File size: 3,159 Bytes
31d41d7 |
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 |
---
license: apache-2.0
datasets:
- Ashenone3/LM-Searcher-Trajectory-228K
language:
- en
metrics:
- accuracy
base_model:
- meta-llama/Llama-3.1-8B
pipeline_tag: text-generation
tags:
- nas
- optimization
- agent
---
# LM-Searcher: Cross-domain Neural Architecture Search with LLMs via Unified Numerical Encoding
Repo: [https://github.com/Ashone3/LM-Searcher](https://github.com/Ashone3/LM-Searcher)
## Introduction
We introduce LM-Searcher, a task-agnostic neural architecture search framework powered by LLMs.
## Usage
### Deployment
We use [vllm](https://github.com/vllm-project/vllm) to deploy our LM-Searcher for inference.
After installing the dependencies required by vllm. You can deploy the model using `vllm_deploy.sh`:
```shell
vllm serve path-to-the-checkpoint --dtype auto --api-key token-abc123 --chat-template template.jinja
```
### Inference
An example is provided to show how LM-Searcher can be used to search for the optimal solution to a given problem：
```python
import os
import re
import json
import time
import random
import argparse
from decimal import Decimal
from openai import OpenAI
from utils import generate_random_cell, sample_new_cell
# -----------------------------
# Argument parser configuration
# -----------------------------
parser = argparse.ArgumentParser()
parser.add_argument('--output_dir', type=str, default='history', help="Directory to save search results.")
parser.add_argument('--chat_model', type=str, default='path-to-the-checkpoint', help="LLM model used for sampling new cells.")
parser.add_argument('--trial_num', type=int, default=192, help="Number of search trials to run.")
args = parser.parse_args()
print(args)
# -----------------------------
# Define the search space here
# (Customize according to your task)
# -----------------------------
search_space = [5, 5, 5, 5, 5, 5, 5, 5, 5, 5] # Search space with 5^10 solutions
performance_history = []
trial_dict = {}
# -----------------------------
# Create output directory if it doesn’t exist
# -----------------------------
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
num_iters = 0
for iteration in range(num_iters, args.trial_num):
# Control number of previous trials referenced by the model
if iteration <= 200:
output_num = iteration
else:
output_num = 200
# First few trials are random
if iteration <= 4:
cell = generate_random_cell(search_space, trial_dict)
# Later trials sample based on history
else:
cell = sample_new_cell(trial_dict, output_num, args.chat_model)
# -----------------------------
# Here the "reward function" is defined.
# Replace this with your custom evaluation metric.
# -----------------------------
val_acc = random.uniform(0, 100)
# Record results for the current trial
trial_dict[f"Trial{iteration+1}"] = {}
trial_dict[f"Trial{iteration+1}"]["cell"] = cell
trial_dict[f"Trial{iteration+1}"]["prediction"] = val_acc
# Save all historical results to file
with open('{}/historical_results.json'.format(args.output_dir), 'w') as f:
json.dump(trial_dict, f)
``` |