-
-
Notifications
You must be signed in to change notification settings - Fork 4
/
generate.py
157 lines (118 loc) · 6.58 KB
/
generate.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import argparse
import asyncio
import json
import logging
import re
from tqdm import tqdm
from conversation import ConversationBatchService
from conversation.huggingface import HuggingfaceBatchService
from conversation.llama_cpp import LLaMACppBatchService
from conversation.openai import OpenAIBatchService
from enum_definitions import *
from data_definitions import *
from typing import Iterator
import yaml
import transformers
import torch
from dataset import coco, open_images
from sample_generation import *
def main():
parser = argparse.ArgumentParser(description="Generate a question-answer instruction tuning dataset.")
# Mandatory parameter
parser.add_argument("sources", nargs='+', type=Source, choices=list(Source),
help="Datasets to use as sources. Available: " + ", ".join([s.value for s in Source]))
# Optional parameters
parser.add_argument("--dataset_storage_path", default="./cache/", type=str,
help="Storage path for downloaded dataset files")
parser.add_argument("--output_path", default="instruct.jsonl", type=str,
help="Output file path for the created dataset (default: llava_instruct.json)")
parser.add_argument("--output_format", default=OutputFormat.JSONL, type=OutputFormat,
choices=list(OutputFormat),
help="Output format of the dataset (default: llava-instruct). Available: " + ", ".join(
[o.value for o in OutputFormat]))
# parser.add_argument("--context", default=[ContextProperties.CAPTIONS, ContextProperties.BOXES], nargs='*',
# type=ContextProperties,
# choices=list(ContextProperties),
# help="Context appended to question and response generation queries (default: captions, boxes). Available: " + ", ".join(
# [c.value for c in ContextProperties]))
#
# parser.add_argument("--question_context", default=[], nargs='*', type=QuestionContext,
# choices=list(QuestionContext),
# help="Context appended to output questions, visible at test time (default: none). Available: " + ", ".join(
# [qc.value for qc in QuestionContext]))
parser.add_argument("--model_source", default="huggingface",
type=ModelSourceType,
choices=list(ModelSourceType),
help="if huggingface: model must support conversational interface, "
"see https://huggingface.co/models?filter=conversational")
parser.add_argument("--model", default="meta-llama/Llama-2-7b-chat-hf", type=str,
help="LLM to use for dataset generation.")
parser.add_argument("--openai_base_url", default=None, type=str,
help="Base url of OpenAI compatible endpoint. Default is official OpenAI API endpoint.")
parser.add_argument("--prompt_config", default="prompt_config_llava.yaml", type=str,
help="Prompt config")
args = parser.parse_args()
# Placeholder for the dataset generation logic.
print("Generating dataset with the following parameters:")
print("Sources: ", args.sources)
print("Output Path: ", args.output_path)
print("Output Format: ", args.output_format)
loop = asyncio.get_event_loop()
loop.run_until_complete(run(args))
async def run(args):
# model inference
conversation_service: ConversationBatchService
if args.model_source == ModelSourceType.OPENAI:
conversation_service = OpenAIBatchService(args.model, args.openai_base_url)
elif args.model_source == ModelSourceType.LLAMA_CPP:
conversation_service = LLaMACppBatchService(args.model)
elif args.model_source == ModelSourceType.HUGGINGFACE:
conversation_service = HuggingfaceBatchService(args.model)
else:
raise ValueError
pipe_name = re.sub(r'\W+', '_', args.model).lower()
with open(args.prompt_config, 'r') as file:
prompt_configs = yaml.safe_load(file)
output_file = open(args.output_path, 'w')
def on_result(question_id: str, result: str, context: Context, prompt_config):
samples = process_llm_result(question_id, result, context, prompt_config)
for sample in samples:
write_str = json.dumps(sample, default=lambda x: x.__dict__)
output_file.write(write_str + '\n')
conversation_service.set_on_result(on_result)
# process datasets
for source in args.sources:
# data loader
generator: Iterator[Context]
if source == Source.COCO2014 or source == Source.COCO2017:
generator = coco.COCOLoader(source.value, args.dataset_storage_path)
elif source == Source.OPENIMAGESV7:
generator: Iterator[Context] = open_images.OpenImagesLoader(args.dataset_storage_path)
else:
raise ValueError
num_expected_requests = len(generator) * len(prompt_configs.values())
with tqdm(total=num_expected_requests, desc="Generating samples") as progress_bar:
def on_result_progress(*vargs, **kwargs):
progress_bar.update(1)
progress_bar.set_description(f"{conversation_service.num_in_progress} running, {conversation_service.num_temp_failed} temp errors, {conversation_service.num_failed} failed (completed from cache {conversation_service.num_completed_from_cache})")
progress_bar.refresh()
on_result(*vargs, **kwargs)
conversation_service.set_on_result(on_result_progress)
tasks: List[asyncio.Task] = []
for i, context in enumerate(generator):
for prompt_config in prompt_configs.values():
messages = generate_samples(context, prompt_config)
# run pipeline and cache
messages_hash = hashlib.sha256(json.dumps(messages, sort_keys=True).encode('utf-8')).hexdigest()
question_id = f"{pipe_name}_{generator.name}_{messages_hash}_{prompt_config['type']}"
task = await conversation_service.submit(question_id, messages, context=context, prompt_config=prompt_config)
tasks.append(task)
if len(tasks) > 1024:
logging.info(f"Gathering {len(tasks)} tasks.")
await asyncio.gather(*tasks)
tasks.clear()
await asyncio.gather(*tasks)
await conversation_service.finish()
output_file.close()
if __name__ == "__main__":
main()