Spaces:
Running
Running
wissamantoun
commited on
Commit
•
3fed02b
1
Parent(s):
0266e41
Update backend/modeling_gpt2.py
Browse files- backend/modeling_gpt2.py +650 -436
backend/modeling_gpt2.py
CHANGED
@@ -13,51 +13,54 @@
|
|
13 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
# See the License for the specific language governing permissions and
|
15 |
# limitations under the License.
|
|
|
16 |
|
17 |
-
|
18 |
-
PyTorch OpenAI GPT-2 model.
|
19 |
-
Adapted from https://github.com/huggingface/transformers/blob/v4.0.1/src/transformers/models/gpt2/modeling_gpt2.py
|
20 |
-
and https://github.com/ghosthamlet/gpt2-ml-torch/blob/master/gpt2_ml_torch/modeling_gpt2.py
|
21 |
-
"""
|
22 |
-
|
23 |
-
|
24 |
-
import logging
|
25 |
import os
|
26 |
from dataclasses import dataclass
|
27 |
-
from typing import
|
28 |
|
29 |
import torch
|
30 |
-
import torch.
|
31 |
-
from
|
32 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
from transformers.activations import ACT2FN
|
34 |
-
from transformers.file_utils import (
|
35 |
-
ModelOutput,
|
36 |
-
add_code_sample_docstrings,
|
37 |
-
add_start_docstrings,
|
38 |
-
add_start_docstrings_to_model_forward,
|
39 |
-
replace_return_docstrings,
|
40 |
-
)
|
41 |
from transformers.modeling_outputs import (
|
42 |
BaseModelOutputWithPastAndCrossAttentions,
|
43 |
CausalLMOutputWithCrossAttentions,
|
44 |
SequenceClassifierOutputWithPast,
|
45 |
TokenClassifierOutput,
|
46 |
)
|
47 |
-
from transformers.modeling_utils import
|
|
|
48 |
Conv1D,
|
49 |
-
PreTrainedModel,
|
50 |
-
SequenceSummary,
|
51 |
find_pruneable_heads_and_indices,
|
52 |
prune_conv1d_layer,
|
53 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
|
|
|
55 |
|
56 |
-
# THe Difference from Transformers is code under _USE_GROVER
|
57 |
-
_USE_GROVER = True
|
58 |
|
59 |
-
logger = logging.
|
60 |
|
|
|
61 |
_CONFIG_FOR_DOC = "GPT2Config"
|
62 |
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
|
63 |
|
@@ -70,11 +73,6 @@ GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
|
70 |
# See all GPT-2 models at https://huggingface.co/models?filter=gpt2
|
71 |
]
|
72 |
|
73 |
-
logger.setLevel(logging.INFO)
|
74 |
-
console = logging.StreamHandler()
|
75 |
-
console.setLevel(logging.INFO)
|
76 |
-
logger.addHandler(console)
|
77 |
-
|
78 |
_GPT2_ML_TF_TO_TORCH = {
|
79 |
"LayerNorm_embed_norm": "emb_norm",
|
80 |
"pos_embed": "wpe.weight",
|
@@ -126,7 +124,6 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
|
|
126 |
"""Load tf checkpoints in a pytorch model"""
|
127 |
try:
|
128 |
import re
|
129 |
-
|
130 |
import tensorflow as tf
|
131 |
except ImportError:
|
132 |
logger.error(
|
@@ -206,11 +203,7 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
|
|
206 |
d = torch.from_numpy(array)
|
207 |
is_bias = len(shape) == 1
|
208 |
end = int(shape[0 if is_bias else 1] / 3)
|
209 |
-
m = dict(
|
210 |
-
query_layer=0,
|
211 |
-
key_layer=end,
|
212 |
-
value_layer=end * 2,
|
213 |
-
)
|
214 |
start = m[attn_layer]
|
215 |
end = start + end
|
216 |
if is_bias:
|
@@ -232,39 +225,54 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
|
|
232 |
return model
|
233 |
|
234 |
|
235 |
-
class
|
236 |
-
def __init__(self,
|
237 |
super().__init__()
|
238 |
|
239 |
-
|
240 |
-
# [switch nx => n_state from Block to Attention to keep identical to TF implem]
|
241 |
-
assert n_state % config.n_head == 0
|
242 |
self.register_buffer(
|
243 |
"bias",
|
244 |
-
torch.tril(
|
245 |
-
|
246 |
-
),
|
247 |
)
|
248 |
self.register_buffer("masked_bias", torch.tensor(-1e4))
|
249 |
-
|
250 |
-
self.
|
251 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
self.is_cross_attention = is_cross_attention
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
if self.is_cross_attention:
|
254 |
-
self.c_attn = Conv1D(2 *
|
255 |
-
self.q_attn = Conv1D(
|
256 |
else:
|
257 |
-
self.c_attn = Conv1D(3 *
|
258 |
-
self.c_proj = Conv1D(
|
|
|
259 |
self.attn_dropout = nn.Dropout(config.attn_pdrop)
|
260 |
self.resid_dropout = nn.Dropout(config.resid_pdrop)
|
|
|
261 |
self.pruned_heads = set()
|
262 |
|
263 |
def prune_heads(self, heads):
|
264 |
if len(heads) == 0:
|
265 |
return
|
266 |
heads, index = find_pruneable_heads_and_indices(
|
267 |
-
heads, self.
|
268 |
)
|
269 |
index_attn = torch.cat(
|
270 |
[index, index + self.split_size, index + (2 * self.split_size)]
|
@@ -275,67 +283,163 @@ class Attention(nn.Module):
|
|
275 |
self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
|
276 |
|
277 |
# Update hyper params
|
278 |
-
self.split_size = (self.split_size // self.
|
279 |
-
|
|
|
|
|
280 |
self.pruned_heads = self.pruned_heads.union(heads)
|
281 |
|
282 |
-
def _attn(
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
|
|
|
|
289 |
|
290 |
if not self.is_cross_attention:
|
291 |
# if only "normal" attention layer implements causal mask
|
292 |
-
|
293 |
-
|
|
|
|
|
|
|
|
|
|
|
294 |
|
295 |
if attention_mask is not None:
|
296 |
# Apply the attention mask
|
297 |
-
|
298 |
|
299 |
-
|
300 |
-
|
|
|
|
|
|
|
301 |
|
302 |
# Mask heads if we want to
|
303 |
if head_mask is not None:
|
304 |
-
|
305 |
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
else:
|
322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
323 |
|
324 |
def forward(
|
325 |
self,
|
326 |
-
hidden_states,
|
327 |
-
layer_past=None,
|
328 |
-
attention_mask=None,
|
329 |
-
head_mask=None,
|
330 |
-
encoder_hidden_states=None,
|
331 |
-
encoder_attention_mask=None,
|
332 |
-
use_cache=False,
|
333 |
-
output_attentions=False,
|
334 |
-
):
|
335 |
if encoder_hidden_states is not None:
|
336 |
-
|
337 |
-
|
338 |
-
|
|
|
|
|
|
|
339 |
query = self.q_attn(hidden_states)
|
340 |
key, value = self.c_attn(encoder_hidden_states).split(
|
341 |
self.split_size, dim=2
|
@@ -344,80 +448,97 @@ class Attention(nn.Module):
|
|
344 |
else:
|
345 |
query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
|
346 |
|
347 |
-
query = self.
|
348 |
-
key = self.
|
349 |
-
value = self.
|
|
|
350 |
if layer_past is not None:
|
351 |
-
past_key, past_value =
|
352 |
-
|
353 |
-
layer_past[1],
|
354 |
-
) # transpose back cf below
|
355 |
-
key = torch.cat((past_key, key), dim=-1)
|
356 |
value = torch.cat((past_value, value), dim=-2)
|
357 |
|
358 |
if use_cache is True:
|
359 |
-
present =
|
360 |
-
(key.transpose(-2, -1), value)
|
361 |
-
) # transpose to have same shapes for stacking
|
362 |
else:
|
363 |
-
present =
|
364 |
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
369 |
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
|
374 |
-
outputs = [a, present] + attn_outputs[1:]
|
375 |
return outputs # a, present, (attentions)
|
376 |
|
377 |
|
378 |
-
class
|
379 |
-
def __init__(self,
|
380 |
super().__init__()
|
381 |
-
|
382 |
-
self.c_fc = Conv1D(
|
383 |
-
self.c_proj = Conv1D(
|
384 |
self.act = ACT2FN[config.activation_function]
|
385 |
self.dropout = nn.Dropout(config.resid_pdrop)
|
386 |
|
387 |
-
def forward(
|
388 |
-
|
389 |
-
|
390 |
-
|
|
|
|
|
|
|
|
|
391 |
|
392 |
|
393 |
-
class
|
394 |
-
def __init__(self,
|
395 |
super().__init__()
|
396 |
-
hidden_size = config.
|
397 |
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
|
|
|
398 |
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
399 |
-
self.attn =
|
400 |
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
|
|
401 |
if config.add_cross_attention:
|
402 |
-
self.crossattention =
|
403 |
-
|
404 |
)
|
405 |
self.ln_cross_attn = nn.LayerNorm(
|
406 |
hidden_size, eps=config.layer_norm_epsilon
|
407 |
)
|
408 |
-
|
|
|
409 |
|
410 |
def forward(
|
411 |
self,
|
412 |
-
hidden_states,
|
413 |
-
layer_past=None,
|
414 |
-
attention_mask=None,
|
415 |
-
head_mask=None,
|
416 |
-
encoder_hidden_states=None,
|
417 |
-
encoder_attention_mask=None,
|
418 |
-
use_cache=False,
|
419 |
-
output_attentions=False,
|
420 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
attn_outputs = self.attn(
|
422 |
hidden_states,
|
423 |
layer_past=layer_past,
|
@@ -433,11 +554,16 @@ class Block(nn.Module):
|
|
433 |
|
434 |
if encoder_hidden_states is not None:
|
435 |
# add one self-attention block for cross-attention
|
436 |
-
|
437 |
-
|
438 |
-
|
|
|
|
|
|
|
|
|
|
|
439 |
cross_attn_outputs = self.crossattention(
|
440 |
-
|
441 |
attention_mask=attention_mask,
|
442 |
head_mask=head_mask,
|
443 |
encoder_hidden_states=encoder_hidden_states,
|
@@ -446,18 +572,24 @@ class Block(nn.Module):
|
|
446 |
)
|
447 |
attn_output = cross_attn_outputs[0]
|
448 |
# residual connection
|
449 |
-
hidden_states =
|
450 |
outputs = (
|
451 |
outputs + cross_attn_outputs[2:]
|
452 |
) # add cross attentions if we output attention weights
|
453 |
|
454 |
-
|
|
|
|
|
455 |
# residual connection
|
456 |
-
hidden_states =
|
457 |
|
458 |
-
hidden_states = self.ln_2(hidden_states)
|
|
|
|
|
|
|
|
|
|
|
459 |
|
460 |
-
outputs = [hidden_states] + outputs
|
461 |
return outputs # hidden_states, present, (attentions, cross_attentions)
|
462 |
|
463 |
|
@@ -471,22 +603,48 @@ class GPT2PreTrainedModel(PreTrainedModel):
|
|
471 |
load_tf_weights = load_tf_weights_in_gpt2
|
472 |
base_model_prefix = "transformer"
|
473 |
is_parallelizable = True
|
|
|
474 |
|
475 |
def __init__(self, *inputs, **kwargs):
|
476 |
super().__init__(*inputs, **kwargs)
|
477 |
|
478 |
def _init_weights(self, module):
|
479 |
"""Initialize the weights."""
|
480 |
-
if isinstance(module, (nn.Linear,
|
481 |
# Slightly different from the TF version which uses truncated_normal for initialization
|
482 |
# cf https://github.com/pytorch/pytorch/pull/5617
|
483 |
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
484 |
-
if
|
485 |
module.bias.data.zero_()
|
|
|
|
|
|
|
|
|
486 |
elif isinstance(module, nn.LayerNorm):
|
487 |
module.bias.data.zero_()
|
488 |
module.weight.data.fill_(1.0)
|
489 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
490 |
|
491 |
@dataclass
|
492 |
class GPT2DoubleHeadsModelOutput(ModelOutput):
|
@@ -494,125 +652,125 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
|
|
494 |
Base class for outputs of models predicting if two sentences are consecutive or not.
|
495 |
|
496 |
Args:
|
497 |
-
loss (
|
498 |
Language modeling loss.
|
499 |
-
mc_loss (
|
500 |
Multiple choice classification loss.
|
501 |
-
logits (
|
502 |
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
503 |
-
mc_logits (
|
504 |
Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
|
505 |
-
past_key_values (
|
506 |
-
|
507 |
-
|
508 |
|
509 |
Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
|
510 |
-
|
511 |
-
hidden_states (
|
512 |
-
Tuple of
|
513 |
-
|
514 |
|
515 |
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
516 |
-
attentions (
|
517 |
-
Tuple of
|
518 |
-
sequence_length
|
519 |
|
520 |
-
|
521 |
-
heads.
|
522 |
"""
|
523 |
|
524 |
loss: Optional[torch.FloatTensor] = None
|
525 |
mc_loss: Optional[torch.FloatTensor] = None
|
526 |
logits: torch.FloatTensor = None
|
527 |
mc_logits: torch.FloatTensor = None
|
528 |
-
past_key_values: Optional[
|
529 |
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
530 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
531 |
|
532 |
|
533 |
GPT2_START_DOCSTRING = r"""
|
534 |
|
535 |
-
This model inherits from
|
536 |
-
|
537 |
-
|
538 |
|
539 |
-
This model is also a PyTorch
|
540 |
-
|
541 |
-
|
542 |
|
543 |
Parameters:
|
544 |
-
config (
|
545 |
Initializing with a config file does not load the weights associated with the model, only the
|
546 |
-
configuration. Check out the
|
547 |
-
weights.
|
548 |
"""
|
549 |
|
550 |
GPT2_INPUTS_DOCSTRING = r"""
|
551 |
Args:
|
552 |
-
input_ids (
|
553 |
-
|
554 |
-
|
555 |
sequence tokens in the vocabulary.
|
556 |
|
557 |
-
If
|
558 |
-
|
559 |
|
560 |
-
Indices can be obtained using
|
561 |
-
|
562 |
-
details.
|
563 |
|
564 |
-
|
565 |
-
past_key_values (
|
566 |
Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
|
567 |
-
|
568 |
-
|
569 |
-
|
570 |
-
|
571 |
-
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
|
572 |
|
573 |
- 1 for tokens that are **not masked**,
|
574 |
- 0 for tokens that are **masked**.
|
575 |
|
576 |
-
`
|
577 |
-
|
578 |
-
|
579 |
-
1]``:
|
580 |
|
581 |
-
|
582 |
-
|
|
|
|
|
583 |
|
584 |
-
|
585 |
-
|
586 |
-
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
|
587 |
-
config.max_position_embeddings - 1]``.
|
588 |
|
589 |
-
|
590 |
-
|
591 |
-
|
|
|
|
|
|
|
|
|
|
|
592 |
|
593 |
- 1 indicates the head is **not masked**,
|
594 |
- 0 indicates the head is **masked**.
|
595 |
|
596 |
-
inputs_embeds (
|
597 |
-
Optionally, instead of passing
|
598 |
-
|
599 |
-
|
600 |
-
|
601 |
-
If
|
602 |
-
|
603 |
-
use_cache (
|
604 |
-
If set to
|
605 |
-
|
606 |
-
output_attentions (
|
607 |
-
Whether or not to return the attentions tensors of all attention layers. See
|
608 |
tensors for more detail.
|
609 |
-
output_hidden_states (
|
610 |
-
Whether or not to return the hidden states of all layers. See
|
611 |
more detail.
|
612 |
-
return_dict (
|
613 |
-
Whether or not to return a
|
614 |
"""
|
615 |
-
|
616 |
PARALLELIZE_DOCSTRING = r"""
|
617 |
This is an experimental feature and is a subject to change at a moment's notice.
|
618 |
|
@@ -620,7 +778,7 @@ PARALLELIZE_DOCSTRING = r"""
|
|
620 |
it will evenly distribute blocks across all devices.
|
621 |
|
622 |
Args:
|
623 |
-
device_map (
|
624 |
A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
|
625 |
automatically mapped to the first device (for esoteric reasons). That means that the first device should
|
626 |
have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
|
@@ -631,31 +789,37 @@ PARALLELIZE_DOCSTRING = r"""
|
|
631 |
- gpt2-large: 36
|
632 |
- gpt2-xl: 48
|
633 |
|
634 |
-
Example
|
635 |
-
|
636 |
-
|
637 |
-
|
638 |
-
|
639 |
-
|
640 |
-
|
641 |
-
|
642 |
-
|
643 |
-
|
|
|
|
|
|
|
644 |
"""
|
645 |
DEPARALLELIZE_DOCSTRING = r"""
|
646 |
Moves the model to cpu from a model parallel state.
|
647 |
|
648 |
-
Example
|
649 |
-
|
650 |
-
|
651 |
-
|
652 |
-
|
653 |
-
|
654 |
-
|
655 |
-
|
656 |
-
|
657 |
-
|
658 |
-
|
|
|
|
|
|
|
659 |
"""
|
660 |
|
661 |
|
@@ -664,26 +828,32 @@ DEPARALLELIZE_DOCSTRING = r"""
|
|
664 |
GPT2_START_DOCSTRING,
|
665 |
)
|
666 |
class GPT2Model(GPT2PreTrainedModel):
|
|
|
|
|
667 |
def __init__(self, config):
|
668 |
super().__init__(config)
|
669 |
|
670 |
-
self.
|
671 |
-
self.wpe = nn.Embedding(config.n_positions, config.n_embd)
|
672 |
-
if _USE_GROVER:
|
673 |
-
self.emb_norm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
674 |
|
|
|
|
|
|
|
|
|
|
|
675 |
self.drop = nn.Dropout(config.embd_pdrop)
|
676 |
self.h = nn.ModuleList(
|
677 |
-
[
|
678 |
)
|
679 |
-
|
680 |
-
|
681 |
-
|
682 |
-
self.init_weights()
|
683 |
|
684 |
# Model parallel
|
685 |
self.model_parallel = False
|
686 |
self.device_map = None
|
|
|
|
|
|
|
|
|
687 |
|
688 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
689 |
def parallelize(self, device_map=None):
|
@@ -703,13 +873,22 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
703 |
self.last_device = "cuda:" + str(max(self.device_map.keys()))
|
704 |
self.wte = self.wte.to(self.first_device)
|
705 |
self.wpe = self.wpe.to(self.first_device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
706 |
# Load onto devices
|
707 |
for k, v in self.device_map.items():
|
708 |
for block in v:
|
709 |
cuda_device = "cuda:" + str(k)
|
710 |
self.h[block] = self.h[block].to(cuda_device)
|
711 |
# ln_f to last
|
712 |
-
|
|
|
713 |
|
714 |
@add_start_docstrings(DEPARALLELIZE_DOCSTRING)
|
715 |
def deparallelize(self):
|
@@ -719,9 +898,12 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
719 |
self.last_device = "cpu"
|
720 |
self.wte = self.wte.to("cpu")
|
721 |
self.wpe = self.wpe.to("cpu")
|
|
|
|
|
722 |
for index in range(len(self.h)):
|
723 |
self.h[index] = self.h[index].to("cpu")
|
724 |
-
|
|
|
725 |
torch.cuda.empty_cache()
|
726 |
|
727 |
def get_input_embeddings(self):
|
@@ -739,27 +921,27 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
739 |
|
740 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
741 |
@add_code_sample_docstrings(
|
742 |
-
|
743 |
-
checkpoint=
|
744 |
output_type=BaseModelOutputWithPastAndCrossAttentions,
|
745 |
config_class=_CONFIG_FOR_DOC,
|
746 |
)
|
747 |
def forward(
|
748 |
self,
|
749 |
-
input_ids=None,
|
750 |
-
past_key_values=None,
|
751 |
-
attention_mask=None,
|
752 |
-
token_type_ids=None,
|
753 |
-
position_ids=None,
|
754 |
-
head_mask=None,
|
755 |
-
inputs_embeds=None,
|
756 |
-
encoder_hidden_states=None,
|
757 |
-
encoder_attention_mask=None,
|
758 |
-
use_cache=None,
|
759 |
-
output_attentions=None,
|
760 |
-
output_hidden_states=None,
|
761 |
-
return_dict=None,
|
762 |
-
):
|
763 |
output_attentions = (
|
764 |
output_attentions
|
765 |
if output_attentions is not None
|
@@ -789,6 +971,8 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
789 |
else:
|
790 |
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
791 |
|
|
|
|
|
792 |
if token_type_ids is not None:
|
793 |
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
794 |
if position_ids is not None:
|
@@ -796,11 +980,10 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
796 |
|
797 |
if past_key_values is None:
|
798 |
past_length = 0
|
799 |
-
past_key_values = [None] * len(self.h)
|
800 |
else:
|
801 |
past_length = past_key_values[0][0].size(-2)
|
802 |
if position_ids is None:
|
803 |
-
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
804 |
position_ids = torch.arange(
|
805 |
past_length,
|
806 |
input_shape[-1] + past_length,
|
@@ -809,7 +992,7 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
809 |
)
|
810 |
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
|
811 |
|
812 |
-
#
|
813 |
if attention_mask is not None:
|
814 |
if batch_size <= 0:
|
815 |
raise ValueError("batch_size has to be defined and > 0")
|
@@ -829,7 +1012,7 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
829 |
attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
|
830 |
attention_mask = (1.0 - attention_mask) * -10000.0
|
831 |
|
832 |
-
# If a 2D
|
833 |
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
834 |
if self.config.add_cross_attention and encoder_hidden_states is not None:
|
835 |
(
|
@@ -860,8 +1043,9 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
860 |
hidden_states = hidden_states + token_type_embeds
|
861 |
|
862 |
hidden_states = self.drop(hidden_states)
|
863 |
-
|
864 |
-
|
|
|
865 |
output_shape = input_shape + (hidden_states.size(-1),)
|
866 |
|
867 |
presents = () if use_cache else None
|
@@ -885,28 +1069,28 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
885 |
attention_mask = attention_mask.to(hidden_states.device)
|
886 |
if isinstance(head_mask, torch.Tensor):
|
887 |
head_mask = head_mask.to(hidden_states.device)
|
888 |
-
|
889 |
if output_hidden_states:
|
890 |
-
all_hidden_states = all_hidden_states + (
|
891 |
-
hidden_states.view(*output_shape),
|
892 |
-
)
|
893 |
|
894 |
-
if
|
|
|
|
|
|
|
|
|
|
|
|
|
895 |
|
896 |
def create_custom_forward(module):
|
897 |
def custom_forward(*inputs):
|
898 |
-
#
|
899 |
-
return
|
900 |
-
output
|
901 |
-
for output in module(*inputs, use_cache, output_attentions)
|
902 |
-
)
|
903 |
|
904 |
return custom_forward
|
905 |
|
906 |
outputs = torch.utils.checkpoint.checkpoint(
|
907 |
create_custom_forward(block),
|
908 |
hidden_states,
|
909 |
-
|
910 |
attention_mask,
|
911 |
head_mask[i],
|
912 |
encoder_hidden_states,
|
@@ -924,9 +1108,9 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
924 |
output_attentions=output_attentions,
|
925 |
)
|
926 |
|
927 |
-
hidden_states
|
928 |
if use_cache is True:
|
929 |
-
presents = presents + (
|
930 |
|
931 |
if output_attentions:
|
932 |
all_self_attentions = all_self_attentions + (
|
@@ -943,10 +1127,10 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
943 |
if i == v[-1] and "cuda:" + str(k) != self.last_device:
|
944 |
hidden_states = hidden_states.to("cuda:" + str(k + 1))
|
945 |
|
946 |
-
|
947 |
-
|
948 |
|
949 |
-
hidden_states = hidden_states.view(
|
950 |
# Add last hidden state
|
951 |
if output_hidden_states:
|
952 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
@@ -981,19 +1165,24 @@ class GPT2Model(GPT2PreTrainedModel):
|
|
981 |
GPT2_START_DOCSTRING,
|
982 |
)
|
983 |
class GPT2LMHeadModel(GPT2PreTrainedModel):
|
984 |
-
_keys_to_ignore_on_load_missing = [
|
|
|
|
|
|
|
|
|
985 |
|
986 |
def __init__(self, config):
|
987 |
super().__init__(config)
|
988 |
self.transformer = GPT2Model(config)
|
989 |
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
990 |
|
991 |
-
self.init_weights()
|
992 |
-
|
993 |
# Model parallel
|
994 |
self.model_parallel = False
|
995 |
self.device_map = None
|
996 |
|
|
|
|
|
|
|
997 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
998 |
def parallelize(self, device_map=None):
|
999 |
self.device_map = (
|
@@ -1017,6 +1206,9 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
|
|
1017 |
def get_output_embeddings(self):
|
1018 |
return self.lm_head
|
1019 |
|
|
|
|
|
|
|
1020 |
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
|
1021 |
token_type_ids = kwargs.get("token_type_ids", None)
|
1022 |
# only last token for inputs_ids if past is defined in kwargs
|
@@ -1047,33 +1239,33 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
|
|
1047 |
|
1048 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
1049 |
@add_code_sample_docstrings(
|
1050 |
-
|
1051 |
-
checkpoint=
|
1052 |
output_type=CausalLMOutputWithCrossAttentions,
|
1053 |
config_class=_CONFIG_FOR_DOC,
|
1054 |
)
|
1055 |
def forward(
|
1056 |
self,
|
1057 |
-
input_ids=None,
|
1058 |
-
past_key_values=None,
|
1059 |
-
attention_mask=None,
|
1060 |
-
token_type_ids=None,
|
1061 |
-
position_ids=None,
|
1062 |
-
head_mask=None,
|
1063 |
-
inputs_embeds=None,
|
1064 |
-
encoder_hidden_states=None,
|
1065 |
-
encoder_attention_mask=None,
|
1066 |
-
labels=None,
|
1067 |
-
use_cache=None,
|
1068 |
-
output_attentions=None,
|
1069 |
-
output_hidden_states=None,
|
1070 |
-
return_dict=None,
|
1071 |
-
):
|
1072 |
r"""
|
1073 |
-
labels (
|
1074 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
1075 |
-
|
1076 |
-
|
1077 |
"""
|
1078 |
return_dict = (
|
1079 |
return_dict if return_dict is not None else self.config.use_return_dict
|
@@ -1132,9 +1324,9 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
|
|
1132 |
past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
1133 |
) -> Tuple[Tuple[torch.Tensor]]:
|
1134 |
"""
|
1135 |
-
This function is used to re-order the
|
1136 |
-
|
1137 |
-
|
1138 |
"""
|
1139 |
return tuple(
|
1140 |
tuple(
|
@@ -1155,6 +1347,12 @@ input sequence).
|
|
1155 |
GPT2_START_DOCSTRING,
|
1156 |
)
|
1157 |
class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
|
|
|
|
|
|
|
|
|
|
|
1158 |
def __init__(self, config):
|
1159 |
super().__init__(config)
|
1160 |
config.num_labels = 1
|
@@ -1162,12 +1360,13 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
1162 |
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
1163 |
self.multiple_choice_head = SequenceSummary(config)
|
1164 |
|
1165 |
-
self.init_weights()
|
1166 |
-
|
1167 |
# Model parallel
|
1168 |
self.model_parallel = False
|
1169 |
self.device_map = None
|
1170 |
|
|
|
|
|
|
|
1171 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
1172 |
def parallelize(self, device_map=None):
|
1173 |
self.device_map = (
|
@@ -1195,6 +1394,9 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
1195 |
def get_output_embeddings(self):
|
1196 |
return self.lm_head
|
1197 |
|
|
|
|
|
|
|
1198 |
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
|
1199 |
token_type_ids = kwargs.get("token_type_ids", None)
|
1200 |
# only last token for inputs_ids if past is defined in kwargs
|
@@ -1230,62 +1432,61 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
1230 |
)
|
1231 |
def forward(
|
1232 |
self,
|
1233 |
-
input_ids=None,
|
1234 |
-
past_key_values=None,
|
1235 |
-
attention_mask=None,
|
1236 |
-
token_type_ids=None,
|
1237 |
-
position_ids=None,
|
1238 |
-
head_mask=None,
|
1239 |
-
inputs_embeds=None,
|
1240 |
-
mc_token_ids=None,
|
1241 |
-
labels=None,
|
1242 |
-
mc_labels=None,
|
1243 |
-
use_cache=None,
|
1244 |
-
output_attentions=None,
|
1245 |
-
output_hidden_states=None,
|
1246 |
-
return_dict=None,
|
1247 |
**kwargs,
|
1248 |
-
):
|
1249 |
r"""
|
1250 |
-
mc_token_ids (
|
1251 |
-
Index of the classification token in each input sequence. Selected in the range
|
1252 |
-
1[
|
1253 |
-
labels (
|
1254 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
1255 |
-
|
1256 |
-
|
1257 |
-
mc_labels (
|
1258 |
-
Labels for computing the multiple choice classification loss. Indices should be in
|
1259 |
-
|
1260 |
-
`input_ids` above)
|
1261 |
|
1262 |
Return:
|
1263 |
|
1264 |
-
Example
|
1265 |
-
|
1266 |
-
>>> import torch
|
1267 |
-
>>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
|
1268 |
-
|
1269 |
-
>>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
|
1270 |
-
>>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
|
1271 |
|
1272 |
-
|
1273 |
-
|
|
|
1274 |
|
1275 |
-
|
|
|
1276 |
|
1277 |
-
|
1278 |
-
|
1279 |
-
|
|
|
1280 |
|
1281 |
-
|
1282 |
-
|
|
|
1283 |
|
1284 |
-
|
1285 |
-
|
1286 |
-
>>> mc_logits = outputs.mc_logits
|
1287 |
|
1288 |
-
|
|
|
|
|
|
|
1289 |
return_dict = (
|
1290 |
return_dict if return_dict is not None else self.config.use_return_dict
|
1291 |
)
|
@@ -1350,9 +1551,9 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
1350 |
past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
1351 |
) -> Tuple[Tuple[torch.Tensor]]:
|
1352 |
"""
|
1353 |
-
This function is used to re-order the
|
1354 |
-
|
1355 |
-
|
1356 |
"""
|
1357 |
return tuple(
|
1358 |
tuple(
|
@@ -1367,14 +1568,14 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
|
1367 |
"""
|
1368 |
The GPT2 Model transformer with a sequence classification head on top (linear layer).
|
1369 |
|
1370 |
-
|
1371 |
-
|
1372 |
|
1373 |
Since it does classification on the last token, it requires to know the position of the last token. If a
|
1374 |
-
|
1375 |
-
|
1376 |
-
|
1377 |
-
|
1378 |
""",
|
1379 |
GPT2_START_DOCSTRING,
|
1380 |
)
|
@@ -1387,39 +1588,42 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
|
|
1387 |
self.transformer = GPT2Model(config)
|
1388 |
self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
|
1389 |
|
1390 |
-
self.init_weights()
|
1391 |
-
|
1392 |
# Model parallel
|
1393 |
self.model_parallel = False
|
1394 |
self.device_map = None
|
1395 |
|
|
|
|
|
|
|
1396 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
1397 |
@add_code_sample_docstrings(
|
1398 |
-
|
1399 |
-
checkpoint="microsoft/
|
1400 |
output_type=SequenceClassifierOutputWithPast,
|
1401 |
config_class=_CONFIG_FOR_DOC,
|
|
|
|
|
1402 |
)
|
1403 |
def forward(
|
1404 |
self,
|
1405 |
-
input_ids=None,
|
1406 |
-
past_key_values=None,
|
1407 |
-
attention_mask=None,
|
1408 |
-
token_type_ids=None,
|
1409 |
-
position_ids=None,
|
1410 |
-
head_mask=None,
|
1411 |
-
inputs_embeds=None,
|
1412 |
-
labels=None,
|
1413 |
-
use_cache=None,
|
1414 |
-
output_attentions=None,
|
1415 |
-
output_hidden_states=None,
|
1416 |
-
return_dict=None,
|
1417 |
-
):
|
1418 |
r"""
|
1419 |
-
labels (
|
1420 |
-
Labels for computing the sequence classification/regression loss. Indices should be in
|
1421 |
-
config.num_labels - 1]`. If
|
1422 |
-
|
1423 |
"""
|
1424 |
return_dict = (
|
1425 |
return_dict if return_dict is not None else self.config.use_return_dict
|
@@ -1460,23 +1664,39 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
|
|
1460 |
sequence_lengths = -1
|
1461 |
logger.warning(
|
1462 |
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
|
1463 |
-
|
1464 |
)
|
1465 |
|
1466 |
-
pooled_logits = logits[
|
|
|
|
|
1467 |
|
1468 |
loss = None
|
1469 |
if labels is not None:
|
1470 |
-
if self.
|
1471 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1472 |
loss_fct = MSELoss()
|
1473 |
-
|
1474 |
-
|
|
|
|
|
|
|
1475 |
loss_fct = CrossEntropyLoss()
|
1476 |
loss = loss_fct(
|
1477 |
pooled_logits.view(-1, self.num_labels), labels.view(-1)
|
1478 |
)
|
1479 |
-
|
|
|
|
|
1480 |
if not return_dict:
|
1481 |
output = (pooled_logits,) + transformer_outputs[1:]
|
1482 |
return ((loss,) + output) if loss is not None else output
|
@@ -1515,39 +1735,44 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
|
|
1515 |
self.dropout = nn.Dropout(classifier_dropout)
|
1516 |
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
1517 |
|
1518 |
-
self.init_weights()
|
1519 |
-
|
1520 |
# Model parallel
|
1521 |
self.model_parallel = False
|
1522 |
self.device_map = None
|
1523 |
|
|
|
|
|
|
|
1524 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
|
|
1525 |
@add_code_sample_docstrings(
|
1526 |
-
|
1527 |
-
checkpoint="
|
1528 |
output_type=TokenClassifierOutput,
|
1529 |
config_class=_CONFIG_FOR_DOC,
|
|
|
|
|
1530 |
)
|
|
|
1531 |
def forward(
|
1532 |
self,
|
1533 |
-
input_ids=None,
|
1534 |
-
past_key_values=None,
|
1535 |
-
attention_mask=None,
|
1536 |
-
token_type_ids=None,
|
1537 |
-
position_ids=None,
|
1538 |
-
head_mask=None,
|
1539 |
-
inputs_embeds=None,
|
1540 |
-
labels=None,
|
1541 |
-
use_cache=None,
|
1542 |
-
output_attentions=None,
|
1543 |
-
output_hidden_states=None,
|
1544 |
-
return_dict=None,
|
1545 |
-
):
|
1546 |
r"""
|
1547 |
-
labels (
|
1548 |
-
Labels for computing the sequence classification/regression loss. Indices should be in
|
1549 |
-
config.num_labels - 1]`. If
|
1550 |
-
|
1551 |
"""
|
1552 |
return_dict = (
|
1553 |
return_dict if return_dict is not None else self.config.use_return_dict
|
@@ -1574,18 +1799,7 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
|
|
1574 |
loss = None
|
1575 |
if labels is not None:
|
1576 |
loss_fct = CrossEntropyLoss()
|
1577 |
-
|
1578 |
-
if attention_mask is not None:
|
1579 |
-
active_loss = attention_mask.view(-1) == 1
|
1580 |
-
active_logits = logits.view(-1, self.num_labels)
|
1581 |
-
active_labels = torch.where(
|
1582 |
-
active_loss,
|
1583 |
-
labels.view(-1),
|
1584 |
-
torch.tensor(loss_fct.ignore_index).type_as(labels),
|
1585 |
-
)
|
1586 |
-
loss = loss_fct(active_logits, active_labels)
|
1587 |
-
else:
|
1588 |
-
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
1589 |
|
1590 |
if not return_dict:
|
1591 |
output = (logits,) + transformer_outputs[2:]
|
@@ -1596,4 +1810,4 @@ class GPT2ForTokenClassification(GPT2PreTrainedModel):
|
|
1596 |
logits=logits,
|
1597 |
hidden_states=transformer_outputs.hidden_states,
|
1598 |
attentions=transformer_outputs.attentions,
|
1599 |
-
)
|
|
|
13 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
# See the License for the specific language governing permissions and
|
15 |
# limitations under the License.
|
16 |
+
"""PyTorch GROVER model."""
|
17 |
|
18 |
+
import math
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
import os
|
20 |
from dataclasses import dataclass
|
21 |
+
from typing import Optional, Tuple, Union
|
22 |
|
23 |
import torch
|
24 |
+
import torch.utils.checkpoint
|
25 |
+
from packaging import version
|
26 |
+
from torch import nn
|
27 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
28 |
+
|
29 |
+
|
30 |
+
if version.parse(torch.__version__) >= version.parse("1.6"):
|
31 |
+
is_amp_available = True
|
32 |
+
from torch.cuda.amp import autocast
|
33 |
+
else:
|
34 |
+
is_amp_available = False
|
35 |
+
|
36 |
from transformers.activations import ACT2FN
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
from transformers.modeling_outputs import (
|
38 |
BaseModelOutputWithPastAndCrossAttentions,
|
39 |
CausalLMOutputWithCrossAttentions,
|
40 |
SequenceClassifierOutputWithPast,
|
41 |
TokenClassifierOutput,
|
42 |
)
|
43 |
+
from transformers.modeling_utils import PreTrainedModel, SequenceSummary
|
44 |
+
from transformers.pytorch_utils import (
|
45 |
Conv1D,
|
|
|
|
|
46 |
find_pruneable_heads_and_indices,
|
47 |
prune_conv1d_layer,
|
48 |
)
|
49 |
+
from transformers.utils import (
|
50 |
+
ModelOutput,
|
51 |
+
add_code_sample_docstrings,
|
52 |
+
add_start_docstrings,
|
53 |
+
add_start_docstrings_to_model_forward,
|
54 |
+
logging,
|
55 |
+
replace_return_docstrings,
|
56 |
+
)
|
57 |
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
|
58 |
+
from transformers import GPT2Config
|
59 |
|
|
|
|
|
60 |
|
61 |
+
logger = logging.get_logger(__name__)
|
62 |
|
63 |
+
_CHECKPOINT_FOR_DOC = "gpt2"
|
64 |
_CONFIG_FOR_DOC = "GPT2Config"
|
65 |
_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
|
66 |
|
|
|
73 |
# See all GPT-2 models at https://huggingface.co/models?filter=gpt2
|
74 |
]
|
75 |
|
|
|
|
|
|
|
|
|
|
|
76 |
_GPT2_ML_TF_TO_TORCH = {
|
77 |
"LayerNorm_embed_norm": "emb_norm",
|
78 |
"pos_embed": "wpe.weight",
|
|
|
124 |
"""Load tf checkpoints in a pytorch model"""
|
125 |
try:
|
126 |
import re
|
|
|
127 |
import tensorflow as tf
|
128 |
except ImportError:
|
129 |
logger.error(
|
|
|
203 |
d = torch.from_numpy(array)
|
204 |
is_bias = len(shape) == 1
|
205 |
end = int(shape[0 if is_bias else 1] / 3)
|
206 |
+
m = dict(query_layer=0, key_layer=end, value_layer=end * 2,)
|
|
|
|
|
|
|
|
|
207 |
start = m[attn_layer]
|
208 |
end = start + end
|
209 |
if is_bias:
|
|
|
225 |
return model
|
226 |
|
227 |
|
228 |
+
class GPT2Attention(nn.Module):
|
229 |
+
def __init__(self, config, is_cross_attention=False, layer_idx=None):
|
230 |
super().__init__()
|
231 |
|
232 |
+
max_positions = config.max_position_embeddings
|
|
|
|
|
233 |
self.register_buffer(
|
234 |
"bias",
|
235 |
+
torch.tril(
|
236 |
+
torch.ones((max_positions, max_positions), dtype=torch.uint8)
|
237 |
+
).view(1, 1, max_positions, max_positions),
|
238 |
)
|
239 |
self.register_buffer("masked_bias", torch.tensor(-1e4))
|
240 |
+
|
241 |
+
self.embed_dim = config.hidden_size
|
242 |
+
self.num_heads = config.num_attention_heads
|
243 |
+
self.head_dim = self.embed_dim // self.num_heads
|
244 |
+
self.split_size = self.embed_dim
|
245 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
246 |
+
raise ValueError(
|
247 |
+
f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
248 |
+
f" {self.num_heads})."
|
249 |
+
)
|
250 |
+
|
251 |
+
self.scale_attn_weights = config.scale_attn_weights
|
252 |
self.is_cross_attention = is_cross_attention
|
253 |
+
|
254 |
+
# Layer-wise attention scaling, reordering, and upcasting
|
255 |
+
self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
|
256 |
+
self.layer_idx = layer_idx
|
257 |
+
self.reorder_and_upcast_attn = config.reorder_and_upcast_attn
|
258 |
+
|
259 |
if self.is_cross_attention:
|
260 |
+
self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
|
261 |
+
self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
|
262 |
else:
|
263 |
+
self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
|
264 |
+
self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
|
265 |
+
|
266 |
self.attn_dropout = nn.Dropout(config.attn_pdrop)
|
267 |
self.resid_dropout = nn.Dropout(config.resid_pdrop)
|
268 |
+
|
269 |
self.pruned_heads = set()
|
270 |
|
271 |
def prune_heads(self, heads):
|
272 |
if len(heads) == 0:
|
273 |
return
|
274 |
heads, index = find_pruneable_heads_and_indices(
|
275 |
+
heads, self.num_heads, self.head_dim, self.pruned_heads
|
276 |
)
|
277 |
index_attn = torch.cat(
|
278 |
[index, index + self.split_size, index + (2 * self.split_size)]
|
|
|
283 |
self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
|
284 |
|
285 |
# Update hyper params
|
286 |
+
self.split_size = (self.split_size // self.num_heads) * (
|
287 |
+
self.num_heads - len(heads)
|
288 |
+
)
|
289 |
+
self.num_heads = self.num_heads - len(heads)
|
290 |
self.pruned_heads = self.pruned_heads.union(heads)
|
291 |
|
292 |
+
def _attn(self, query, key, value, attention_mask=None, head_mask=None):
|
293 |
+
attn_weights = torch.matmul(query, key.transpose(-1, -2))
|
294 |
+
|
295 |
+
if self.scale_attn_weights:
|
296 |
+
attn_weights = attn_weights / (value.size(-1) ** 0.5)
|
297 |
+
|
298 |
+
# Layer-wise attention scaling
|
299 |
+
if self.scale_attn_by_inverse_layer_idx:
|
300 |
+
attn_weights = attn_weights / float(self.layer_idx + 1)
|
301 |
|
302 |
if not self.is_cross_attention:
|
303 |
# if only "normal" attention layer implements causal mask
|
304 |
+
query_length, key_length = query.size(-2), key.size(-2)
|
305 |
+
causal_mask = self.bias[
|
306 |
+
:, :, key_length - query_length : key_length, :key_length
|
307 |
+
].bool()
|
308 |
+
attn_weights = torch.where(
|
309 |
+
causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype)
|
310 |
+
)
|
311 |
|
312 |
if attention_mask is not None:
|
313 |
# Apply the attention mask
|
314 |
+
attn_weights = attn_weights + attention_mask
|
315 |
|
316 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
317 |
+
|
318 |
+
# Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
|
319 |
+
attn_weights = attn_weights.type(value.dtype)
|
320 |
+
attn_weights = self.attn_dropout(attn_weights)
|
321 |
|
322 |
# Mask heads if we want to
|
323 |
if head_mask is not None:
|
324 |
+
attn_weights = attn_weights * head_mask
|
325 |
|
326 |
+
attn_output = torch.matmul(attn_weights, value)
|
327 |
+
|
328 |
+
return attn_output, attn_weights
|
329 |
+
|
330 |
+
def _upcast_and_reordered_attn(
|
331 |
+
self, query, key, value, attention_mask=None, head_mask=None
|
332 |
+
):
|
333 |
+
# Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
|
334 |
+
bsz, num_heads, q_seq_len, dk = query.size()
|
335 |
+
_, _, k_seq_len, _ = key.size()
|
336 |
+
|
337 |
+
# Preallocate attn_weights for `baddbmm`
|
338 |
+
attn_weights = torch.empty(
|
339 |
+
bsz * num_heads,
|
340 |
+
q_seq_len,
|
341 |
+
k_seq_len,
|
342 |
+
dtype=torch.float32,
|
343 |
+
device=query.device,
|
344 |
+
)
|
345 |
+
|
346 |
+
# Compute Scale Factor
|
347 |
+
scale_factor = 1.0
|
348 |
+
if self.scale_attn_weights:
|
349 |
+
scale_factor /= float(value.size(-1)) ** 0.5
|
350 |
+
|
351 |
+
if self.scale_attn_by_inverse_layer_idx:
|
352 |
+
scale_factor /= float(self.layer_idx + 1)
|
353 |
+
|
354 |
+
# Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
|
355 |
+
if is_amp_available:
|
356 |
+
with autocast(enabled=False):
|
357 |
+
q, k = (
|
358 |
+
query.reshape(-1, q_seq_len, dk),
|
359 |
+
key.transpose(-1, -2).reshape(-1, dk, k_seq_len),
|
360 |
+
)
|
361 |
+
attn_weights = torch.baddbmm(
|
362 |
+
attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
|
363 |
+
)
|
364 |
+
attn_weights = attn_weights.reshape(
|
365 |
+
bsz, num_heads, q_seq_len, k_seq_len
|
366 |
+
)
|
367 |
else:
|
368 |
+
q, k = (
|
369 |
+
query.reshape(-1, q_seq_len, dk),
|
370 |
+
key.transpose(-1, -2).reshape(-1, dk, k_seq_len),
|
371 |
+
)
|
372 |
+
attn_weights = torch.baddbmm(
|
373 |
+
attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
|
374 |
+
)
|
375 |
+
attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
|
376 |
+
|
377 |
+
if not self.is_cross_attention:
|
378 |
+
# if only "normal" attention layer implements causal mask
|
379 |
+
query_length, key_length = query.size(-2), key.size(-2)
|
380 |
+
causal_mask = self.bias[
|
381 |
+
:, :, key_length - query_length : key_length, :key_length
|
382 |
+
].bool()
|
383 |
+
attn_weights = torch.where(
|
384 |
+
causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype)
|
385 |
+
)
|
386 |
+
|
387 |
+
if attention_mask is not None:
|
388 |
+
# Apply the attention mask
|
389 |
+
attn_weights = attn_weights + attention_mask
|
390 |
+
|
391 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
392 |
+
|
393 |
+
# Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
|
394 |
+
if attn_weights.dtype != torch.float32:
|
395 |
+
raise RuntimeError(
|
396 |
+
"Error with upcasting, attn_weights does not have dtype torch.float32"
|
397 |
+
)
|
398 |
+
attn_weights = attn_weights.type(value.dtype)
|
399 |
+
attn_weights = self.attn_dropout(attn_weights)
|
400 |
+
|
401 |
+
# Mask heads if we want to
|
402 |
+
if head_mask is not None:
|
403 |
+
attn_weights = attn_weights * head_mask
|
404 |
+
|
405 |
+
attn_output = torch.matmul(attn_weights, value)
|
406 |
+
|
407 |
+
return attn_output, attn_weights
|
408 |
+
|
409 |
+
def _split_heads(self, tensor, num_heads, attn_head_size):
|
410 |
+
"""
|
411 |
+
Splits hidden_size dim into attn_head_size and num_heads
|
412 |
+
"""
|
413 |
+
new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
|
414 |
+
tensor = tensor.view(new_shape)
|
415 |
+
return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
|
416 |
+
|
417 |
+
def _merge_heads(self, tensor, num_heads, attn_head_size):
|
418 |
+
"""
|
419 |
+
Merges attn_head_size dim and num_attn_heads dim into hidden_size
|
420 |
+
"""
|
421 |
+
tensor = tensor.permute(0, 2, 1, 3).contiguous()
|
422 |
+
new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
|
423 |
+
return tensor.view(new_shape)
|
424 |
|
425 |
def forward(
|
426 |
self,
|
427 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]],
|
428 |
+
layer_past: Optional[Tuple[torch.Tensor]] = None,
|
429 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
430 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
431 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
432 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
433 |
+
use_cache: Optional[bool] = False,
|
434 |
+
output_attentions: Optional[bool] = False,
|
435 |
+
) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
|
436 |
if encoder_hidden_states is not None:
|
437 |
+
if not hasattr(self, "q_attn"):
|
438 |
+
raise ValueError(
|
439 |
+
"If class is used as cross attention, the weights `q_attn` have to be defined. "
|
440 |
+
"Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
|
441 |
+
)
|
442 |
+
|
443 |
query = self.q_attn(hidden_states)
|
444 |
key, value = self.c_attn(encoder_hidden_states).split(
|
445 |
self.split_size, dim=2
|
|
|
448 |
else:
|
449 |
query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
|
450 |
|
451 |
+
query = self._split_heads(query, self.num_heads, self.head_dim)
|
452 |
+
key = self._split_heads(key, self.num_heads, self.head_dim)
|
453 |
+
value = self._split_heads(value, self.num_heads, self.head_dim)
|
454 |
+
|
455 |
if layer_past is not None:
|
456 |
+
past_key, past_value = layer_past
|
457 |
+
key = torch.cat((past_key, key), dim=-2)
|
|
|
|
|
|
|
458 |
value = torch.cat((past_value, value), dim=-2)
|
459 |
|
460 |
if use_cache is True:
|
461 |
+
present = (key, value)
|
|
|
|
|
462 |
else:
|
463 |
+
present = None
|
464 |
|
465 |
+
if self.reorder_and_upcast_attn:
|
466 |
+
attn_output, attn_weights = self._upcast_and_reordered_attn(
|
467 |
+
query, key, value, attention_mask, head_mask
|
468 |
+
)
|
469 |
+
else:
|
470 |
+
attn_output, attn_weights = self._attn(
|
471 |
+
query, key, value, attention_mask, head_mask
|
472 |
+
)
|
473 |
+
|
474 |
+
attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
|
475 |
+
attn_output = self.c_proj(attn_output)
|
476 |
+
attn_output = self.resid_dropout(attn_output)
|
477 |
|
478 |
+
outputs = (attn_output, present)
|
479 |
+
if output_attentions:
|
480 |
+
outputs += (attn_weights,)
|
481 |
|
|
|
482 |
return outputs # a, present, (attentions)
|
483 |
|
484 |
|
485 |
+
class GPT2MLP(nn.Module):
|
486 |
+
def __init__(self, intermediate_size, config):
|
487 |
super().__init__()
|
488 |
+
embed_dim = config.hidden_size
|
489 |
+
self.c_fc = Conv1D(intermediate_size, embed_dim)
|
490 |
+
self.c_proj = Conv1D(embed_dim, intermediate_size)
|
491 |
self.act = ACT2FN[config.activation_function]
|
492 |
self.dropout = nn.Dropout(config.resid_pdrop)
|
493 |
|
494 |
+
def forward(
|
495 |
+
self, hidden_states: Optional[Tuple[torch.FloatTensor]]
|
496 |
+
) -> torch.FloatTensor:
|
497 |
+
hidden_states = self.c_fc(hidden_states)
|
498 |
+
hidden_states = self.act(hidden_states)
|
499 |
+
hidden_states = self.c_proj(hidden_states)
|
500 |
+
hidden_states = self.dropout(hidden_states)
|
501 |
+
return hidden_states
|
502 |
|
503 |
|
504 |
+
class GPT2Block(nn.Module):
|
505 |
+
def __init__(self, config, layer_idx=None):
|
506 |
super().__init__()
|
507 |
+
hidden_size = config.hidden_size
|
508 |
inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
|
509 |
+
|
510 |
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
511 |
+
self.attn = GPT2Attention(config, layer_idx=layer_idx)
|
512 |
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
513 |
+
|
514 |
if config.add_cross_attention:
|
515 |
+
self.crossattention = GPT2Attention(
|
516 |
+
config, is_cross_attention=True, layer_idx=layer_idx
|
517 |
)
|
518 |
self.ln_cross_attn = nn.LayerNorm(
|
519 |
hidden_size, eps=config.layer_norm_epsilon
|
520 |
)
|
521 |
+
|
522 |
+
self.mlp = GPT2MLP(inner_dim, config)
|
523 |
|
524 |
def forward(
|
525 |
self,
|
526 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]],
|
527 |
+
layer_past: Optional[Tuple[torch.Tensor]] = None,
|
528 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
529 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
530 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
531 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
532 |
+
use_cache: Optional[bool] = False,
|
533 |
+
output_attentions: Optional[bool] = False,
|
534 |
+
) -> Union[
|
535 |
+
Tuple[torch.Tensor],
|
536 |
+
Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]],
|
537 |
+
]:
|
538 |
+
|
539 |
+
# removed in GROVER
|
540 |
+
# residual = hidden_states
|
541 |
+
# hidden_states = self.ln_1(hidden_states)
|
542 |
attn_outputs = self.attn(
|
543 |
hidden_states,
|
544 |
layer_past=layer_past,
|
|
|
554 |
|
555 |
if encoder_hidden_states is not None:
|
556 |
# add one self-attention block for cross-attention
|
557 |
+
if not hasattr(self, "crossattention"):
|
558 |
+
raise ValueError(
|
559 |
+
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
|
560 |
+
"cross-attention layers by setting `config.add_cross_attention=True`"
|
561 |
+
)
|
562 |
+
# removed in GROVER
|
563 |
+
# residual = hidden_states
|
564 |
+
# hidden_states = self.ln_cross_attn(hidden_states)
|
565 |
cross_attn_outputs = self.crossattention(
|
566 |
+
hidden_states,
|
567 |
attention_mask=attention_mask,
|
568 |
head_mask=head_mask,
|
569 |
encoder_hidden_states=encoder_hidden_states,
|
|
|
572 |
)
|
573 |
attn_output = cross_attn_outputs[0]
|
574 |
# residual connection
|
575 |
+
hidden_states = attn_output + hidden_states
|
576 |
outputs = (
|
577 |
outputs + cross_attn_outputs[2:]
|
578 |
) # add cross attentions if we output attention weights
|
579 |
|
580 |
+
residual = hidden_states
|
581 |
+
hidden_states = self.ln_1(hidden_states)
|
582 |
+
feed_forward_hidden_states = self.mlp(hidden_states)
|
583 |
# residual connection
|
584 |
+
hidden_states = residual + feed_forward_hidden_states
|
585 |
|
586 |
+
hidden_states = self.ln_2(hidden_states) # Added in GROVER
|
587 |
+
|
588 |
+
if use_cache:
|
589 |
+
outputs = (hidden_states,) + outputs
|
590 |
+
else:
|
591 |
+
outputs = (hidden_states,) + outputs[1:]
|
592 |
|
|
|
593 |
return outputs # hidden_states, present, (attentions, cross_attentions)
|
594 |
|
595 |
|
|
|
603 |
load_tf_weights = load_tf_weights_in_gpt2
|
604 |
base_model_prefix = "transformer"
|
605 |
is_parallelizable = True
|
606 |
+
supports_gradient_checkpointing = True
|
607 |
|
608 |
def __init__(self, *inputs, **kwargs):
|
609 |
super().__init__(*inputs, **kwargs)
|
610 |
|
611 |
def _init_weights(self, module):
|
612 |
"""Initialize the weights."""
|
613 |
+
if isinstance(module, (nn.Linear, Conv1D)):
|
614 |
# Slightly different from the TF version which uses truncated_normal for initialization
|
615 |
# cf https://github.com/pytorch/pytorch/pull/5617
|
616 |
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
617 |
+
if module.bias is not None:
|
618 |
module.bias.data.zero_()
|
619 |
+
elif isinstance(module, nn.Embedding):
|
620 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
621 |
+
if module.padding_idx is not None:
|
622 |
+
module.weight.data[module.padding_idx].zero_()
|
623 |
elif isinstance(module, nn.LayerNorm):
|
624 |
module.bias.data.zero_()
|
625 |
module.weight.data.fill_(1.0)
|
626 |
|
627 |
+
# Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
|
628 |
+
# > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
|
629 |
+
# > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
|
630 |
+
# > -- GPT-2 :: https://openai.com/blog/better-language-models/
|
631 |
+
#
|
632 |
+
# Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
|
633 |
+
for name, p in module.named_parameters():
|
634 |
+
if "c_proj" in name and "weight" in name:
|
635 |
+
# Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
|
636 |
+
p.data.normal_(
|
637 |
+
mean=0.0,
|
638 |
+
std=(
|
639 |
+
self.config.initializer_range
|
640 |
+
/ math.sqrt(2 * self.config.n_layer)
|
641 |
+
),
|
642 |
+
)
|
643 |
+
|
644 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
645 |
+
if isinstance(module, GPT2Model):
|
646 |
+
module.gradient_checkpointing = value
|
647 |
+
|
648 |
|
649 |
@dataclass
|
650 |
class GPT2DoubleHeadsModelOutput(ModelOutput):
|
|
|
652 |
Base class for outputs of models predicting if two sentences are consecutive or not.
|
653 |
|
654 |
Args:
|
655 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
|
656 |
Language modeling loss.
|
657 |
+
mc_loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `mc_labels` is provided):
|
658 |
Multiple choice classification loss.
|
659 |
+
logits (`torch.FloatTensor` of shape `(batch_size, num_choices, sequence_length, config.vocab_size)`):
|
660 |
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
661 |
+
mc_logits (`torch.FloatTensor` of shape `(batch_size, num_choices)`):
|
662 |
Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
|
663 |
+
past_key_values (`Tuple[Tuple[torch.Tensor]]`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
664 |
+
Tuple of length `config.n_layers`, containing tuples of tensors of shape `(batch_size, num_heads,
|
665 |
+
sequence_length, embed_size_per_head)`).
|
666 |
|
667 |
Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
|
668 |
+
`past_key_values` input) to speed up sequential decoding.
|
669 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
670 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
|
671 |
+
shape `(batch_size, sequence_length, hidden_size)`.
|
672 |
|
673 |
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
674 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
675 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
676 |
+
sequence_length)`.
|
677 |
|
678 |
+
GPT2Attentions weights after the attention softmax, used to compute the weighted average in the
|
679 |
+
self-attention heads.
|
680 |
"""
|
681 |
|
682 |
loss: Optional[torch.FloatTensor] = None
|
683 |
mc_loss: Optional[torch.FloatTensor] = None
|
684 |
logits: torch.FloatTensor = None
|
685 |
mc_logits: torch.FloatTensor = None
|
686 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
|
687 |
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
688 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
689 |
|
690 |
|
691 |
GPT2_START_DOCSTRING = r"""
|
692 |
|
693 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
694 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
695 |
+
etc.)
|
696 |
|
697 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
698 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
699 |
+
and behavior.
|
700 |
|
701 |
Parameters:
|
702 |
+
config ([`GPT2Config`]): Model configuration class with all the parameters of the model.
|
703 |
Initializing with a config file does not load the weights associated with the model, only the
|
704 |
+
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
|
|
705 |
"""
|
706 |
|
707 |
GPT2_INPUTS_DOCSTRING = r"""
|
708 |
Args:
|
709 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
710 |
+
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
711 |
+
`past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
|
712 |
sequence tokens in the vocabulary.
|
713 |
|
714 |
+
If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
|
715 |
+
`input_ids`.
|
716 |
|
717 |
+
Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
718 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
|
|
719 |
|
720 |
+
[What are input IDs?](../glossary#input-ids)
|
721 |
+
past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
|
722 |
Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
|
723 |
+
`past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
|
724 |
+
their past given to this model should not be passed as `input_ids` as they have already been computed.
|
725 |
+
attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
726 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
|
|
727 |
|
728 |
- 1 for tokens that are **not masked**,
|
729 |
- 0 for tokens that are **masked**.
|
730 |
|
731 |
+
If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
|
732 |
+
`past_key_values`. In other words, the `attention_mask` always has to have the length:
|
733 |
+
`len(past_key_values) + len(input_ids)`
|
|
|
734 |
|
735 |
+
[What are attention masks?](../glossary#attention-mask)
|
736 |
+
token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
|
737 |
+
Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
|
738 |
+
1]`:
|
739 |
|
740 |
+
- 0 corresponds to a *sentence A* token,
|
741 |
+
- 1 corresponds to a *sentence B* token.
|
|
|
|
|
742 |
|
743 |
+
[What are token type IDs?](../glossary#token-type-ids)
|
744 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
745 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
746 |
+
config.max_position_embeddings - 1]`.
|
747 |
+
|
748 |
+
[What are position IDs?](../glossary#position-ids)
|
749 |
+
head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
|
750 |
+
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
|
751 |
|
752 |
- 1 indicates the head is **not masked**,
|
753 |
- 0 indicates the head is **masked**.
|
754 |
|
755 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
756 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
757 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
758 |
+
model's internal embedding lookup matrix.
|
759 |
+
|
760 |
+
If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
|
761 |
+
`past_key_values`).
|
762 |
+
use_cache (`bool`, *optional*):
|
763 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
764 |
+
`past_key_values`).
|
765 |
+
output_attentions (`bool`, *optional*):
|
766 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
767 |
tensors for more detail.
|
768 |
+
output_hidden_states (`bool`, *optional*):
|
769 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
770 |
more detail.
|
771 |
+
return_dict (`bool`, *optional*):
|
772 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
773 |
"""
|
|
|
774 |
PARALLELIZE_DOCSTRING = r"""
|
775 |
This is an experimental feature and is a subject to change at a moment's notice.
|
776 |
|
|
|
778 |
it will evenly distribute blocks across all devices.
|
779 |
|
780 |
Args:
|
781 |
+
device_map (`Dict[int, list]`, optional, defaults to None):
|
782 |
A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
|
783 |
automatically mapped to the first device (for esoteric reasons). That means that the first device should
|
784 |
have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
|
|
|
789 |
- gpt2-large: 36
|
790 |
- gpt2-xl: 48
|
791 |
|
792 |
+
Example:
|
793 |
+
|
794 |
+
```python
|
795 |
+
# Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
|
796 |
+
model = GPT2LMHeadModel.from_pretrained("gpt2-xl")
|
797 |
+
device_map = {
|
798 |
+
0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
|
799 |
+
1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
|
800 |
+
2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
|
801 |
+
3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
|
802 |
+
}
|
803 |
+
model.parallelize(device_map)
|
804 |
+
```
|
805 |
"""
|
806 |
DEPARALLELIZE_DOCSTRING = r"""
|
807 |
Moves the model to cpu from a model parallel state.
|
808 |
|
809 |
+
Example:
|
810 |
+
|
811 |
+
```python
|
812 |
+
# On a 4 GPU machine with gpt2-large:
|
813 |
+
model = GPT2LMHeadModel.from_pretrained("gpt2-large")
|
814 |
+
device_map = {
|
815 |
+
0: [0, 1, 2, 3, 4, 5, 6, 7],
|
816 |
+
1: [8, 9, 10, 11, 12, 13, 14, 15],
|
817 |
+
2: [16, 17, 18, 19, 20, 21, 22, 23],
|
818 |
+
3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
|
819 |
+
}
|
820 |
+
model.parallelize(device_map) # Splits the model across several devices
|
821 |
+
model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
|
822 |
+
```
|
823 |
"""
|
824 |
|
825 |
|
|
|
828 |
GPT2_START_DOCSTRING,
|
829 |
)
|
830 |
class GPT2Model(GPT2PreTrainedModel):
|
831 |
+
_keys_to_ignore_on_load_missing = ["attn.masked_bias"]
|
832 |
+
|
833 |
def __init__(self, config):
|
834 |
super().__init__(config)
|
835 |
|
836 |
+
self.embed_dim = config.hidden_size
|
|
|
|
|
|
|
837 |
|
838 |
+
self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
|
839 |
+
self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
|
840 |
+
self.emb_norm = nn.LayerNorm(
|
841 |
+
config.n_embd, eps=config.layer_norm_epsilon
|
842 |
+
) # Added in GROVER
|
843 |
self.drop = nn.Dropout(config.embd_pdrop)
|
844 |
self.h = nn.ModuleList(
|
845 |
+
[GPT2Block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
|
846 |
)
|
847 |
+
# Removed in GROVER
|
848 |
+
# self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
|
|
|
|
|
849 |
|
850 |
# Model parallel
|
851 |
self.model_parallel = False
|
852 |
self.device_map = None
|
853 |
+
self.gradient_checkpointing = False
|
854 |
+
|
855 |
+
# Initialize weights and apply final processing
|
856 |
+
self.post_init()
|
857 |
|
858 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
859 |
def parallelize(self, device_map=None):
|
|
|
873 |
self.last_device = "cuda:" + str(max(self.device_map.keys()))
|
874 |
self.wte = self.wte.to(self.first_device)
|
875 |
self.wpe = self.wpe.to(self.first_device)
|
876 |
+
|
877 |
+
# Added in GROVER
|
878 |
+
# Wissam: not sure if is fine being on cpu or Better on GPU
|
879 |
+
self.emb_norm = self.emb_norm.to(
|
880 |
+
"cuda:" + str(min(self.device_map.keys()))
|
881 |
+
) # GPU
|
882 |
+
# self.emb_norm = self.emb_norm.to(self.first_device) # CPU
|
883 |
+
|
884 |
# Load onto devices
|
885 |
for k, v in self.device_map.items():
|
886 |
for block in v:
|
887 |
cuda_device = "cuda:" + str(k)
|
888 |
self.h[block] = self.h[block].to(cuda_device)
|
889 |
# ln_f to last
|
890 |
+
# Removed in GROVER
|
891 |
+
# self.ln_f = self.ln_f.to(self.last_device)
|
892 |
|
893 |
@add_start_docstrings(DEPARALLELIZE_DOCSTRING)
|
894 |
def deparallelize(self):
|
|
|
898 |
self.last_device = "cpu"
|
899 |
self.wte = self.wte.to("cpu")
|
900 |
self.wpe = self.wpe.to("cpu")
|
901 |
+
# Added in GROVER
|
902 |
+
self.emb_norm = self.emb_norm.to("cpu")
|
903 |
for index in range(len(self.h)):
|
904 |
self.h[index] = self.h[index].to("cpu")
|
905 |
+
# Removed in GROVER
|
906 |
+
# self.ln_f = self.ln_f.to("cpu")
|
907 |
torch.cuda.empty_cache()
|
908 |
|
909 |
def get_input_embeddings(self):
|
|
|
921 |
|
922 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
923 |
@add_code_sample_docstrings(
|
924 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
925 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
926 |
output_type=BaseModelOutputWithPastAndCrossAttentions,
|
927 |
config_class=_CONFIG_FOR_DOC,
|
928 |
)
|
929 |
def forward(
|
930 |
self,
|
931 |
+
input_ids: Optional[torch.LongTensor] = None,
|
932 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
933 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
934 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
935 |
+
position_ids: Optional[torch.LongTensor] = None,
|
936 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
937 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
938 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
939 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
940 |
+
use_cache: Optional[bool] = None,
|
941 |
+
output_attentions: Optional[bool] = None,
|
942 |
+
output_hidden_states: Optional[bool] = None,
|
943 |
+
return_dict: Optional[bool] = None,
|
944 |
+
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
945 |
output_attentions = (
|
946 |
output_attentions
|
947 |
if output_attentions is not None
|
|
|
971 |
else:
|
972 |
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
973 |
|
974 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
975 |
+
|
976 |
if token_type_ids is not None:
|
977 |
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
978 |
if position_ids is not None:
|
|
|
980 |
|
981 |
if past_key_values is None:
|
982 |
past_length = 0
|
983 |
+
past_key_values = tuple([None] * len(self.h))
|
984 |
else:
|
985 |
past_length = past_key_values[0][0].size(-2)
|
986 |
if position_ids is None:
|
|
|
987 |
position_ids = torch.arange(
|
988 |
past_length,
|
989 |
input_shape[-1] + past_length,
|
|
|
992 |
)
|
993 |
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
|
994 |
|
995 |
+
# GPT2Attention mask.
|
996 |
if attention_mask is not None:
|
997 |
if batch_size <= 0:
|
998 |
raise ValueError("batch_size has to be defined and > 0")
|
|
|
1012 |
attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
|
1013 |
attention_mask = (1.0 - attention_mask) * -10000.0
|
1014 |
|
1015 |
+
# If a 2D or 3D attention mask is provided for the cross-attention
|
1016 |
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
1017 |
if self.config.add_cross_attention and encoder_hidden_states is not None:
|
1018 |
(
|
|
|
1043 |
hidden_states = hidden_states + token_type_embeds
|
1044 |
|
1045 |
hidden_states = self.drop(hidden_states)
|
1046 |
+
# Added in Grover
|
1047 |
+
hidden_states = self.emb_norm(hidden_states)
|
1048 |
+
|
1049 |
output_shape = input_shape + (hidden_states.size(-1),)
|
1050 |
|
1051 |
presents = () if use_cache else None
|
|
|
1069 |
attention_mask = attention_mask.to(hidden_states.device)
|
1070 |
if isinstance(head_mask, torch.Tensor):
|
1071 |
head_mask = head_mask.to(hidden_states.device)
|
|
|
1072 |
if output_hidden_states:
|
1073 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
|
|
1074 |
|
1075 |
+
if self.gradient_checkpointing and self.training:
|
1076 |
+
|
1077 |
+
if use_cache:
|
1078 |
+
logger.warning(
|
1079 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
1080 |
+
)
|
1081 |
+
use_cache = False
|
1082 |
|
1083 |
def create_custom_forward(module):
|
1084 |
def custom_forward(*inputs):
|
1085 |
+
# None for past_key_value
|
1086 |
+
return module(*inputs, use_cache, output_attentions)
|
|
|
|
|
|
|
1087 |
|
1088 |
return custom_forward
|
1089 |
|
1090 |
outputs = torch.utils.checkpoint.checkpoint(
|
1091 |
create_custom_forward(block),
|
1092 |
hidden_states,
|
1093 |
+
None,
|
1094 |
attention_mask,
|
1095 |
head_mask[i],
|
1096 |
encoder_hidden_states,
|
|
|
1108 |
output_attentions=output_attentions,
|
1109 |
)
|
1110 |
|
1111 |
+
hidden_states = outputs[0]
|
1112 |
if use_cache is True:
|
1113 |
+
presents = presents + (outputs[1],)
|
1114 |
|
1115 |
if output_attentions:
|
1116 |
all_self_attentions = all_self_attentions + (
|
|
|
1127 |
if i == v[-1] and "cuda:" + str(k) != self.last_device:
|
1128 |
hidden_states = hidden_states.to("cuda:" + str(k + 1))
|
1129 |
|
1130 |
+
# Removed in Grover
|
1131 |
+
# hidden_states = self.ln_f(hidden_states)
|
1132 |
|
1133 |
+
hidden_states = hidden_states.view(output_shape)
|
1134 |
# Add last hidden state
|
1135 |
if output_hidden_states:
|
1136 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
1165 |
GPT2_START_DOCSTRING,
|
1166 |
)
|
1167 |
class GPT2LMHeadModel(GPT2PreTrainedModel):
|
1168 |
+
_keys_to_ignore_on_load_missing = [
|
1169 |
+
r"attn.masked_bias",
|
1170 |
+
r"attn.bias",
|
1171 |
+
r"lm_head.weight",
|
1172 |
+
]
|
1173 |
|
1174 |
def __init__(self, config):
|
1175 |
super().__init__(config)
|
1176 |
self.transformer = GPT2Model(config)
|
1177 |
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
1178 |
|
|
|
|
|
1179 |
# Model parallel
|
1180 |
self.model_parallel = False
|
1181 |
self.device_map = None
|
1182 |
|
1183 |
+
# Initialize weights and apply final processing
|
1184 |
+
self.post_init()
|
1185 |
+
|
1186 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
1187 |
def parallelize(self, device_map=None):
|
1188 |
self.device_map = (
|
|
|
1206 |
def get_output_embeddings(self):
|
1207 |
return self.lm_head
|
1208 |
|
1209 |
+
def set_output_embeddings(self, new_embeddings):
|
1210 |
+
self.lm_head = new_embeddings
|
1211 |
+
|
1212 |
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
|
1213 |
token_type_ids = kwargs.get("token_type_ids", None)
|
1214 |
# only last token for inputs_ids if past is defined in kwargs
|
|
|
1239 |
|
1240 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
1241 |
@add_code_sample_docstrings(
|
1242 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1243 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
1244 |
output_type=CausalLMOutputWithCrossAttentions,
|
1245 |
config_class=_CONFIG_FOR_DOC,
|
1246 |
)
|
1247 |
def forward(
|
1248 |
self,
|
1249 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1250 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
1251 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1252 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1253 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1254 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1255 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1256 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1257 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
1258 |
+
labels: Optional[torch.LongTensor] = None,
|
1259 |
+
use_cache: Optional[bool] = None,
|
1260 |
+
output_attentions: Optional[bool] = None,
|
1261 |
+
output_hidden_states: Optional[bool] = None,
|
1262 |
+
return_dict: Optional[bool] = None,
|
1263 |
+
) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
|
1264 |
r"""
|
1265 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1266 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
1267 |
+
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
|
1268 |
+
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
|
1269 |
"""
|
1270 |
return_dict = (
|
1271 |
return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
1324 |
past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
1325 |
) -> Tuple[Tuple[torch.Tensor]]:
|
1326 |
"""
|
1327 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
1328 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
1329 |
+
beam_idx at every generation step.
|
1330 |
"""
|
1331 |
return tuple(
|
1332 |
tuple(
|
|
|
1347 |
GPT2_START_DOCSTRING,
|
1348 |
)
|
1349 |
class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
|
1350 |
+
_keys_to_ignore_on_load_missing = [
|
1351 |
+
r"attn.masked_bias",
|
1352 |
+
r"attn.bias",
|
1353 |
+
r"lm_head.weight",
|
1354 |
+
]
|
1355 |
+
|
1356 |
def __init__(self, config):
|
1357 |
super().__init__(config)
|
1358 |
config.num_labels = 1
|
|
|
1360 |
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
1361 |
self.multiple_choice_head = SequenceSummary(config)
|
1362 |
|
|
|
|
|
1363 |
# Model parallel
|
1364 |
self.model_parallel = False
|
1365 |
self.device_map = None
|
1366 |
|
1367 |
+
# Initialize weights and apply final processing
|
1368 |
+
self.post_init()
|
1369 |
+
|
1370 |
@add_start_docstrings(PARALLELIZE_DOCSTRING)
|
1371 |
def parallelize(self, device_map=None):
|
1372 |
self.device_map = (
|
|
|
1394 |
def get_output_embeddings(self):
|
1395 |
return self.lm_head
|
1396 |
|
1397 |
+
def set_output_embeddings(self, new_embeddings):
|
1398 |
+
self.lm_head = new_embeddings
|
1399 |
+
|
1400 |
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
|
1401 |
token_type_ids = kwargs.get("token_type_ids", None)
|
1402 |
# only last token for inputs_ids if past is defined in kwargs
|
|
|
1432 |
)
|
1433 |
def forward(
|
1434 |
self,
|
1435 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1436 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
1437 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1438 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1439 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1440 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1441 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1442 |
+
mc_token_ids: Optional[torch.LongTensor] = None,
|
1443 |
+
labels: Optional[torch.LongTensor] = None,
|
1444 |
+
mc_labels: Optional[torch.LongTensor] = None,
|
1445 |
+
use_cache: Optional[bool] = None,
|
1446 |
+
output_attentions: Optional[bool] = None,
|
1447 |
+
output_hidden_states: Optional[bool] = None,
|
1448 |
+
return_dict: Optional[bool] = None,
|
1449 |
**kwargs,
|
1450 |
+
) -> Union[Tuple, GPT2DoubleHeadsModelOutput]:
|
1451 |
r"""
|
1452 |
+
mc_token_ids (`torch.LongTensor` of shape `(batch_size, num_choices)`, *optional*, default to index of the last token of the input):
|
1453 |
+
Index of the classification token in each input sequence. Selected in the range `[0, input_ids.size(-1) -
|
1454 |
+
1[`.
|
1455 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1456 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
1457 |
+
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size - 1]` All labels set to
|
1458 |
+
`-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size - 1]`
|
1459 |
+
mc_labels (`torch.LongTensor` of shape `(batch_size)`, *optional*):
|
1460 |
+
Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]`
|
1461 |
+
where *num_choices* is the size of the second dimension of the input tensors. (see *input_ids* above)
|
|
|
1462 |
|
1463 |
Return:
|
1464 |
|
1465 |
+
Example:
|
|
|
|
|
|
|
|
|
|
|
|
|
1466 |
|
1467 |
+
```python
|
1468 |
+
>>> import torch
|
1469 |
+
>>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
|
1470 |
|
1471 |
+
>>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
1472 |
+
>>> model = GPT2DoubleHeadsModel.from_pretrained("gpt2")
|
1473 |
|
1474 |
+
>>> # Add a [CLS] to the vocabulary (we should train it also!)
|
1475 |
+
>>> num_added_tokens = tokenizer.add_special_tokens({"cls_token": "[CLS]"})
|
1476 |
+
>>> # Update the model embeddings with the new vocabulary size
|
1477 |
+
>>> embedding_layer = model.resize_token_embeddings(len(tokenizer))
|
1478 |
|
1479 |
+
>>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
|
1480 |
+
>>> encoded_choices = [tokenizer.encode(s) for s in choices]
|
1481 |
+
>>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]
|
1482 |
|
1483 |
+
>>> input_ids = torch.tensor(encoded_choices).unsqueeze(0) # Batch size: 1, number of choices: 2
|
1484 |
+
>>> mc_token_ids = torch.tensor([cls_token_location]) # Batch size: 1
|
|
|
1485 |
|
1486 |
+
>>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
|
1487 |
+
>>> lm_logits = outputs.logits
|
1488 |
+
>>> mc_logits = outputs.mc_logits
|
1489 |
+
```"""
|
1490 |
return_dict = (
|
1491 |
return_dict if return_dict is not None else self.config.use_return_dict
|
1492 |
)
|
|
|
1551 |
past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
1552 |
) -> Tuple[Tuple[torch.Tensor]]:
|
1553 |
"""
|
1554 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
1555 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
1556 |
+
beam_idx at every generation step.
|
1557 |
"""
|
1558 |
return tuple(
|
1559 |
tuple(
|
|
|
1568 |
"""
|
1569 |
The GPT2 Model transformer with a sequence classification head on top (linear layer).
|
1570 |
|
1571 |
+
[`GPT2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
1572 |
+
(e.g. GPT-1) do.
|
1573 |
|
1574 |
Since it does classification on the last token, it requires to know the position of the last token. If a
|
1575 |
+
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
1576 |
+
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
1577 |
+
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
1578 |
+
each row of the batch).
|
1579 |
""",
|
1580 |
GPT2_START_DOCSTRING,
|
1581 |
)
|
|
|
1588 |
self.transformer = GPT2Model(config)
|
1589 |
self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
|
1590 |
|
|
|
|
|
1591 |
# Model parallel
|
1592 |
self.model_parallel = False
|
1593 |
self.device_map = None
|
1594 |
|
1595 |
+
# Initialize weights and apply final processing
|
1596 |
+
self.post_init()
|
1597 |
+
|
1598 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
1599 |
@add_code_sample_docstrings(
|
1600 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1601 |
+
checkpoint="microsoft/DialogRPT-updown",
|
1602 |
output_type=SequenceClassifierOutputWithPast,
|
1603 |
config_class=_CONFIG_FOR_DOC,
|
1604 |
+
expected_output="'LABEL_0'",
|
1605 |
+
expected_loss=5.28,
|
1606 |
)
|
1607 |
def forward(
|
1608 |
self,
|
1609 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1610 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
1611 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1612 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1613 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1614 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1615 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1616 |
+
labels: Optional[torch.LongTensor] = None,
|
1617 |
+
use_cache: Optional[bool] = None,
|
1618 |
+
output_attentions: Optional[bool] = None,
|
1619 |
+
output_hidden_states: Optional[bool] = None,
|
1620 |
+
return_dict: Optional[bool] = None,
|
1621 |
+
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
1622 |
r"""
|
1623 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1624 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1625 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1626 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1627 |
"""
|
1628 |
return_dict = (
|
1629 |
return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
1664 |
sequence_lengths = -1
|
1665 |
logger.warning(
|
1666 |
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
|
1667 |
+
"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
|
1668 |
)
|
1669 |
|
1670 |
+
pooled_logits = logits[
|
1671 |
+
torch.arange(batch_size, device=self.device), sequence_lengths
|
1672 |
+
]
|
1673 |
|
1674 |
loss = None
|
1675 |
if labels is not None:
|
1676 |
+
if self.config.problem_type is None:
|
1677 |
+
if self.num_labels == 1:
|
1678 |
+
self.config.problem_type = "regression"
|
1679 |
+
elif self.num_labels > 1 and (
|
1680 |
+
labels.dtype == torch.long or labels.dtype == torch.int
|
1681 |
+
):
|
1682 |
+
self.config.problem_type = "single_label_classification"
|
1683 |
+
else:
|
1684 |
+
self.config.problem_type = "multi_label_classification"
|
1685 |
+
|
1686 |
+
if self.config.problem_type == "regression":
|
1687 |
loss_fct = MSELoss()
|
1688 |
+
if self.num_labels == 1:
|
1689 |
+
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
1690 |
+
else:
|
1691 |
+
loss = loss_fct(pooled_logits, labels)
|
1692 |
+
elif self.config.problem_type == "single_label_classification":
|
1693 |
loss_fct = CrossEntropyLoss()
|
1694 |
loss = loss_fct(
|
1695 |
pooled_logits.view(-1, self.num_labels), labels.view(-1)
|
1696 |
)
|
1697 |
+
elif self.config.problem_type == "multi_label_classification":
|
1698 |
+
loss_fct = BCEWithLogitsLoss()
|
1699 |
+
loss = loss_fct(pooled_logits, labels)
|
1700 |
if not return_dict:
|
1701 |
output = (pooled_logits,) + transformer_outputs[1:]
|
1702 |
return ((loss,) + output) if loss is not None else output
|
|
|
1735 |
self.dropout = nn.Dropout(classifier_dropout)
|
1736 |
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
1737 |
|
|
|
|
|
1738 |
# Model parallel
|
1739 |
self.model_parallel = False
|
1740 |
self.device_map = None
|
1741 |
|
1742 |
+
# Initialize weights and apply final processing
|
1743 |
+
self.post_init()
|
1744 |
+
|
1745 |
@add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
|
1746 |
+
# fmt: off
|
1747 |
@add_code_sample_docstrings(
|
1748 |
+
processor_class=_TOKENIZER_FOR_DOC,
|
1749 |
+
checkpoint="brad1141/gpt2-finetuned-comp2",
|
1750 |
output_type=TokenClassifierOutput,
|
1751 |
config_class=_CONFIG_FOR_DOC,
|
1752 |
+
expected_loss=0.25,
|
1753 |
+
expected_output=["Lead", "Lead", "Lead", "Position", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead"],
|
1754 |
)
|
1755 |
+
# fmt: on
|
1756 |
def forward(
|
1757 |
self,
|
1758 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1759 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
1760 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1761 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1762 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1763 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1764 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1765 |
+
labels: Optional[torch.LongTensor] = None,
|
1766 |
+
use_cache: Optional[bool] = None,
|
1767 |
+
output_attentions: Optional[bool] = None,
|
1768 |
+
output_hidden_states: Optional[bool] = None,
|
1769 |
+
return_dict: Optional[bool] = None,
|
1770 |
+
) -> Union[Tuple, TokenClassifierOutput]:
|
1771 |
r"""
|
1772 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1773 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1774 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1775 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1776 |
"""
|
1777 |
return_dict = (
|
1778 |
return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
1799 |
loss = None
|
1800 |
if labels is not None:
|
1801 |
loss_fct = CrossEntropyLoss()
|
1802 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1803 |
|
1804 |
if not return_dict:
|
1805 |
output = (logits,) + transformer_outputs[2:]
|
|
|
1810 |
logits=logits,
|
1811 |
hidden_states=transformer_outputs.hidden_states,
|
1812 |
attentions=transformer_outputs.attentions,
|
1813 |
+
)
|