File size: 5,258 Bytes
ff66cf3 |
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 |
import cv2
import numpy as np
import IPython
import os
import openai
import pandas as pd
import json
import subprocess
from gensim.utils import set_gpt_model, clear_messages, format_finetune_prompt, format_finetune_prompt_codeonly
def format_completion_codeonly(task_name, descriptions, code):
completion_text = " \n```python\n" + code + "\n```\n\nSTOP"
return completion_text
def format_completion(task_name, descriptions, code):
completion_text = f" \n {task_name}: {descriptions}```\n\n###"
completion_text += "\n```python\n" + code + "\n```\n\nSTOP"
return completion_text
# test if using the finetuned model can generate better task coed than the base model
# https://platform.openai.com/docs/guides/fine-tuning
data_path = 'prompts/data'
def load_offline_memory():
"""get the current task descriptions, assets, and code"""
base_task_path = os.path.join(data_path, "base_tasks.json")
base_asset_path = os.path.join(data_path, "base_assets.json")
base_task_code_path = os.path.join(data_path, "base_task_codes.json")
base_tasks = json.load(open(base_task_path))
base_assets = json.load(open(base_asset_path))
base_task_codes = json.load(open(base_task_code_path))
generated_task_path = os.path.join(data_path, "generated_tasks.json")
generated_asset_path = os.path.join(data_path, "generated_assets.json")
generated_task_code_path = os.path.join(data_path, "generated_task_codes.json")
# print("original base task num:", len(base_tasks))
base_tasks.update(json.load(open(generated_task_path)))
# base_assets.update(json.load(open(generated_asset_path)))
for task in json.load(open(generated_task_code_path)):
if task not in base_task_codes:
base_task_codes.append(task)
# print("current base task num:", len(base_tasks))
return base_tasks, base_assets, base_task_codes
code_buffer = {}
base_tasks, base_assets, base_task_codes = load_offline_memory()
TOTAL_DATASET_TOKENS = 0
added_tasks = []
df = pd.DataFrame()
file_name = 'prompts/finetune_data_new.jsonl'
file = open(file_name, 'w')
for task_file in base_task_codes:
## TODO(lirui): consider adding more structure here.
task_name = task_file[:-3].replace("_", "-")
if task_name in added_tasks:
continue
if task_name not in base_tasks:
print(f"{task_name} missing")
continue
added_tasks.append(task_name)
task_description = base_tasks[task_name]
if os.path.exists("cliport/tasks/" + task_file):
task_code = open("cliport/tasks/" + task_file).read()
# the generated cliport task path
elif os.path.exists("cliport/generated_tasks/" + task_file):
task_code = open("cliport/generated_tasks/" + task_file).read()
# prompt = format_finetune_prompt(task_name)
# completion = format_completion(task_name, task_description, task_code)
prompt = format_finetune_prompt_codeonly(task_name)
completion = format_completion_codeonly(task_name, task_description, task_code)
# rough estimates
TOTAL_DATASET_TOKENS += len(prompt) / 4
TOTAL_DATASET_TOKENS += len(completion) / 4
# new_row = { 'prompt': prompt,
# 'completion': completion}
# new_row = {"role": "system", "content": "You are an AI in robot simulation code and task design."}
# new_row = pd.DataFrame([new_row])
# df = pd.concat([df, new_row], axis=0, ignore_index=True)
# new_row = {"role": "user", "content": prompt}
# new_row = pd.DataFrame([new_row])
# df = pd.concat([df, new_row], axis=0, ignore_index=True)
# new_row = {"role": "assistant", "content": completion}
# new_row = pd.DataFrame([new_row])
# df = pd.concat([df, new_row], axis=0, ignore_index=True)
data = ({"messages": [{"role": "system", "content": "You are an AI in robot simulation code and task design."},
{"role": "user", "content": prompt},
{"role": "assistant", "content": completion}]})
# print(data)
file.write(json.dumps(data)+"\n") # jsonl
# df.to_csv("prompts/finetune_data.csv",index=False)
print("======================================")
print("estimate number of tokens:", TOTAL_DATASET_TOKENS)
print("estimate price for davinci:", TOTAL_DATASET_TOKENS / 1000 * 0.03)
print("total number of instructions:", len(df))
print("======================================")
# actual finetuning
## prepared_data.csv --> prepared_data_prepared.json
# subprocess.run('openai tools fine_tunes.prepare_data --file prompts/finetune_data.csv --quiet'.split())
print("now you can run \n python misc/job_create.py")
print("check file!:", file_name)
# Model Training Usage
# Ada $0.0004 / 1K tokens $0.0016 / 1K tokens
# Curie $0.0030 / 1K tokens $0.0120 / 1K tokens
# Davinci $0.0300 / 1K tokens $0.1200 / 1K tokens
# ## Start fine-tuning
# openai api fine_tunes.create --training_file prompts/finetune_data_new.jsonl --model gpt-3.5-turbo --suffix "GenSimNew"
# subprocess.run('openai api fine_tunes.create --training_file output/finetune_data_prepared.jsonl --model davinci --suffix "GenSim"'.split())
# Tracking Finetune Status
# openai api fine_tunes.follow -i
# openai api fine_tunes.get -i
# openai wandb sync |