-
Notifications
You must be signed in to change notification settings - Fork 3
/
self_training.py
175 lines (151 loc) · 8.15 KB
/
self_training.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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
import json
import numpy as np
import pandas as pd
import os
import jsonlines
import random
import argparse
import logging
import subprocess
logger = logging.getLogger('self_training_logger')
# logger.setLevel(logging.DEBUG)
def parse_args():
parser = argparse.ArgumentParser(description="self-training framework combining symbolic solver and llms")
parser.add_argument(
"--domain",
type=str,
default="math",
help="The name of the domain [math,agent,logic].",
)
parser.add_argument(
"--iter_num",
type=int,
default=10,
help="The number of iteration for the self-training.",
)
parser.add_argument(
"--vllm_batchsize",
type=int,
default=4,
help="batchsize for vllm",
)
parser.add_argument(
"--task_prefix",
type=str,
default="gsm_math_full_llama2chat",
help="The prefix for the dataset name, directory name and save path",
)
parser.add_argument(
"--model_size",
type=str,
default="7B",
help="Model size",
)
parser.add_argument(
"--base_model",
type=str,
default="llama2chat",
help="base model",
)
args = parser.parse_args()
return args
def main():
args = parse_args()
base_model = args.base_model
# ======================================================================== #
# Generate the initial steps
# ======================================================================== #
processes = []
for part_id in range(1,9): # split the dataset into 8 parts (on 8 GPUs)
env = os.environ.copy()
env['CUDA_VISIBLE_DEVICES'] = str(part_id - 1)
process = subprocess.Popen(
['python', 'ENVISIONS/generate_symbol_output/generate_candidates_vLLM_self_training.py', \
"--cur_iter", "0", "--few_shot", "--part_id", str(part_id), "--task_prefix", args.task_prefix, "--base_model",
base_model, "--model_size", args.model_size], env=env)
processes.append(process)
# wait for all the process to be completed
for process in processes:
process.wait()
# ensure candidates are generated successfully
for i in range(8):
assert os.path.exists(
f"ENVISIONS/score_memory/{args.task_prefix}/{args.task_prefix}_part{i + 1}_iter0.json") == True, "generated candidates file does not exist..."
# label preferences for the data before
subprocess.call(["python", "ENVISIONS/pal/scripts/label_preference.py", \
"--task_prefix", args.task_prefix, "--cur_iter", "0", "--few_shot"])
for i in range(8):
assert os.path.exists(f"ENVISIONS/score_memory/{args.task_prefix}/scores_{args.task_prefix}_part{i+1}_iter0.npy") == True
# ======================================================================== #
# Start the self-training loops
# ======================================================================== #
for cur_iter in range(0,args.iter_num):
logger.info(f"Current iteration: {cur_iter}")
# ======================================================================== #
# Step 1: Generate Samples
# ======================================================================== #
logger.info(f"Start to generate samples for iteration-{cur_iter}")
subprocess.call(["python", f"ENVISIONS/self-training/organize_preference_data_{base_model}.py", \
"--task_prefix", args.task_prefix, "--cur_iter", str(cur_iter), \
"--model_size", args.model_size])
assert os.path.exists(f"ENVISIONS/open-instruct/data/{args.task_prefix}_sft_iter{cur_iter}.jsonl") == True, "training set does not exist..."
# ======================================================================== #
# Step 2: Training LLM (call open-instruct)
# ======================================================================== #
logger.info(f"Start to train LLM for iteration-{cur_iter}")
training_bash_script = "ENVISIONS/open-instruct/scripts/finetune_with_accelerate_self_training.sh"
# call the training scripts
subprocess.call(["bash", training_bash_script, args.task_prefix, str(cur_iter), base_model, args.model_size])
assert len(os.listdir((f"ENVISIONS/open-instruct/output/{args.task_prefix}_sft_iter{cur_iter}_sft_tune_{base_model}_{args.model_size}"))) != 0, "The checkpoint does not exist"
# ======================================================================== #
# Step 3: Generate Candidates with vLLM
# ======================================================================== #
logger.info(f"Start to generate candidates for iteration-{cur_iter}")
processes = []
for part_id in range(1,9):
env = os.environ.copy()
env['CUDA_VISIBLE_DEVICES'] = str(part_id - 1)
process = subprocess.Popen(['python', 'ENVISIONS/generate_symbol_output/generate_candidates_vLLM_self_training.py', \
"--cur_iter", str(cur_iter), "--part_id", str(part_id), "--task_prefix", args.task_prefix, \
"--base_model", base_model,"--vllm_batchsize", str(args.vllm_batchsize), "--model_size", args.model_size], env=env)
processes.append(process)
# wait for all the process to be completed
for process in processes:
process.wait()
# ensure candidates are generated successfully
for i in range(8):
assert os.path.exists(f"new_generated_data/{args.task_prefix}_part{i+1}_iter{cur_iter+1}.json") == True, "generated candidates file does not exist..."
# ======================================================================== #
# Step 4: Repair Candidates with vLLM
# ======================================================================== #
logger.info(f"Start to repair candidates for iteration-{cur_iter}")
processes = []
for part_id in range(1,9):
env = os.environ.copy()
env['CUDA_VISIBLE_DEVICES'] = str(part_id - 1)
process = subprocess.Popen(['python', 'ENVISIONS/generate_symbol_output/generate_repaired_math_vLLM_self_training.py', \
"--cur_iter", str(cur_iter), "--part_id", str(part_id), "--task_prefix", args.task_prefix, "--base_model", base_model, \
"--model_size", args.model_size],env=env)
processes.append(process)
# wait for all the process to be completed
for process in processes:
process.wait()
# ensure repaired candidates are generated successfully
for i in range(8):
assert os.path.exists(f"new_generated_data/{args.task_prefix}_part{i+1}_iter{cur_iter+1}_repaired.json") == True, "generated candidates file does not exist..."
# ======================================================================== #
# Step 5: Check the correctness of candidates
# ======================================================================== #
# label preferences for the data before
subprocess.call(["python", "ENVISIONS/pal/scripts/label_preference.py", \
"--task_prefix", args.task_prefix, "--cur_iter", str(cur_iter)])
for i in range(8):
assert os.path.exists(f"ENVISIONS/score_memory/{args.task_prefix}/scores_{args.task_prefix}_part{i+1}_iter{cur_iter+1}.npy") == True
# label perference for the repaired data
subprocess.call(["python", "ENVISIONS/pal/scripts/label_preference.py", \
"--task_prefix", args.task_prefix, "--cur_iter", str(cur_iter), "--repaired",])
for i in range(8):
assert os.path.exists(f"ENVISIONS/score_memory/{args.task_prefix}/scores_{args.task_prefix}_part{i+1}_iter{cur_iter+1}_repaired.npy") == True
logger.info("Self-Training process has finished successfully ! ! !")
if __name__ == "__main__":
main()