FailureSensorIQ/run_eval.py at main · IBM/FailureSensorIQ · GitHub

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
#!/usr/bin/env python
# coding: utf-8

# In[1]:


# %load_ext autoreload
# %autoreload 2


# In[1]:


import sys
import os
sys.path.append('perteval')
sys.path.append('uq_score')


# In[2]:


import json
from utils import get_requests, get_results
from perteval.perteval_end_to_end import get_perteval_results
# from uq_score.uq_end_to_end import run_uq
from llm_uncertainty_bench.uq_end_to_end import run_uq_benchmark
import datetime
import perteval.transition_analysis as tas
import time
import subprocess


# In[3]:


# try killing old processes that utilize the gpu
try:
    subprocess.run(["nvidia-smi | grep 'python' | awk '{ print $5 }' | xargs -n1 kill -9"], shell=True)
except:
    pass


# In[4]:


if len(sys.argv) == 1 or sys.argv[1] == '-f':
    # -f is to avoid the case that this runs as ipynb
    all_request_models = get_requests('cc4718/requests')
    all_result_models = get_results('cc4718/results')
    models_todo = set(all_request_models) - set(all_result_models)
    # dataset can be full/sample
    dataset = 'full'
else:
    # huggingface models
    models_todo = [sys.argv[1]]
    # dataset full/sample
    dataset = sys.argv[2]
print('models to evaluate:')
print(models_todo)
print('dataset:')
print(dataset)


# In[ ]:


for model_name in models_todo:
    # perteval
    original_log_path, test_path = get_perteval_results(model_name,
                                                        mode='original',
                                                        cot='cot_standard',
                                                        dataset=dataset)
    print('Finished perteval on original data')
    # wait a few seconds to kill vllm and spin up a new
    time.sleep(10)
    perturb_log_path, _ = get_perteval_results(model_name,
                                               mode='perturb',
                                               cot='cot_standard',
                                               dataset=dataset)
    print('Finished perteval on perturbed data')
    original, perturb, consist = tas.transition_analysis(original_log_path,
                                                         perturb_log_path,
                                                         subjects=["failure_mode_sensor_analysis"])
    asset_scores = tas.get_record_id_for_correct_answer(original_log_path,
                                                        dimention='asset_name',
                                                        fdata_path=test_path)
    relevancy_scores = tas.get_record_id_for_correct_answer(original_log_path,
                                                            dimention='relevancy',
                                                            fdata_path=test_path)
    # uq bench
    print('Running LLM Uncertainty Bench')
    uq_scores = run_uq_benchmark(model_name, prompt_type='chat', dataset=dataset)
    if dataset == 'sample':
        acc_sel_key = 'relevant_sensor_for_failure_mode'
        acc_el_key = 'irrelevant_sensor_for_failure_mode'
    else:
        acc_sel_key = 'relevant_sensors_for_failure_mode'
        acc_el_key = 'irrelevant_sensors_for_failure_mode'
    result_dict = {
        "config": {
            "model_dtype": "torch.bfloat16",
            "model_name": model_name,
            "model_sha": "main"
        },
        "results": {
            "acc_overall": {
                "acc": original
            },
            "acc_sel": {
                "acc_sel": relevancy_scores[acc_sel_key]
            },
            "acc_el": {
                "acc_el": relevancy_scores[acc_el_key]
            },
            "acc_perturb": {
                "perturb_score": perturb
            },
            "score_consistency": {
                "consist_score": consist
            }
        },
    }
    print(uq_scores)
    for k, v in uq_scores.items():
        result_dict['results'][k] = {k: v}
    for asset in asset_scores:
        if not asset:
            asset_lower = 'other'
        else:
            asset_lower = asset.lower().replace(' ', '_')
        result_dict['results'][f'acc_{asset_lower}'] = {f'acc_{asset_lower}': asset_scores[asset]}
    out_model_name = model_name.replace('/', '--')
    out_fname = f'results/demo-leaderboard/gpt2-demo/results_{out_model_name}.json'
    with open(out_fname, 'w') as f:
        f.write(json.dumps(result_dict))
    print(f'all results written to {out_fname}')


# In[ ]:


# !nvidia-smi | grep 'python' | awk '{ print $5 }' | xargs -n1 kill -9


# In[ ]: