You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

418 lines
18 KiB

# Copyright 2022, UCLA NLP, The Facebook AI Research Team and The HuggingFace Inc. team. All rights reserved.
#
# 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.
"""PyTorch PLBART model."""
import math
import torch
from torch import nn
from torch.nn import CrossEntropyLoss
from ... import initialization as init
from ...cache_utils import Cache
from ...generation import GenerationMixin
from ...modeling_outputs import (
BaseModelOutput,
Seq2SeqLMOutput,
Seq2SeqModelOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import auto_docstring
from ..bart.modeling_bart import (
BartClassificationHead,
BartDecoder,
BartEncoder,
BartForCausalLM,
BartScaledWordEmbedding,
)
from ..bigbird_pegasus.modeling_bigbird_pegasus import BigBirdPegasusForSequenceClassification
from ..mbart.modeling_mbart import shift_tokens_right
from .configuration_plbart import PLBartConfig
class PLBartScaledWordEmbedding(BartScaledWordEmbedding):
pass
@auto_docstring
class PLBartPreTrainedModel(PreTrainedModel):
config: PLBartConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["PLBartDecoderLayer", "PLBartEncoderLayer"]
_supports_flash_attn = True
_supports_sdpa = True
_supports_flex_attn = True
def _init_weights(self, module):
super()._init_weights(module)
if isinstance(module, PLBartForConditionalGeneration):
init.zeros_(module.final_logits_bias)
class PLBartEncoder(BartEncoder):
pass
class PLBartDecoder(BartDecoder):
pass
@auto_docstring
class PLBartModel(PLBartPreTrainedModel):
_tied_weights_keys = {
"encoder.embed_tokens.weight": "shared.weight",
"decoder.embed_tokens.weight": "shared.weight",
}
def __init__(self, config: PLBartConfig):
super().__init__(config)
padding_idx, vocab_size = config.pad_token_id, config.vocab_size
embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
self.shared = PLBartScaledWordEmbedding(vocab_size, config.d_model, padding_idx, embed_scale=embed_scale)
self.encoder = PLBartEncoder(config)
self.decoder = PLBartDecoder(config)
self.post_init()
def get_input_embeddings(self):
return self.shared
def set_input_embeddings(self, value):
self.shared = value
self.encoder.embed_tokens = self.shared
self.decoder.embed_tokens = self.shared
@auto_docstring
def forward(
self,
input_ids: torch.LongTensor | None = None,
attention_mask: torch.LongTensor | None = None,
decoder_input_ids: torch.LongTensor | None = None,
decoder_attention_mask: torch.Tensor | None = None,
encoder_outputs: list[torch.FloatTensor] | None = None,
past_key_values: Cache | None = None,
inputs_embeds: torch.FloatTensor | None = None,
decoder_inputs_embeds: torch.FloatTensor | None = None,
use_cache: bool | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
cache_position: torch.LongTensor | None = None,
**kwargs,
) -> tuple[torch.Tensor] | Seq2SeqModelOutput:
r"""
decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`] or [`PLBartMultiTokenizer`] depending on the checkpoint.
See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are decoder input IDs?](../glossary#decoder-input-ids)
PLBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
varies according to source and target language, *e.g.* 50003 for *en_XX*, and 50001 for *java*. If
`past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
For translation and summarization training, `decoder_input_ids` should be provided. If no
`decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
for denoising pre-training following the paper.
decoder_attention_mask (:
obj:*torch.LongTensor* of shape `(batch_size, target_sequence_length)`, *optional*):
Default behavior:
generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# different to other models, PLBart automatically creates decoder_input_ids from
# input_ids if no decoder_input_ids are provided
if decoder_input_ids is None and decoder_inputs_embeds is None:
decoder_input_ids = shift_tokens_right(input_ids, self.config.pad_token_id)
if encoder_outputs is None:
encoder_outputs = self.encoder(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
# If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
encoder_outputs = BaseModelOutput(
last_hidden_state=encoder_outputs[0],
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
)
# decoder outputs consists of (dec_features, past_key_values, dec_hidden, dec_attn)
decoder_outputs = self.decoder(
input_ids=decoder_input_ids,
attention_mask=decoder_attention_mask,
encoder_hidden_states=encoder_outputs[0],
encoder_attention_mask=attention_mask,
past_key_values=past_key_values,
inputs_embeds=decoder_inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
if not return_dict:
return decoder_outputs + encoder_outputs
return Seq2SeqModelOutput(
last_hidden_state=decoder_outputs.last_hidden_state,
past_key_values=decoder_outputs.past_key_values,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
)
@auto_docstring(
custom_intro="""
The PLBART Model with a language modeling head. Can be used for code-to-text, text-to-code and code-to-code.
"""
)
class PLBartForConditionalGeneration(PLBartPreTrainedModel, GenerationMixin):
base_model_prefix = "model"
_keys_to_ignore_on_load_missing = ["final_logits_bias"]
_tied_weights_keys = {
"lm_head.weight": "model.shared.weight",
}
def __init__(self, config: PLBartConfig):
super().__init__(config)
self.model = PLBartModel(config)
self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)
self.post_init()
def resize_token_embeddings(
self, new_num_tokens: int, pad_to_multiple_of: int | None = None, mean_resizing: bool = True
) -> nn.Embedding:
new_embeddings = super().resize_token_embeddings(new_num_tokens, pad_to_multiple_of, mean_resizing)
self._resize_final_logits_bias(new_embeddings.weight.shape[0])
return new_embeddings
def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
old_num_tokens = self.final_logits_bias.shape[-1]
if new_num_tokens <= old_num_tokens:
new_bias = self.final_logits_bias[:, :new_num_tokens]
else:
extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
self.register_buffer("final_logits_bias", new_bias)
@auto_docstring
def forward(
self,
input_ids: torch.LongTensor | None = None,
attention_mask: torch.LongTensor | None = None,
decoder_input_ids: torch.LongTensor | None = None,
decoder_attention_mask: torch.Tensor | None = None,
encoder_outputs: list[torch.FloatTensor] | None = None,
past_key_values: Cache | None = None,
inputs_embeds: torch.FloatTensor | None = None,
decoder_inputs_embeds: torch.FloatTensor | None = None,
labels: torch.Tensor | None = None,
use_cache: bool | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
cache_position: torch.LongTensor | None = None,
**kwargs,
) -> tuple[torch.Tensor] | Seq2SeqLMOutput:
r"""
decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`] or [`PLBartMultiTokenizer`] depending on the checkpoint.
See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are decoder input IDs?](../glossary#decoder-input-ids)
PLBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
varies according to source and target language, *e.g.* 50003 for *en_XX*, and 50001 for *java*. If
`past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
For translation and summarization training, `decoder_input_ids` should be provided. If no
`decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
for denoising pre-training following the paper.
decoder_attention_mask (:
obj:*torch.LongTensor* of shape `(batch_size, target_sequence_length)`, *optional*):
Default behavior:
generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default.
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Example Mask-filling:
```python
>>> from transformers import AutoTokenizer, PLBartForConditionalGeneration
>>> model = PLBartForConditionalGeneration.from_pretrained("uclanlp/plbart-base")
>>> tokenizer = AutoTokenizer.from_pretrained("uclanlp/plbart-base")
>>> # en_XX is the language symbol id <LID> for English
>>> TXT = "<s> Is 0 the <mask> Fibonacci number ? </s> en_XX"
>>> input_ids = tokenizer([TXT], add_special_tokens=False, return_tensors="pt").input_ids
>>> logits = model(input_ids).logits
>>> masked_index = (input_ids[0] == tokenizer.mask_token_id).nonzero().item()
>>> probs = logits[0, masked_index].softmax(dim=0)
>>> values, predictions = probs.topk(5)
>>> tokenizer.decode(predictions).split()
['first', 'same', 'highest', 'result', 'number']
```
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if labels is not None:
if decoder_input_ids is None and decoder_inputs_embeds is None:
decoder_input_ids = shift_tokens_right(labels, self.config.pad_token_id)
outputs = self.model(
input_ids,
attention_mask=attention_mask,
decoder_input_ids=decoder_input_ids,
encoder_outputs=encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
decoder_inputs_embeds=decoder_inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
lm_logits = self.lm_head(outputs[0])
lm_logits = lm_logits + self.final_logits_bias.to(lm_logits.device)
masked_lm_loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
if not return_dict:
output = (lm_logits,) + outputs[1:]
return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
return Seq2SeqLMOutput(
loss=masked_lm_loss,
logits=lm_logits,
past_key_values=outputs.past_key_values,
decoder_hidden_states=outputs.decoder_hidden_states,
decoder_attentions=outputs.decoder_attentions,
cross_attentions=outputs.cross_attentions,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
encoder_hidden_states=outputs.encoder_hidden_states,
encoder_attentions=outputs.encoder_attentions,
)
def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
return shift_tokens_right(labels, self.config.pad_token_id)
class PLBartClassificationHead(BartClassificationHead):
pass
class PLBartForSequenceClassification(BigBirdPegasusForSequenceClassification):
def forward(**super_kwargs):
r"""
decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`] or [`PLBartMultiTokenizer`] depending on the checkpoint.
See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
[What are decoder input IDs?](../glossary#decoder-input-ids)
PLBart uses a specific language id token as the starting token for `decoder_input_ids` generation that
varies according to source and target language, *e.g.* 50003 for *en_XX*, and 50001 for *java*. If
`past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
For translation and summarization training, `decoder_input_ids` should be provided. If no
`decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right
for denoising pre-training following the paper.
decoder_attention_mask (:
obj:*torch.LongTensor* of shape `(batch_size, target_sequence_length)`, *optional*):
Default behavior:
generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default.
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
super().forward(**super_kwargs)
class PLBartForCausalLM(BartForCausalLM):
@auto_docstring
def forward(**super_kwargs):
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Example:
```python
>>> from transformers import AutoTokenizer, PLBartForCausalLM
>>> tokenizer = AutoTokenizer.from_pretrained("uclanlp/plbart-base")
>>> model = PLBartForCausalLM.from_pretrained("uclanlp/plbart-base")
>>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
>>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> expected_shape = [1, inputs.input_ids.shape[-1], model.config.vocab_size]
>>> list(logits.shape) == expected_shape
True
```"""
super().forward(**super_kwargs)
__all__ = [
"PLBartForCausalLM",
"PLBartForConditionalGeneration",
"PLBartForSequenceClassification",
"PLBartModel",
"PLBartPreTrainedModel",
]