-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathNLI_judge.py
133 lines (122 loc) · 6.56 KB
/
NLI_judge.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
import json
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
import argparse
from tqdm import tqdm
def read_args():
parser = argparse.ArgumentParser()
parser.add_argument("--model", required=True)
parser.add_argument("--compression_model_path", required=True)
parser.add_argument("--dataset", required=True)
parser.add_argument("--max_token_len", help='max token length for ICL demos', default=750)
parser.add_argument("--compression_rate", default=0.1)
parser.add_argument("--target_device", default='cuda')
parser.add_argument("--compression_device", default='cpu')
args = parser.parse_args()
return args
def runkey(subtree, k, printres=True, modeleval=MODEL+'_triplets', yesc=[], noc=[], printjudge=False, judgen=None):
lsofres1=[]
triplets = [t.replace('(','').replace(')','').replace(',','')+'.' for t in subtree[k]['triplets']]
embeddings = model.encode(triplets)
if(printres): print(' '.join(triplets),'\n')
c = 0
pbar = tqdm(range(len(subtree[k]['instance'])))
for i in pbar:
lsofres1.append([])
if('triplets' not in modeleval):
if('sentences' in subtree[k]['instance'][i].keys()):
evaltriplets = subtree[k]['instance'][i]['sentences']
else:
evaltriplets = subtree[k]['instance'][i][modeleval].split('. ')
evaltriplets = [e if e[-1]=='.' else e+'.' for e in evaltriplets]
subtree[k]['instance'][i]['sentences'] = evaltriplets
else: evaltriplets = subtree[k]['instance'][i][modeleval]
#objectsent = 'There are '+', '.join(list(set(tree['minigpt4'][k]['object']).union(set(tree['minigpt4'][k]['all_object']))))+'.'
for tdx,t in enumerate(evaltriplets):
if('triplets' not in modeleval): t = [t]
if(printres): print(c, [' '.join(t)])
c+=1
src = model.encode([' '.join(t)])
res = cosine_similarity(src,embeddings)
# filter not useful triplets
filtered = [' '.join([triplets[idx] for idx in np.nonzero(res>0.5)[1]])]
if(filtered==['']):
oriembeddings = model.encode([' '.join(triplets)])
topk = (-res).argsort()[0,:3]
if(printres): print('filtered',[triplets[tdx] for tdx in topk])
oriembeddings = model.encode([' '.join([triplets[tdx] for tdx in topk])])
entailval = cosine_similarity(src,oriembeddings)[0][0]
#newembeddings = model.encode(filtered)
#entailval = cosine_similarity(src,newembeddings)[0][0]
else:
if(printres): print('filtered',filtered)
oriembeddings = model.encode([' '.join(triplets)])
topk = (-res).argsort()[0,:3]
oriembeddings = model.encode([' '.join([triplets[tdx] for tdx in topk])])
orival = cosine_similarity(src,oriembeddings)[0][0]
newembeddings = model.encode(filtered)
entailval = cosine_similarity(src,newembeddings)[0][0]
entailval = max(entailval, orival)
if(entailval>0.6):
if(printres): print(entailval,'yes')
lsofres1[-1].append('yes')
else:
if(printres): print(entailval,'no')
lsofres1[-1].append('no')
if(printres):
if(printjudge and judgen in subtree[k]['instance'][i].keys() and tdx<len(subtree[k]['instance'][i][judgen])): print(subtree[k]['instance'][i][judgen][tdx])
print()
if(judgen!=None): subtree[k]['instance'][i][judgen+'_nli'] = lsofres1[-1] #add new keys of nli judgement
lsofresall = []
for l in lsofres1:
lsofresall += l
yes_c, no_c = len([r for r in lsofresall if r=='yes']), len([r for r in lsofresall if r=='no'])
pbar.set_postfix({'entail': (yes_c+sum(yesc))/(yes_c+sum(yesc)+no_c+sum(noc))})
return lsofres1
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Example script with flags and arguments.")
parser.add_argument('--model', help='model for evaluation')
args = parser.parse_args()
MODEL = args.model
model = SentenceTransformer('sentence-transformers/all-mpnet-base-v2')
tree = {
MODEL: json.load(open(MODEL+'_with_triplets.json')),
}
ks = list(tree[MODEL].keys())
treens = list(tree.keys())
modelevalns = [MODEL+'_triplets']
judgens = [MODEL+'_triplets_judgements']
# NLIeval on triplets
for x in range(len(treens)):
yesc, noc = [], []
for k in ks:
# print('key:',k)
res = runkey(tree[treens[x]], k, printres=False, modeleval=modelevalns[x], yesc=yesc, noc=noc, printjudge=True, judgen=judgens[x])
print('The result of NLI Judgement is updated in the json file.')
########### topn ablation ###########
# from nltk.tokenize.treebank import TreebankWordDetokenizer
# modeln = ['shikra', 'internlm', 'instructblip', 'llava15', 'llava1', 'minigpt4']
# modelans = ['shikra-7b', 'InternLM_XComposer2_VL', 'instruct_blip_7b', 'llava-1.5-7b', 'llava-1.1-7b', 'minigpt-4-7b(vicuna)']
# tripletsn = ['shikra_7b_triplets', 'internlm_triplets', 'instruct_blip_7b_triplets', 'llava-1.5_7b_triplets', 'llava-1.1_7b_triplets', 'minigpt4_7b(vicuna)_triplets']
# from nltk.tokenize import word_tokenize
# for no in range(len(modeln)):
# for k in tqdm(ks):
# for idx,i in enumerate(tree[modeln[no]][k]['instance']):
# maxtop50 = []
# tokenized = word_tokenize(i[modelans[no]])
# #print(tokenized)
# for topw in topls:
# b50, a50 = TreebankWordDetokenizer().detokenize(tokenized[:topw]), TreebankWordDetokenizer().detokenize(tokenized[topw:])
# eb50, ea50 = model.encode(b50.split('.')), model.encode(a50.split('.'))
# #print(b50,'\n',a50)
# top50triplets = []
# for jdx,j in enumerate(i[tripletsn[no]]):
# t = ' '.join(j)+'.'
# t = model.encode([t])
# if(cosine_similarity(t,eb50).max()<cosine_similarity(t,ea50).max()):
# break
# top50triplets.append(j)
# if(len(top50triplets)>len(maxtop50)): maxtop50 = top50triplets
# tree[modeln[no]][k]['instance'][idx][tripletsn[no]+'top'+str(topw)+'w'] = maxtop50
# tree[modeln[no]][k]['instance'][idx][modelans[no]+'top'+str(topw)+'w'] = b50