marouni commited on
Commit
768ac84
1 Parent(s): f8a5605

chore(): add py pipeline

Browse files
Files changed (1) hide show
  1. instruct_pipeline.py +212 -0
instruct_pipeline.py ADDED
@@ -0,0 +1,212 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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:"
17
+ END_KEY = "### End"
18
+ INTRO_BLURB = (
19
+ "Below is an instruction that describes a task. Write a response that appropriately completes the request."
20
+ )
21
+
22
+ # This is the prompt that is used for generating responses using an already trained model. It ends with the response
23
+ # key, where the job of the model is to provide the completion that follows it (i.e. the response itself).
24
+ PROMPT_FOR_GENERATION_FORMAT = """{intro}
25
+
26
+ {instruction_key}
27
+ {instruction}
28
+
29
+ {response_key}
30
+ """.format(
31
+ intro=INTRO_BLURB,
32
+ instruction_key=INSTRUCTION_KEY,
33
+ instruction="{instruction}",
34
+ response_key=RESPONSE_KEY,
35
+ )
36
+
37
+
38
+ def get_special_token_id(tokenizer: PreTrainedTokenizer, key: str) -> int:
39
+ """Gets the token ID for a given string that has been added to the tokenizer as a special token.
40
+
41
+ When training, we configure the tokenizer so that the sequences like "### Instruction:" and "### End" are
42
+ treated specially and converted to a single, new token. This retrieves the token ID each of these keys map to.
43
+
44
+ Args:
45
+ tokenizer (PreTrainedTokenizer): the tokenizer
46
+ key (str): the key to convert to a single token
47
+
48
+ Raises:
49
+ RuntimeError: if more than one ID was generated
50
+
51
+ Returns:
52
+ int: the token ID for the given key
53
+ """
54
+ token_ids = tokenizer.encode(key)
55
+ if len(token_ids) > 1:
56
+ raise ValueError(f"Expected only a single token for '{key}' but found {token_ids}")
57
+ return token_ids[0]
58
+
59
+
60
+ class InstructionTextGenerationPipeline(Pipeline):
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
83
+ # append a newline to yield a single token. find whatever token is configured for the response key.
84
+ tokenizer_response_key = next(
85
+ (token for token in self.tokenizer.additional_special_tokens if token.startswith(RESPONSE_KEY)), None
86
+ )
87
+
88
+ response_key_token_id = None
89
+ end_key_token_id = None
90
+ if tokenizer_response_key:
91
+ try:
92
+ response_key_token_id = get_special_token_id(self.tokenizer, tokenizer_response_key)
93
+ end_key_token_id = get_special_token_id(self.tokenizer, END_KEY)
94
+
95
+ # Ensure generation stops once it generates "### End"
96
+ generate_kwargs["eos_token_id"] = end_key_token_id
97
+ except ValueError:
98
+ pass
99
+
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):
112
+ prompt_text = PROMPT_FOR_GENERATION_FORMAT.format(instruction=instruction_text)
113
+ inputs = self.tokenizer(
114
+ prompt_text,
115
+ return_tensors="pt",
116
+ )
117
+ inputs["prompt_text"] = prompt_text
118
+ inputs["instruction_text"] = instruction_text
119
+ return inputs
120
+
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