bourdoiscatie
commited on
Commit
•
ee94040
1
Parent(s):
a4c7177
Update custom_heads_flash_t5.py
Browse files- custom_heads_flash_t5.py +1 -92
custom_heads_flash_t5.py
CHANGED
@@ -257,6 +257,7 @@ class FlashT5ForQuestionAnswering(FlashT5PreTrainedModel):
|
|
257 |
Labels for position (index) of the end of the labelled span for computing the token classification loss.
|
258 |
Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence
|
259 |
are not taken into account for computing the loss.
|
|
|
260 |
Returns:
|
261 |
"""
|
262 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
@@ -306,95 +307,3 @@ class FlashT5ForQuestionAnswering(FlashT5PreTrainedModel):
|
|
306 |
hidden_states=encoder_outputs.hidden_states,
|
307 |
attentions=encoder_outputs.attentions,
|
308 |
)
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
class FlashT5ForQuestionAnswering(FlashT5PreTrainedModel):
|
313 |
-
_keys_to_ignore_on_load_missing = [r"encoder.embed_tokens.weight"]
|
314 |
-
|
315 |
-
def __init__(self, config: FlashT5Config):
|
316 |
-
super().__init__(config)
|
317 |
-
self.shared = nn.Embedding(config.vocab_size, config.d_model)
|
318 |
-
|
319 |
-
encoder_config = copy.deepcopy(config)
|
320 |
-
encoder_config.is_decoder = False
|
321 |
-
encoder_config.is_encoder_decoder = False
|
322 |
-
self.encoder = FlashT5Stack(encoder_config, self.shared)
|
323 |
-
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
|
324 |
-
|
325 |
-
# Initialize weights and apply final processing
|
326 |
-
self.post_init()
|
327 |
-
|
328 |
-
self.qa_outputs.weight.data.normal_(mean=0.0, std=config.initializer_factor * 1.0)
|
329 |
-
self.qa_outputs.bias.data.zero_()
|
330 |
-
|
331 |
-
self.model_parallel = False
|
332 |
-
|
333 |
-
def forward(
|
334 |
-
self,
|
335 |
-
input_ids: Optional[torch.LongTensor] = None,
|
336 |
-
attention_mask: Optional[torch.FloatTensor] = None,
|
337 |
-
head_mask: Optional[torch.FloatTensor] = None,
|
338 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
339 |
-
start_positions: Optional[torch.LongTensor] = None,
|
340 |
-
end_positions: Optional[torch.LongTensor] = None,
|
341 |
-
output_attentions: Optional[bool] = None,
|
342 |
-
output_hidden_states: Optional[bool] = None,
|
343 |
-
return_dict: Optional[bool] = None,
|
344 |
-
) -> Union[Tuple, QuestionAnsweringModelOutput]:
|
345 |
-
r"""
|
346 |
-
Returns:
|
347 |
-
Example:
|
348 |
-
```python
|
349 |
-
>>> from transformers import AutoTokenizer, MTxEncoderForQuestionAnswering
|
350 |
-
>>> tokenizer = AutoTokenizer.from_pretrained("MTx-small")
|
351 |
-
>>> model = MTxEncoderForQuestionAnswering.from_pretrained("MTx-small")
|
352 |
-
>>> input_ids = tokenizer(
|
353 |
-
... "Studies have been shown that owning a dog is good for you", return_tensors="pt"
|
354 |
-
... ).input_ids # Batch size 1
|
355 |
-
>>> outputs = model(input_ids=input_ids)
|
356 |
-
>>> start_logits = outputs.start_logits
|
357 |
-
>>> end_logits = outputs.end_logits
|
358 |
-
```"""
|
359 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
360 |
-
|
361 |
-
outputs = self.encoder(
|
362 |
-
input_ids,
|
363 |
-
attention_mask=attention_mask,
|
364 |
-
inputs_embeds=inputs_embeds,
|
365 |
-
)
|
366 |
-
sequence_output = outputs[0]
|
367 |
-
|
368 |
-
logits = self.qa_outputs(sequence_output)
|
369 |
-
start_logits, end_logits = logits.split(1, dim=-1)
|
370 |
-
start_logits = start_logits.squeeze(-1).contiguous()
|
371 |
-
end_logits = end_logits.squeeze(-1).contiguous()
|
372 |
-
|
373 |
-
total_loss = None
|
374 |
-
if start_positions is not None and end_positions is not None:
|
375 |
-
# If we are on multi-GPU, split add a dimension
|
376 |
-
if len(start_positions.size()) > 1:
|
377 |
-
start_positions = start_positions.squeeze(-1).to(start_logits.device)
|
378 |
-
if len(end_positions.size()) > 1:
|
379 |
-
end_positions = end_positions.squeeze(-1).to(end_logits.device)
|
380 |
-
# sometimes the start/end positions are outside our model inputs, we ignore these terms
|
381 |
-
ignored_index = start_logits.size(1)
|
382 |
-
start_positions = start_positions.clamp(0, ignored_index)
|
383 |
-
end_positions = end_positions.clamp(0, ignored_index)
|
384 |
-
|
385 |
-
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
|
386 |
-
start_loss = loss_fct(start_logits, start_positions)
|
387 |
-
end_loss = loss_fct(end_logits, end_positions)
|
388 |
-
total_loss = (start_loss + end_loss) / 2
|
389 |
-
|
390 |
-
if not return_dict:
|
391 |
-
output = (start_logits, end_logits) + outputs[1:]
|
392 |
-
return ((total_loss,) + output) if total_loss is not None else output
|
393 |
-
|
394 |
-
return QuestionAnsweringModelOutput(
|
395 |
-
loss=total_loss,
|
396 |
-
start_logits=start_logits,
|
397 |
-
end_logits=end_logits,
|
398 |
-
hidden_states=outputs.hidden_states,
|
399 |
-
attentions=outputs.attentions,
|
400 |
-
)
|
|
|
257 |
Labels for position (index) of the end of the labelled span for computing the token classification loss.
|
258 |
Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence
|
259 |
are not taken into account for computing the loss.
|
260 |
+
|
261 |
Returns:
|
262 |
"""
|
263 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
307 |
hidden_states=encoder_outputs.hidden_states,
|
308 |
attentions=encoder_outputs.attentions,
|
309 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|