-
Notifications
You must be signed in to change notification settings - Fork 0
/
metrics.py
183 lines (143 loc) · 5.66 KB
/
metrics.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
import json
import numpy as np
import torch
from numpy import mean
from nltk.translate import meteor_score
from pycocoevalcap.bleu.bleu import Bleu
from pycocoevalcap.cider.cider import Cider
# from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.rouge.rouge import Rouge
from pycocoevalcap.spice.spice import Spice
from torchmetrics.functional.multimodal.clip_score import _get_model_and_processor
from tqdm import tqdm
def bleu_scores(gts, res, n=4):
results = {}
scorer = Bleu(n)
score, scores = scorer.compute_score(gts, res)
if type(score) == list:
for i, s in enumerate(score):
results['BLEU {}'.format(i + 1)] = s
results['BLEU {} scores'.format(i + 1)] = scores[i]
else:
results['BLEU 1'] = score
results['BLEU 1 scores'] = scores
return results
def cider_scores(gts, res):
scorer = Cider()
score, scores = scorer.compute_score(gts, res)
return {'CIDEr': score, 'CIDEr scores': scores}
# This part of code contains some bugs we are unable to fix.
# def meteor_scores(gts, res):
# scorer = Meteor()
# score, scores = scorer.compute_score(gts, res)
#
# return {'METEOR': score, 'METEOR scores':scores}
def meteor_scores(gts, res):
scores = []
for id in gts.keys():
score_m = meteor_score.single_meteor_score(gts[id][0].split(), res[id][0].split())
scores.append(score_m)
score = mean(scores)
return {'METEOR': score, 'METEOR scores': scores}
def rougel_scores(gts, res):
scorer = Rouge()
score, scores = scorer.compute_score(gts, res)
return {'ROUGE-L': score, 'ROUGE-L scores': scores}
def spice_scores(gts, res):
scorer = Spice()
score, scores = scorer.compute_score(gts, res)
return {'SPICE': score, 'SPICE scores': scores}
def clip_scores(res, img_features):
model, processor = _get_model_and_processor("openai/clip-vit-base-patch32")
model.eval()
model = model.cuda()
b_s = 128
n_samples = len(res)
if isinstance(res, dict):
res = [v[0] for v in res.values()]
elif isinstance(res, list):
res = [v[0] for v in res]
else:
raise RuntimeError("Invalid res!")
txt_features = []
for i in range(0, n_samples, b_s):
if i + b_s <= n_samples:
batch = res[i: i + b_s]
else:
batch = res[i: n_samples]
processed_input = processor(text=batch, return_tensors="pt", padding=True)
with torch.no_grad():
txt_feature = model.get_text_features(
processed_input["input_ids"][:, :77].cuda(), processed_input["attention_mask"][:, :77].cuda()
)
txt_feature = txt_feature / txt_feature.norm(p=2, dim=-1, keepdim=True)
txt_feature = txt_feature.to(img_features.device)
txt_features.append(txt_feature)
txt_features = torch.cat(txt_features, dim=0)
scores = 100 * (txt_features * img_features).sum(-1)
score = scores.mean().item()
score = max(score, 0)
return {'CLIP': score, 'CLIP scores': scores.cpu()}
def show_all_scores(gts, res, n=4, clip_feature=None):
"""
:param gts: diction of ground truths
:param res: diction of references
:return: Language scores
"""
if isinstance(gts, list):
gts = {i: [gts[i]] for i in range(len(gts))}
if isinstance(res, list):
res = {i: [res[i]] for i in range(len(res))}
assert gts.keys() == res.keys(), "ERROR: The keys of references and ground truths are unequal!"
bleu_results = bleu_scores(gts, res, n=n)
meteor_results = meteor_scores(gts, res)
cider_results = cider_scores(gts, res)
rouge_results = rougel_scores(gts, res)
rsum = 0.0
for i in range(n):
rsum += bleu_results['BLEU {}'.format(i + 1)]
print("BLEU {} score: {}".format(i + 1, bleu_results['BLEU {}'.format(i + 1)] * 100))
rsum += meteor_results['METEOR']
print("METEOR score: {}".format(meteor_results['METEOR'] * 100))
rsum += cider_results['CIDEr']
print("CIDEr score: {}".format(cider_results['CIDEr'] * 100))
rsum += rouge_results['ROUGE-L']
print("ROUGE-L score: {}".format(rouge_results['ROUGE-L'] * 100))
print("rSUM score: {}".format(rsum * 100))
if clip_feature is not None:
clip_results = clip_scores(res, clip_feature)
print("CLIP score: {}".format(clip_results['CLIP']))
else:
clip_results = None
return bleu_results, meteor_results, cider_results, rouge_results, rsum, clip_results
def test_stories(end_gen, gold_file="./data/gen/test.txt"):
with open(gold_file, 'r', encoding='utf-8') as gf:
gts = [txt.strip('\n') for txt in gf.readlines()]
res = [txt.lower() for txt in end_gen]
return show_all_scores(gts, res)
def test_clip(end_gen, dataset="VIST-E"):
story = json.load(open("data/{}/story_test.json".format(dataset)))
story = {str(s["story_id"]): s["last_img_id"] for s in story}
res = json.load(open(end_gen))
new_res = []
img_features = []
for k in tqdm(res.keys()):
try:
img_id = story[k]
img_features.append(np.load("data/{}/clip_features/{}.npy".format(dataset, img_id)))
new_res.append(res[k])
except:
continue
img_features = torch.tensor(np.stack(img_features))
score = clip_scores(new_res, img_features)
return score
if __name__ == '__main__':
# Input stories
gold_file = ".results/VIST-E/gts.txt"
pred_file = "results/VIST-E/res_gpt4_test.txt"
tgt = "minigpt4_1s.json"
#score = test_clip("results/VIST-E/" + tgt, dataset="VIST-E")
#print(score)
gts = json.load(open("results/VIST-E/gts.json"))
res = json.load(open("results/VIST-E/" + tgt))
show_all_scores(gts, res)