-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
228 lines (194 loc) · 11.9 KB
/
eval.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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import itertools
from langchain.cache import InMemoryCache, SQLiteCache
import langchain
import pandas as pd
from certa.utils import merge_sources
from certa.explain import CertaExplainer
from datetime import datetime
import os
import ellmer.models
import ellmer.metrics
from time import sleep, time
import json
import traceback
from tqdm import tqdm
import argparse
def eval(cache, samples, num_triangles, explanation_granularity, quantitative, base_dir, dataset_names, model_type,
model_name, deployment_name, tag, temperature):
if cache == "memory":
langchain.llm_cache = InMemoryCache()
elif cache == "sqlite":
langchain.llm_cache = SQLiteCache(database_path=".langchain.db")
llm_config = {"model_type": model_type, "model_name": model_name, "deployment_name": deployment_name, "tag": tag}
pase = ellmer.models.SelfExplainer(explanation_granularity=explanation_granularity,
deployment_name=llm_config['deployment_name'], temperature=temperature,
model_name=llm_config['model_name'], model_type=llm_config['model_type'],
prompts={"pase": "ellmer/prompts/constrained7.txt"})
ptse = ellmer.models.SelfExplainer(explanation_granularity=explanation_granularity,
deployment_name=llm_config['deployment_name'], temperature=temperature,
model_name=llm_config['model_name'], model_type=llm_config['model_type'],
prompts={"ptse": {"er": "ellmer/prompts/er.txt",
"saliency": "ellmer/prompts/er-saliency-lc.txt",
"cf": "ellmer/prompts/er-cf-lc.txt"}})
ptsew = ellmer.models.SelfExplainer(explanation_granularity=explanation_granularity,
deployment_name=llm_config['deployment_name'], temperature=temperature,
model_name=llm_config['model_name'], model_type=llm_config['model_type'],
prompts={
"ptse": {"er": "ellmer/prompts/er.txt",
"why": "ellmer/prompts/er-why.txt",
"saliency": "ellmer/prompts/er-saliency-lc.txt",
"cf": "ellmer/prompts/er-cf-lc.txt"}})
ptn = ellmer.models.SelfExplainer(explanation_granularity=explanation_granularity,
deployment_name=llm_config['deployment_name'], temperature=temperature,
model_name=llm_config['model_name'], model_type=llm_config['model_type'],
prompts={"ptse": {"er": "ellmer/prompts/er.txt"}})
evals = []
for d in dataset_names:
expdir = f'./experiments/{model_type}/{model_name}/{explanation_granularity}/{d}/{datetime.now():%Y%m%d}/{datetime.now():%H_%M}/'
obs_dir = f'experiments/{model_type}/{model_name}/{explanation_granularity}/concordance/{d}//{datetime.now():%Y%m%d}/{datetime.now():%H_%M}'
print(f'using dataset {d}')
dataset_dir = '/'.join([base_dir, d])
lsource = pd.read_csv(dataset_dir + '/tableA.csv')
rsource = pd.read_csv(dataset_dir + '/tableB.csv')
test = pd.read_csv(dataset_dir + '/test.csv')
test_df = merge_sources(test, 'ltable_', 'rtable_', lsource, rsource, ['label'],
[])
certa = CertaExplainer(lsource, rsource)
ellmers = {
"pase_" + llm_config['tag']: pase,
"ptse_" + llm_config['tag']: ptse,
"ptsew_" + llm_config['tag']: ptsew,
"certa(ptse)_" + llm_config['tag']: ellmer.models.FullCerta(explanation_granularity, ptn, certa,
num_triangles),
"certa(pase)_" + llm_config['tag']: ellmer.models.FullCerta(explanation_granularity, pase, certa,
num_triangles),
"ellmer(certa)_" + llm_config['tag']: ellmer.models.HybridCerta(explanation_granularity, ptse, certa,
[pase, ptse, ptsew],
num_triangles=num_triangles),
"ellmer(lm)_" + llm_config['tag']: ellmer.models.HybridGeneric(explanation_granularity, ptse,
[pase, ptse, ptsew], lsource, rsource),
}
result_files = []
for key, llm in ellmers.items():
print(f'{key} on {d}')
curr_llm_results = []
start_time = time()
# generate predictions and explanations
test_data_df = test_df[:samples]
ranged = range(len(test_data_df))
for idx in tqdm(ranged, disable=False):
try:
rand_row = test_df.iloc[[idx]]
ltuple, rtuple = ellmer.utils.get_tuples(rand_row)
ptime = time()
answer_dictionary = llm.predict_and_explain(ltuple, rtuple)
ptime = time() - ptime
prediction = answer_dictionary['prediction']
saliency = answer_dictionary['saliency']
cfs = [answer_dictionary['cf']]
if 'conversation' in answer_dictionary:
conversation = answer_dictionary['conversation']
else:
conversation = ''
row_dict = {"id": idx, "ltuple": ltuple, "rtuple": rtuple, "prediction": prediction,
"label": rand_row['label'].values[0], "saliency": saliency, "cfs": cfs,
"latency": ptime, "conversation": conversation}
if "filter_features" in answer_dictionary:
row_dict["filter_features"] = answer_dictionary["filter_features"]
curr_llm_results.append(row_dict)
except Exception:
traceback.print_exc()
print(f'error, waiting...')
sleep(10)
start_time += 10
total_time = time() - start_time
os.makedirs(expdir, exist_ok=True)
llm_results = {"data": curr_llm_results, "total_time": total_time}
output_file_path = expdir + key + '_results.json'
with open(output_file_path, 'w') as fout:
json.dump(llm_results, fout)
faithfulness = 'nan'
cf_metrics = {}
count_tokens_samples = 'nan'
predictions_samples = 'nan'
if quantitative:
# generate quantitative explainability metrics for each set of generated explanations
# generate saliency metrics
faithfulness = ellmer.metrics.get_faithfulness([key], llm.evaluation, expdir, test_data_df)
print(f'{key} faithfulness({key}):{faithfulness}')
# generate counterfactual metrics
cf_metrics = ellmer.metrics.get_cf_metrics([key], llm.predict, expdir, test_data_df)
print(f'{key} cf_metrics({key}):{cf_metrics}')
metrics_results = {"faithfulness": faithfulness, "counterfactual_metrics": cf_metrics}
count_tokens_samples = llm.count_tokens() / samples
predictions_samples = llm.count_predictions() / samples
llm_results = {"data": curr_llm_results, "total_time": total_time, "metrics": metrics_results,
"tokens": count_tokens_samples, "predictions": predictions_samples}
output_file_path = expdir + key + '_results.json'
with open(output_file_path, 'w') as fout:
json.dump(llm_results, fout)
result_files.append((key, output_file_path))
print(f'{key} data generated in {total_time}s')
row_dict = {"total_time": total_time, "tokens": count_tokens_samples, "predictions": predictions_samples,
"faithfulness": faithfulness, "model": key, "dataset": d}
for cfk, cfv in cf_metrics.items():
row_dict[cfk] = cfv
eval_row = pd.Series(row_dict)
evals.append(eval_row)
# generate concordance statistics for each pair of results
for pair in itertools.combinations(result_files, 2):
p1 = pair[0]
p1_name = p1[0]
p1_file = p1[1]
p2 = pair[1]
p2_name = p2[0]
p2_file = p2[1]
print(f'concordance statistics for {p1_name} - {p2_name}')
observations = ellmer.metrics.get_concordance(p1_file, p2_file)
print(f'{observations}')
os.makedirs(obs_dir, exist_ok=True)
observations.to_csv(f'{obs_dir}/{p1_name}_{p2_name}.csv')
eval_df = pd.DataFrame(evals)
eval_expdir = f'./experiments/{model_type}/{model_name}/{explanation_granularity}/{datetime.now():%Y%m%d}/{datetime.now():%H_%M}/'
os.makedirs(eval_expdir, exist_ok=True)
eval_df.to_csv(eval_expdir + "eval.csv")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Run saliency experiments.')
parser.add_argument('--base_dir', metavar='b', type=str, help='the datasets base directory',
required=True)
parser.add_argument('--model_type', metavar='m', type=str, help='the LLM type to evaluate',
choices=['azure_openai', 'falcon', 'llama2', 'hf'], required=True)
parser.add_argument('--datasets', metavar='d', type=str, nargs='+', required=True,
help='the dataset(s) to be used for the evaluation')
parser.add_argument('--samples', metavar='s', type=int, default=-1,
help='no. of samples from the test set used for the evaluation')
parser.add_argument('--cache', metavar='c', type=str, choices=['', 'sqlite', 'memory'], default='',
help='LLM prediction caching mechanism')
parser.add_argument('--num_triangles', metavar='t', type=int, default=10,
help='no. of open triangles used to generate CERTA explanations')
parser.add_argument('--granularity', metavar='tk', type=str, default='attribute',
choices=['attribute', 'token'], help='explanation granularity')
parser.add_argument('--quantitative', metavar='q', type=bool, default=True,
help='whether to generate quantitative explanation evaluation results')
parser.add_argument('--model_name', metavar='mn', type=str, help='model name/identifier',
default="gpt-3.5-turbo")
parser.add_argument('--deployment_name', metavar='dn', type=str, help='deployment name',
default="gpt-35-turbo")
parser.add_argument('--tag', metavar='tg', type=str, help='run tag', default="sample")
parser.add_argument('--temperature', metavar='tp', type=float, help='LLM temperature', default=0.01)
args = parser.parse_args()
base_datadir = args.base_dir
samples = args.samples
num_triangles = args.num_triangles
temperature = args.temperature
cache = args.cache
explanation_granularity = args.granularity
quantitative = args.quantitative
dataset_names = args.datasets
base_dir = args.base_dir
model_type = args.model_type
model_name = args.model_name
deployment_name = args.deployment_name
tag = args.tag
eval(cache, samples, num_triangles, explanation_granularity, quantitative, base_dir, dataset_names, model_type,
model_name, deployment_name, tag, temperature)