Implement batch generation
Browse files- modeling_chatglm.py +100 -83
- tokenization_chatglm.py +103 -12
modeling_chatglm.py
CHANGED
@@ -13,7 +13,7 @@ import torch.nn.functional as F
|
|
13 |
from torch import nn
|
14 |
from torch.nn import CrossEntropyLoss, LayerNorm
|
15 |
from torch.nn.utils import skip_init
|
16 |
-
from typing import Optional, Tuple, Union, List, Callable
|
17 |
|
18 |
from transformers.utils import (
|
19 |
add_code_sample_docstrings,
|
@@ -28,7 +28,7 @@ from transformers.modeling_outputs import (
|
|
28 |
from transformers.modeling_utils import PreTrainedModel
|
29 |
from transformers.utils import logging
|
30 |
from transformers.generation.logits_process import LogitsProcessor
|
31 |
-
from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig
|
32 |
|
33 |
from .configuration_chatglm import ChatGLMConfig
|
34 |
|
@@ -664,6 +664,39 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
|
|
664 |
"""Initialize the weights."""
|
665 |
return
|
666 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
667 |
def _set_gradient_checkpointing(self, module, value=False):
|
668 |
if isinstance(module, ChatGLMModel):
|
669 |
module.gradient_checkpointing = value
|
@@ -828,39 +861,6 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
828 |
# past_key_values = [(v[0], v[1]) for v in past_key_values]
|
829 |
return past_key_values
|
830 |
|
831 |
-
def get_masks(self, input_ids, device):
|
832 |
-
batch_size, seq_length = input_ids.shape
|
833 |
-
context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
|
834 |
-
attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
|
835 |
-
attention_mask.tril_()
|
836 |
-
for i, context_length in enumerate(context_lengths):
|
837 |
-
attention_mask[i, :, :context_length] = 1
|
838 |
-
attention_mask.unsqueeze_(1)
|
839 |
-
attention_mask = (attention_mask < 0.5).bool()
|
840 |
-
|
841 |
-
return attention_mask
|
842 |
-
|
843 |
-
def get_position_ids(self, input_ids, mask_positions, device, gmask=False):
|
844 |
-
batch_size, seq_length = input_ids.shape
|
845 |
-
context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
|
846 |
-
if self.position_encoding_2d:
|
847 |
-
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).expand(batch_size, seq_length)
|
848 |
-
for i, context_length in enumerate(context_lengths):
|
849 |
-
position_ids[i, context_length:] = mask_positions[i]
|
850 |
-
block_position_ids = [torch.cat((
|
851 |
-
torch.zeros(context_length, dtype=torch.long, device=device),
|
852 |
-
torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
|
853 |
-
)) for context_length in context_lengths]
|
854 |
-
block_position_ids = torch.stack(block_position_ids, dim=0)
|
855 |
-
position_ids = torch.stack((position_ids, block_position_ids), dim=1)
|
856 |
-
else:
|
857 |
-
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).expand(batch_size, seq_length)
|
858 |
-
if not gmask:
|
859 |
-
for i, context_length in enumerate(context_lengths):
|
860 |
-
position_ids[context_length:] = mask_positions[i]
|
861 |
-
|
862 |
-
return position_ids
|
863 |
-
|
864 |
@add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
865 |
@add_code_sample_docstrings(
|
866 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
@@ -1038,35 +1038,39 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1038 |
def set_output_embeddings(self, new_embeddings):
|
1039 |
self.lm_head = new_embeddings
|
1040 |
|
1041 |
-
def
|
1042 |
-
|
1043 |
-
|
1044 |
-
|
1045 |
-
|
1046 |
-
|
1047 |
-
|
1048 |
-
|
1049 |
-
|
|
|
|
|
1050 |
|
1051 |
-
|
1052 |
-
|
1053 |
-
|
1054 |
-
|
1055 |
-
|
1056 |
-
|
1057 |
-
|
1058 |
-
|
1059 |
-
|
1060 |
-
)
|
1061 |
-
block_position_ids = torch.stack(block_position_ids, dim=0)
|
1062 |
-
position_ids = torch.stack((position_ids, block_position_ids), dim=1)
|
1063 |
-
else:
|
1064 |
-
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).expand(batch_size, seq_length)
|
1065 |
-
if not gmask:
|
1066 |
-
for i, context_length in enumerate(context_lengths):
|
1067 |
-
position_ids[context_length:] = mask_positions[i]
|
1068 |
|
1069 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1070 |
|
1071 |
def prepare_inputs_for_generation(
|
1072 |
self,
|
@@ -1074,6 +1078,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1074 |
past: Optional[torch.Tensor] = None,
|
1075 |
past_key_values: Optional[torch.Tensor] = None,
|
1076 |
attention_mask: Optional[torch.Tensor] = None,
|
|
|
1077 |
**kwargs
|
1078 |
) -> dict:
|
1079 |
batch_size, seq_length = input_ids.shape
|
@@ -1085,15 +1090,20 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1085 |
|
1086 |
# only last token for input_ids if past is not None
|
1087 |
if past is not None or past_key_values is not None:
|
1088 |
-
context_lengths = [seq.index(self.config.bos_token_id) for seq in seqs]
|
1089 |
last_token = input_ids[:, -1].unsqueeze(-1)
|
1090 |
-
if
|
1091 |
-
|
1092 |
-
|
1093 |
-
|
1094 |
else:
|
1095 |
-
|
1096 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1097 |
|
1098 |
if past is None:
|
1099 |
past = past_key_values
|
@@ -1101,14 +1111,21 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1101 |
"input_ids": last_token,
|
1102 |
"past_key_values": past,
|
1103 |
"position_ids": position_ids,
|
|
|
1104 |
}
|
1105 |
else:
|
1106 |
-
attention_mask
|
1107 |
-
|
1108 |
-
|
1109 |
-
|
1110 |
-
|
1111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1112 |
|
1113 |
return {
|
1114 |
"input_ids": input_ids,
|
@@ -1226,10 +1243,10 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1226 |
for i, (old_query, response) in enumerate(history):
|
1227 |
prompt += "[Round {}]\n问:{}\n答:{}\n".format(i, old_query, response)
|
1228 |
prompt += "[Round {}]\n问:{}\n答:".format(len(history), query)
|
1229 |
-
|
1230 |
-
|
1231 |
-
outputs = self.generate(**
|
1232 |
-
outputs = outputs.tolist()[0][len(
|
1233 |
response = tokenizer.decode(outputs)
|
1234 |
response = self.process_response(response)
|
1235 |
history = history + [(query, response)]
|
@@ -1252,10 +1269,10 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
|
1252 |
for i, (old_query, response) in enumerate(history):
|
1253 |
prompt += "[Round {}]\n问:{}\n答:{}\n".format(i, old_query, response)
|
1254 |
prompt += "[Round {}]\n问:{}\n答:".format(len(history), query)
|
1255 |
-
|
1256 |
-
|
1257 |
-
for outputs in self.stream_generate(**
|
1258 |
-
outputs = outputs.tolist()[0][len(
|
1259 |
response = tokenizer.decode(outputs)
|
1260 |
response = self.process_response(response)
|
1261 |
new_history = history + [(query, response)]
|
|
|
13 |
from torch import nn
|
14 |
from torch.nn import CrossEntropyLoss, LayerNorm
|
15 |
from torch.nn.utils import skip_init
|
16 |
+
from typing import Optional, Tuple, Union, List, Callable, Dict, Any
|
17 |
|
18 |
from transformers.utils import (
|
19 |
add_code_sample_docstrings,
|
|
|
28 |
from transformers.modeling_utils import PreTrainedModel
|
29 |
from transformers.utils import logging
|
30 |
from transformers.generation.logits_process import LogitsProcessor
|
31 |
+
from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
|
32 |
|
33 |
from .configuration_chatglm import ChatGLMConfig
|
34 |
|
|
|
664 |
"""Initialize the weights."""
|
665 |
return
|
666 |
|
667 |
+
def get_masks(self, input_ids, device):
|
668 |
+
batch_size, seq_length = input_ids.shape
|
669 |
+
context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
|
670 |
+
attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
|
671 |
+
attention_mask.tril_()
|
672 |
+
for i, context_length in enumerate(context_lengths):
|
673 |
+
attention_mask[i, :, :context_length] = 1
|
674 |
+
attention_mask.unsqueeze_(1)
|
675 |
+
attention_mask = (attention_mask < 0.5).bool()
|
676 |
+
|
677 |
+
return attention_mask
|
678 |
+
|
679 |
+
def get_position_ids(self, input_ids, mask_positions, device, gmask=False):
|
680 |
+
batch_size, seq_length = input_ids.shape
|
681 |
+
context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
|
682 |
+
if self.position_encoding_2d:
|
683 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).expand(batch_size, seq_length)
|
684 |
+
for i, context_length in enumerate(context_lengths):
|
685 |
+
position_ids[i, context_length:] = mask_positions[i]
|
686 |
+
block_position_ids = [torch.cat((
|
687 |
+
torch.zeros(context_length, dtype=torch.long, device=device),
|
688 |
+
torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
|
689 |
+
)) for context_length in context_lengths]
|
690 |
+
block_position_ids = torch.stack(block_position_ids, dim=0)
|
691 |
+
position_ids = torch.stack((position_ids, block_position_ids), dim=1)
|
692 |
+
else:
|
693 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).expand(batch_size, seq_length)
|
694 |
+
if not gmask:
|
695 |
+
for i, context_length in enumerate(context_lengths):
|
696 |
+
position_ids[context_length:] = mask_positions[i]
|
697 |
+
|
698 |
+
return position_ids
|
699 |
+
|
700 |
def _set_gradient_checkpointing(self, module, value=False):
|
701 |
if isinstance(module, ChatGLMModel):
|
702 |
module.gradient_checkpointing = value
|
|
|
861 |
# past_key_values = [(v[0], v[1]) for v in past_key_values]
|
862 |
return past_key_values
|
863 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
864 |
@add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
|
865 |
@add_code_sample_docstrings(
|
866 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
|
|
1038 |
def set_output_embeddings(self, new_embeddings):
|
1039 |
self.lm_head = new_embeddings
|
1040 |
|
1041 |
+
def _update_model_kwargs_for_generation(
|
1042 |
+
self,
|
1043 |
+
outputs: ModelOutput,
|
1044 |
+
model_kwargs: Dict[str, Any],
|
1045 |
+
is_encoder_decoder: bool = False,
|
1046 |
+
standardize_cache_format: bool = False,
|
1047 |
+
) -> Dict[str, Any]:
|
1048 |
+
# update past_key_values
|
1049 |
+
model_kwargs["past_key_values"] = self._extract_past_from_model_output(
|
1050 |
+
outputs, standardize_cache_format=standardize_cache_format
|
1051 |
+
)
|
1052 |
|
1053 |
+
# update attention mask
|
1054 |
+
if "attention_mask" in model_kwargs:
|
1055 |
+
attention_mask = model_kwargs["attention_mask"]
|
1056 |
+
attention_mask = torch.cat(
|
1057 |
+
[attention_mask, attention_mask.new_ones((*attention_mask.shape[:3], 1))], dim=3)
|
1058 |
+
new_attention_mask = attention_mask[:, :, -1:].clone()
|
1059 |
+
new_attention_mask[..., -1] = False
|
1060 |
+
model_kwargs["attention_mask"] = torch.cat(
|
1061 |
+
[attention_mask, new_attention_mask], dim=2
|
1062 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1063 |
|
1064 |
+
# update position ids
|
1065 |
+
if "position_ids" in model_kwargs:
|
1066 |
+
position_ids = model_kwargs["position_ids"]
|
1067 |
+
new_position_id = position_ids[..., -1:].clone()
|
1068 |
+
new_position_id[:, 1, :] += 1
|
1069 |
+
model_kwargs["position_ids"] = torch.cat(
|
1070 |
+
[position_ids, new_position_id], dim=-1
|
1071 |
+
)
|
1072 |
+
|
1073 |
+
return model_kwargs
|
1074 |
|
1075 |
def prepare_inputs_for_generation(
|
1076 |
self,
|
|
|
1078 |
past: Optional[torch.Tensor] = None,
|
1079 |
past_key_values: Optional[torch.Tensor] = None,
|
1080 |
attention_mask: Optional[torch.Tensor] = None,
|
1081 |
+
position_ids: Optional[torch.Tensor] = None,
|
1082 |
**kwargs
|
1083 |
) -> dict:
|
1084 |
batch_size, seq_length = input_ids.shape
|
|
|
1090 |
|
1091 |
# only last token for input_ids if past is not None
|
1092 |
if past is not None or past_key_values is not None:
|
|
|
1093 |
last_token = input_ids[:, -1].unsqueeze(-1)
|
1094 |
+
if attention_mask is not None:
|
1095 |
+
attention_mask = attention_mask[:, :, -1:]
|
1096 |
+
if position_ids is not None:
|
1097 |
+
position_ids = position_ids[..., -1:]
|
1098 |
else:
|
1099 |
+
context_lengths = [seq.index(self.config.bos_token_id) for seq in seqs]
|
1100 |
+
if self.position_encoding_2d:
|
1101 |
+
position_ids = torch.tensor(
|
1102 |
+
[[mask_position, seq_length - context_length] for mask_position, context_length in
|
1103 |
+
zip(mask_positions, context_lengths)], dtype=torch.long, device=input_ids.device).unsqueeze(-1)
|
1104 |
+
else:
|
1105 |
+
position_ids = torch.tensor([mask_position for mask_position in mask_positions], dtype=torch.long,
|
1106 |
+
device=input_ids.device).unsqueeze(-1)
|
1107 |
|
1108 |
if past is None:
|
1109 |
past = past_key_values
|
|
|
1111 |
"input_ids": last_token,
|
1112 |
"past_key_values": past,
|
1113 |
"position_ids": position_ids,
|
1114 |
+
"attention_mask": attention_mask
|
1115 |
}
|
1116 |
else:
|
1117 |
+
if attention_mask is None:
|
1118 |
+
attention_mask = self.get_masks(
|
1119 |
+
input_ids,
|
1120 |
+
device=input_ids.device
|
1121 |
+
)
|
1122 |
+
if position_ids is None:
|
1123 |
+
position_ids = self.get_position_ids(
|
1124 |
+
input_ids,
|
1125 |
+
device=input_ids.device,
|
1126 |
+
mask_positions=mask_positions,
|
1127 |
+
gmask=use_gmask
|
1128 |
+
)
|
1129 |
|
1130 |
return {
|
1131 |
"input_ids": input_ids,
|
|
|
1243 |
for i, (old_query, response) in enumerate(history):
|
1244 |
prompt += "[Round {}]\n问:{}\n答:{}\n".format(i, old_query, response)
|
1245 |
prompt += "[Round {}]\n问:{}\n答:".format(len(history), query)
|
1246 |
+
inputs = tokenizer([prompt], return_tensors="pt", padding=True)
|
1247 |
+
inputs = inputs.to(self.device)
|
1248 |
+
outputs = self.generate(**inputs, **gen_kwargs)
|
1249 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
|
1250 |
response = tokenizer.decode(outputs)
|
1251 |
response = self.process_response(response)
|
1252 |
history = history + [(query, response)]
|
|
|
1269 |
for i, (old_query, response) in enumerate(history):
|
1270 |
prompt += "[Round {}]\n问:{}\n答:{}\n".format(i, old_query, response)
|
1271 |
prompt += "[Round {}]\n问:{}\n答:".format(len(history), query)
|
1272 |
+
inputs = tokenizer([prompt], return_tensors="pt", padding=True)
|
1273 |
+
inputs = inputs.to(self.device)
|
1274 |
+
for outputs in self.stream_generate(**inputs, **gen_kwargs):
|
1275 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
|
1276 |
response = tokenizer.decode(outputs)
|
1277 |
response = self.process_response(response)
|
1278 |
new_history = history + [(query, response)]
|
tokenization_chatglm.py
CHANGED
@@ -1,17 +1,14 @@
|
|
1 |
"""Tokenization classes for ChatGLM."""
|
2 |
-
import sys
|
3 |
-
import unicodedata
|
4 |
from typing import List, Optional, Union
|
5 |
-
from functools import lru_cache
|
6 |
import os
|
7 |
-
import collections
|
8 |
-
import re
|
9 |
|
10 |
from transformers.tokenization_utils import PreTrainedTokenizer
|
11 |
from icetk.text_tokenizer import TextTokenizer
|
12 |
-
from icetk.utils import auto_create
|
13 |
import icetk.sentencepiece_model_pb2 as sp_model
|
14 |
-
from transformers.utils import logging
|
|
|
|
|
|
|
15 |
|
16 |
logger = logging.get_logger(__name__)
|
17 |
|
@@ -192,7 +189,7 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
|
|
192 |
eop_token='eop',
|
193 |
mask_token='[MASK]',
|
194 |
gmask_token='[gMASK]',
|
195 |
-
padding_side="
|
196 |
**kwargs
|
197 |
) -> None:
|
198 |
super().__init__(
|
@@ -210,7 +207,7 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
|
|
210 |
self.eos_token = eos_token
|
211 |
self.eop_token = eop_token
|
212 |
self.mask_token = mask_token
|
213 |
-
self.
|
214 |
|
215 |
self.sp_tokenizer = SPTokenizer(vocab_file)
|
216 |
|
@@ -331,10 +328,9 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
|
|
331 |
Returns:
|
332 |
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
|
333 |
"""
|
334 |
-
if token_ids_1 is not None:
|
335 |
-
token_ids_0 += token_ids_1
|
336 |
mask_ids = self.sp_tokenizer[self.mask_token]
|
337 |
-
gmask_ids = self.sp_tokenizer[self.
|
|
|
338 |
if mask_ids not in token_ids_0 and gmask_ids not in token_ids_0:
|
339 |
token_ids_0 += [gmask_ids]
|
340 |
|
@@ -343,4 +339,99 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
|
|
343 |
|
344 |
token_ids_0 += [self.sp_tokenizer[self.bos_token]]
|
345 |
|
|
|
|
|
|
|
|
|
|
|
346 |
return token_ids_0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
"""Tokenization classes for ChatGLM."""
|
|
|
|
|
2 |
from typing import List, Optional, Union
|
|
|
3 |
import os
|
|
|
|
|
4 |
|
5 |
from transformers.tokenization_utils import PreTrainedTokenizer
|
6 |
from icetk.text_tokenizer import TextTokenizer
|
|
|
7 |
import icetk.sentencepiece_model_pb2 as sp_model
|
8 |
+
from transformers.utils import logging, PaddingStrategy
|
9 |
+
from transformers.tokenization_utils_base import EncodedInput, BatchEncoding
|
10 |
+
from typing import Dict
|
11 |
+
import numpy as np
|
12 |
|
13 |
logger = logging.get_logger(__name__)
|
14 |
|
|
|
189 |
eop_token='eop',
|
190 |
mask_token='[MASK]',
|
191 |
gmask_token='[gMASK]',
|
192 |
+
padding_side="left",
|
193 |
**kwargs
|
194 |
) -> None:
|
195 |
super().__init__(
|
|
|
207 |
self.eos_token = eos_token
|
208 |
self.eop_token = eop_token
|
209 |
self.mask_token = mask_token
|
210 |
+
self.gmask_token = gmask_token
|
211 |
|
212 |
self.sp_tokenizer = SPTokenizer(vocab_file)
|
213 |
|
|
|
328 |
Returns:
|
329 |
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
|
330 |
"""
|
|
|
|
|
331 |
mask_ids = self.sp_tokenizer[self.mask_token]
|
332 |
+
gmask_ids = self.sp_tokenizer[self.gmask_token]
|
333 |
+
eop_id = self.sp_tokenizer[self.eop_token]
|
334 |
if mask_ids not in token_ids_0 and gmask_ids not in token_ids_0:
|
335 |
token_ids_0 += [gmask_ids]
|
336 |
|
|
|
339 |
|
340 |
token_ids_0 += [self.sp_tokenizer[self.bos_token]]
|
341 |
|
342 |
+
if token_ids_1 is not None:
|
343 |
+
if token_ids_1[-1] != eop_id:
|
344 |
+
token_ids_1 += [eop_id]
|
345 |
+
token_ids_0 += token_ids_1
|
346 |
+
|
347 |
return token_ids_0
|
348 |
+
|
349 |
+
def _pad(
|
350 |
+
self,
|
351 |
+
encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
|
352 |
+
max_length: Optional[int] = None,
|
353 |
+
padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
|
354 |
+
pad_to_multiple_of: Optional[int] = None,
|
355 |
+
return_attention_mask: Optional[bool] = None,
|
356 |
+
) -> dict:
|
357 |
+
"""
|
358 |
+
Pad encoded inputs (on left/right and up to predefined length or max length in the batch)
|
359 |
+
|
360 |
+
Args:
|
361 |
+
encoded_inputs:
|
362 |
+
Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).
|
363 |
+
max_length: maximum length of the returned list and optionally padding length (see below).
|
364 |
+
Will truncate by taking into account the special tokens.
|
365 |
+
padding_strategy: PaddingStrategy to use for padding.
|
366 |
+
|
367 |
+
- PaddingStrategy.LONGEST Pad to the longest sequence in the batch
|
368 |
+
- PaddingStrategy.MAX_LENGTH: Pad to the max length (default)
|
369 |
+
- PaddingStrategy.DO_NOT_PAD: Do not pad
|
370 |
+
The tokenizer padding sides are defined in self.padding_side:
|
371 |
+
|
372 |
+
- 'left': pads on the left of the sequences
|
373 |
+
- 'right': pads on the right of the sequences
|
374 |
+
pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.
|
375 |
+
This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability
|
376 |
+
`>= 7.5` (Volta).
|
377 |
+
return_attention_mask:
|
378 |
+
(optional) Set to False to avoid returning attention mask (default: set to model specifics)
|
379 |
+
"""
|
380 |
+
# Load from model defaults
|
381 |
+
bos_token_id = self.sp_tokenizer[self.bos_token]
|
382 |
+
mask_token_id = self.sp_tokenizer[self.mask_token]
|
383 |
+
gmask_token_id = self.sp_tokenizer[self.gmask_token]
|
384 |
+
assert self.padding_side == "left"
|
385 |
+
if return_attention_mask is None:
|
386 |
+
return_attention_mask = "attention_mask" in self.model_input_names
|
387 |
+
|
388 |
+
required_input = encoded_inputs[self.model_input_names[0]]
|
389 |
+
seq_length = len(required_input)
|
390 |
+
|
391 |
+
if padding_strategy == PaddingStrategy.LONGEST:
|
392 |
+
max_length = len(required_input)
|
393 |
+
|
394 |
+
if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
|
395 |
+
max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
|
396 |
+
|
397 |
+
needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
|
398 |
+
|
399 |
+
# Initialize attention mask if not present.
|
400 |
+
if needs_to_be_padded or return_attention_mask:
|
401 |
+
context_length = required_input.index(bos_token_id)
|
402 |
+
attention_mask = np.ones((1, seq_length, seq_length))
|
403 |
+
attention_mask = np.tril(attention_mask)
|
404 |
+
attention_mask[:, :, :context_length] = 1
|
405 |
+
attention_mask = np.bool_(attention_mask < 0.5)
|
406 |
+
encoded_inputs["attention_mask"] = attention_mask
|
407 |
+
|
408 |
+
if needs_to_be_padded or return_attention_mask:
|
409 |
+
mask_token = mask_token_id if mask_token_id in required_input else gmask_token_id
|
410 |
+
mask_position = required_input.index(mask_token)
|
411 |
+
context_length = required_input.index(bos_token_id)
|
412 |
+
position_ids = np.arange(seq_length, dtype=np.int64)
|
413 |
+
position_ids[context_length:] = mask_position
|
414 |
+
block_position_ids = np.concatenate(
|
415 |
+
[np.zeros(context_length, dtype=np.int64),
|
416 |
+
np.arange(1, seq_length - context_length + 1, dtype=np.int64)])
|
417 |
+
encoded_inputs["position_ids"] = np.stack([position_ids, block_position_ids], axis=0)
|
418 |
+
|
419 |
+
if needs_to_be_padded:
|
420 |
+
difference = max_length - len(required_input)
|
421 |
+
|
422 |
+
if "attention_mask" in encoded_inputs:
|
423 |
+
encoded_inputs["attention_mask"] = np.pad(encoded_inputs["attention_mask"],
|
424 |
+
pad_width=[(0, 0), (difference, 0), (difference, 0)],
|
425 |
+
mode='constant', constant_values=True)
|
426 |
+
if "token_type_ids" in encoded_inputs:
|
427 |
+
encoded_inputs["token_type_ids"] = [self.pad_token_type_id] * difference + encoded_inputs[
|
428 |
+
"token_type_ids"
|
429 |
+
]
|
430 |
+
if "special_tokens_mask" in encoded_inputs:
|
431 |
+
encoded_inputs["special_tokens_mask"] = [1] * difference + encoded_inputs["special_tokens_mask"]
|
432 |
+
if "position_ids" in encoded_inputs:
|
433 |
+
encoded_inputs["position_ids"] = np.pad(encoded_inputs["position_ids"],
|
434 |
+
pad_width=[(0, 0), (difference, 0)])
|
435 |
+
encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input
|
436 |
+
|
437 |
+
return encoded_inputs
|