jacobrenn commited on
Commit
ebeb179
1 Parent(s): ed7947a

Update instruct_pipeline.py

Browse files
Files changed (1) hide show
  1. instruct_pipeline.py +102 -50
instruct_pipeline.py CHANGED
@@ -1,8 +1,16 @@
 
1
  import re
 
2
 
3
  import numpy as np
4
  from transformers import Pipeline, PreTrainedTokenizer
5
 
 
 
 
 
 
 
6
 
7
  INSTRUCTION_KEY = "### Instruction:"
8
  RESPONSE_KEY = "### Response:"
@@ -53,9 +61,22 @@ class InstructionTextGenerationPipeline(Pipeline):
53
  def __init__(
54
  self, *args, do_sample: bool = True, max_new_tokens: int = 256, top_p: float = 0.92, top_k: int = 0, **kwargs
55
  ):
56
- super().__init__(*args, do_sample=do_sample, max_new_tokens=max_new_tokens, top_p=top_p, top_k=top_k, **kwargs)
57
-
58
- def _sanitize_parameters(self, return_instruction_text=False, **generate_kwargs):
 
 
 
 
 
 
 
 
 
 
 
 
 
59
  preprocess_params = {}
60
 
61
  # newer versions of the tokenizer configure the response key as a special token. newer versions still may
@@ -79,10 +100,12 @@ class InstructionTextGenerationPipeline(Pipeline):
79
  forward_params = generate_kwargs
80
  postprocess_params = {
81
  "response_key_token_id": response_key_token_id,
82
- "end_key_token_id": end_key_token_id,
83
- "return_instruction_text": return_instruction_text,
84
  }
85
 
 
 
 
86
  return preprocess_params, forward_params, postprocess_params
87
 
88
  def preprocess(self, instruction_text, **generate_kwargs):
@@ -98,63 +121,92 @@ class InstructionTextGenerationPipeline(Pipeline):
98
  def _forward(self, model_inputs, **generate_kwargs):
99
  input_ids = model_inputs["input_ids"]
100
  attention_mask = model_inputs.get("attention_mask", None)
 
 
 
 
 
 
 
 
101
  generated_sequence = self.model.generate(
102
  input_ids=input_ids.to(self.model.device),
103
- attention_mask=attention_mask,
104
  pad_token_id=self.tokenizer.pad_token_id,
105
  **generate_kwargs,
106
- )[0].cpu()
 
 
 
 
 
 
 
107
  instruction_text = model_inputs.pop("instruction_text")
108
  return {"generated_sequence": generated_sequence, "input_ids": input_ids, "instruction_text": instruction_text}
109
 
110
- def postprocess(self, model_outputs, response_key_token_id, end_key_token_id, return_instruction_text):
111
- sequence = model_outputs["generated_sequence"]
 
112
  instruction_text = model_outputs["instruction_text"]
113
 
114
- # The response will be set to this variable if we can identify it.
115
- decoded = None
 
116
 
117
- # If we have token IDs for the response and end, then we can find the tokens and only decode between them.
118
- if response_key_token_id and end_key_token_id:
119
- # Find where "### Response:" is first found in the generated tokens. Considering this is part of the
120
- # prompt, we should definitely find it. We will return the tokens found after this token.
121
- response_pos = None
122
- response_positions = np.where(sequence == response_key_token_id)[0]
123
- if len(response_positions) == 0:
124
- pass
125
- else:
126
- response_pos = response_positions[0]
127
-
128
- if response_pos:
129
- # Next find where "### End" is located. The model has been trained to end its responses with this
130
- # sequence (or actually, the token ID it maps to, since it is a special token). We may not find
131
- # this token, as the response could be truncated. If we don't find it then just return everything
132
- # to the end. Note that even though we set eos_token_id, we still see the this token at the end.
133
- end_pos = None
134
- end_positions = np.where(sequence == end_key_token_id)[0]
135
- if len(end_positions) > 0:
136
- end_pos = end_positions[0]
137
-
138
- decoded = self.tokenizer.decode(sequence[response_pos + 1 : end_pos]).strip()
139
- else:
140
- # Otherwise we'll decode everything and use a regex to find the response and end.
141
 
142
- fully_decoded = self.tokenizer.decode(sequence)
 
 
 
 
 
 
 
 
143
 
144
- # The response appears after "### Response:". The model has been trained to append "### End" at the
145
- # end.
146
- m = re.search(r"#+\s*Response:\s*(.+?)#+\s*End", fully_decoded, flags=re.DOTALL)
 
 
 
 
 
 
147
 
148
- if m:
149
- decoded = m.group(1).strip()
150
- else:
151
- # The model might not generate the "### End" sequence before reaching the max tokens. In this case,
152
- # return everything after "### Response:".
153
- m = re.search(r"#+\s*Response:\s*(.+)", fully_decoded, flags=re.DOTALL)
154
- if m:
155
- decoded = m.group(1).strip()
156
 
157
- if return_instruction_text:
158
- return {"instruction_text": instruction_text, "generated_text": decoded}
159
 
160
- return {'generated_text': decoded}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
  import re
3
+ from typing import List
4
 
5
  import numpy as np
6
  from transformers import Pipeline, PreTrainedTokenizer
7
 
8
+ from transformers.utils import is_tf_available
9
+
10
+ if is_tf_available():
11
+ import tensorflow as tf
12
+
13
+ logger = logging.getLogger(__name__)
14
 
15
  INSTRUCTION_KEY = "### Instruction:"
16
  RESPONSE_KEY = "### Response:"
 
61
  def __init__(
62
  self, *args, do_sample: bool = True, max_new_tokens: int = 256, top_p: float = 0.92, top_k: int = 0, **kwargs
63
  ):
64
+ """Initialize the pipeline
65
+
66
+ Args:
67
+ do_sample (bool, optional): Whether or not to use sampling. Defaults to True.
68
+ max_new_tokens (int, optional): Max new tokens after the prompt to generate. Defaults to 128.
69
+ top_p (float, optional): If set to float < 1, only the smallest set of most probable tokens with
70
+ probabilities that add up to top_p or higher are kept for generation. Defaults to 0.92.
71
+ top_k (int, optional): The number of highest probability vocabulary tokens to keep for top-k-filtering.
72
+ Defaults to 0.
73
+ """
74
+ super().__init__(*args, do_sample=do_sample, max_new_tokens=max_new_tokens, top_p=top_p, top_k=top_k,
75
+ **kwargs)
76
+
77
+ def _sanitize_parameters(self,
78
+ return_full_text: bool = None,
79
+ **generate_kwargs):
80
  preprocess_params = {}
81
 
82
  # newer versions of the tokenizer configure the response key as a special token. newer versions still may
 
100
  forward_params = generate_kwargs
101
  postprocess_params = {
102
  "response_key_token_id": response_key_token_id,
103
+ "end_key_token_id": end_key_token_id
 
104
  }
105
 
106
+ if return_full_text is not None:
107
+ postprocess_params["return_full_text"] = return_full_text
108
+
109
  return preprocess_params, forward_params, postprocess_params
110
 
111
  def preprocess(self, instruction_text, **generate_kwargs):
 
121
  def _forward(self, model_inputs, **generate_kwargs):
122
  input_ids = model_inputs["input_ids"]
123
  attention_mask = model_inputs.get("attention_mask", None)
124
+
125
+ if input_ids.shape[1] == 0:
126
+ input_ids = None
127
+ attention_mask = None
128
+ in_b = 1
129
+ else:
130
+ in_b = input_ids.shape[0]
131
+
132
  generated_sequence = self.model.generate(
133
  input_ids=input_ids.to(self.model.device),
134
+ attention_mask=attention_mask.to(self.model.device) if attention_mask is not None else None,
135
  pad_token_id=self.tokenizer.pad_token_id,
136
  **generate_kwargs,
137
+ )
138
+
139
+ out_b = generated_sequence.shape[0]
140
+ if self.framework == "pt":
141
+ generated_sequence = generated_sequence.reshape(in_b, out_b // in_b, *generated_sequence.shape[1:])
142
+ elif self.framework == "tf":
143
+ generated_sequence = tf.reshape(generated_sequence, (in_b, out_b // in_b, *generated_sequence.shape[1:]))
144
+
145
  instruction_text = model_inputs.pop("instruction_text")
146
  return {"generated_sequence": generated_sequence, "input_ids": input_ids, "instruction_text": instruction_text}
147
 
148
+ def postprocess(self, model_outputs, response_key_token_id, end_key_token_id, return_full_text: bool = False):
149
+
150
+ generated_sequence = model_outputs["generated_sequence"][0]
151
  instruction_text = model_outputs["instruction_text"]
152
 
153
+ generated_sequence: List[List[int]] = generated_sequence.numpy().tolist()
154
+ records = []
155
+ for sequence in generated_sequence:
156
 
157
+ # The response will be set to this variable if we can identify it.
158
+ decoded = None
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
159
 
160
+ # If we have token IDs for the response and end, then we can find the tokens and only decode between them.
161
+ if response_key_token_id and end_key_token_id:
162
+ # Find where "### Response:" is first found in the generated tokens. Considering this is part of the
163
+ # prompt, we should definitely find it. We will return the tokens found after this token.
164
+ try:
165
+ response_pos = sequence.index(response_key_token_id)
166
+ except ValueError:
167
+ logger.warn(f"Could not find response key {response_key_token_id} in: {sequence}")
168
+ response_pos = None
169
 
170
+ if response_pos:
171
+ # Next find where "### End" is located. The model has been trained to end its responses with this
172
+ # sequence (or actually, the token ID it maps to, since it is a special token). We may not find
173
+ # this token, as the response could be truncated. If we don't find it then just return everything
174
+ # to the end. Note that even though we set eos_token_id, we still see the this token at the end.
175
+ try:
176
+ end_pos = sequence.index(end_key_token_id)
177
+ except ValueError:
178
+ end_pos = None
179
 
180
+ decoded = self.tokenizer.decode(sequence[response_pos + 1 : end_pos]).strip()
 
 
 
 
 
 
 
181
 
182
+ if not decoded:
183
+ # Otherwise we'll decode everything and use a regex to find the response and end.
184
 
185
+ fully_decoded = self.tokenizer.decode(sequence)
186
+
187
+ # The response appears after "### Response:". The model has been trained to append "### End" at the
188
+ # end.
189
+ m = re.search(r"#+\s*Response:\s*(.+?)#+\s*End", fully_decoded, flags=re.DOTALL)
190
+
191
+ if m:
192
+ decoded = m.group(1).strip()
193
+ else:
194
+ # The model might not generate the "### End" sequence before reaching the max tokens. In this case,
195
+ # return everything after "### Response:".
196
+ m = re.search(r"#+\s*Response:\s*(.+)", fully_decoded, flags=re.DOTALL)
197
+ if m:
198
+ decoded = m.group(1).strip()
199
+ else:
200
+ logger.warn(f"Failed to find response in:\n{fully_decoded}")
201
+
202
+ # If the full text is requested, then append the decoded text to the original instruction.
203
+ # This technically isn't the full text, as we format the instruction in the prompt the model has been
204
+ # trained on, but to the client it will appear to be the full text.
205
+ if return_full_text:
206
+ decoded = f"{instruction_text}\n{decoded}"
207
+
208
+ rec = {"generated_text": decoded}
209
+
210
+ records.append(rec)
211
+
212
+ return records