@ -1132,7 +1132,7 @@ class {{cookiecutter.camelcase_modelname}}ForCausalLM({{cookiecutter.camelcase_m ...@@ -1132,7 +1132,7 @@ class {{cookiecutter.camelcase_modelname}}ForCausalLM({{cookiecutter.camelcase_m
if past is not None: if past is not None:
input_ids = input_ids[:, -1:] input_ids = input_ids[:, -1:]
return {"input_ids": input_ids, "attention_mask": attention_mask} return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
def _reorder_cache(self, past, beam_idx): def _reorder_cache(self, past, beam_idx):
reordered_past = () reordered_past = ()
......
tests/test_generation_utils.py
View file @ b8462b5b
...@@ -522,6 +522,7 @@ class GenerationTesterMixin: ...@@ -522,6 +522,7 @@ class GenerationTesterMixin:
return return
config.use_cache = True config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval() model = model_class(config).to(torch_device).eval()
output_greedy, output_generate = self._greedy_generate( output_greedy, output_generate = self._greedy_generate(
model=model, model=model,
...@@ -730,6 +731,7 @@ class GenerationTesterMixin: ...@@ -730,6 +731,7 @@ class GenerationTesterMixin:
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length) beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
config.use_cache = True config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval() model = model_class(config).to(torch_device).eval()
output_beam, output_generate = self._beam_search_generate( output_beam, output_generate = self._beam_search_generate(
model=model, model=model,
...@@ -962,12 +964,7 @@ class GenerationTesterMixin: ...@@ -962,12 +964,7 @@ class GenerationTesterMixin:
# Attentions # Attentions
if config.is_encoder_decoder: if config.is_encoder_decoder:
# encoder # encoder
encoder_expected_shape = (batch_size, config.num_attention_heads, seq_length, seq_length) self._check_encoder_attention_for_generate(output.encoder_attentions, batch_size, config, seq_length)
self.assertIsInstance(output.encoder_attentions, tuple)
self.assertListEqual(
[layer_attentions.shape for layer_attentions in output.encoder_attentions],
[encoder_expected_shape] * len(output.encoder_attentions),
)
# decoder # decoder
self._check_attentions_for_generate( self._check_attentions_for_generate(
num_sequences_in_output, num_sequences_in_output,
...@@ -993,11 +990,8 @@ class GenerationTesterMixin: ...@@ -993,11 +990,8 @@ class GenerationTesterMixin:
# Hidden States # Hidden States
if config.is_encoder_decoder: if config.is_encoder_decoder:
# encoder # encoder
encoder_expected_shape = (batch_size, seq_length, config.hidden_size) self._check_encoder_hidden_states_for_generate(
self.assertIsInstance(output.encoder_hidden_states, tuple) output.encoder_hidden_states, batch_size, config, seq_length
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in output.encoder_hidden_states],
[encoder_expected_shape] * len(output.encoder_hidden_states),
) )
# decoder # decoder
...@@ -1052,6 +1046,14 @@ class GenerationTesterMixin: ...@@ -1052,6 +1046,14 @@ class GenerationTesterMixin:
[layer_attention.shape for layer_attention in iter_attentions], [expected_shape] * len(iter_attentions) [layer_attention.shape for layer_attention in iter_attentions], [expected_shape] * len(iter_attentions)
) )
def _check_encoder_attention_for_generate(self, attentions, batch_size, config, seq_length):
encoder_expected_shape = (batch_size, config.num_attention_heads, seq_length, seq_length)
self.assertIsInstance(attentions, tuple)
self.assertListEqual(
[layer_attentions.shape for layer_attentions in attentions],
[encoder_expected_shape] * len(attentions),
)
def _check_hidden_states_for_generate( def _check_hidden_states_for_generate(
self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1 self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
): ):
...@@ -1071,6 +1073,14 @@ class GenerationTesterMixin: ...@@ -1071,6 +1073,14 @@ class GenerationTesterMixin:
[expected_shape] * len(iter_hidden_states), [expected_shape] * len(iter_hidden_states),
) )
def _check_encoder_hidden_states_for_generate(self, hidden_states, batch_size, config, seq_length):
encoder_expected_shape = (batch_size, seq_length, config.hidden_size)
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in hidden_states],
[encoder_expected_shape] * len(hidden_states),
)
@require_torch @require_torch
class UtilsFunctionsTest(unittest.TestCase): class UtilsFunctionsTest(unittest.TestCase):
......
tests/test_modeling_led.py
View file @ b8462b5b
...@@ -327,6 +327,32 @@ class LEDModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase): ...@@ -327,6 +327,32 @@ class LEDModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
# longformer cannot keep gradients in attentions or hidden states # longformer cannot keep gradients in attentions or hidden states
return return
def _check_encoder_attention_for_generate(self, attentions, batch_size, config, seq_length):
# make sure tgt_length is padded
tgt_length = (
seq_length // config.attention_window[0] + (seq_length % config.attention_window[0] != 0)
) * config.attention_window[0]
encoder_expected_shape = (batch_size, config.num_attention_heads, tgt_length, seq_length)
self.assertIsInstance(attentions, tuple)
self.assertListEqual(
[layer_attentions.shape for layer_attentions in attentions],
[encoder_expected_shape] * len(attentions),
)
def _check_encoder_hidden_states_for_generate(self, hidden_states, batch_size, config, seq_length):
# make sure seq_length is padded
seq_length = (
seq_length // config.attention_window[0] + (seq_length % config.attention_window[0] != 0)
) * config.attention_window[0]
encoder_expected_shape = (batch_size, seq_length, config.hidden_size)
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in hidden_states],
[encoder_expected_shape] * len(hidden_states),
)
def test_attention_outputs(self): def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True config.return_dict = True
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment