Spaces:
Sleeping
Sleeping
AhmedSSabir
commited on
Commit
•
f38b41e
1
Parent(s):
97ec5b9
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
from doctest import OutputChecker
|
3 |
+
import sys
|
4 |
+
import torch
|
5 |
+
import re
|
6 |
+
import os
|
7 |
+
import gradio as gr
|
8 |
+
import requests
|
9 |
+
|
10 |
+
#url = "https://github.com/simonepri/lm-scorer/tree/master/lm_scorer/models"
|
11 |
+
#resp = requests.get(url)
|
12 |
+
|
13 |
+
from sentence_transformers import SentenceTransformer, util
|
14 |
+
#from sentence_transformers import SentenceTransformer, util
|
15 |
+
#from sklearn.metrics.pairwise import cosine_similarity
|
16 |
+
#from lm_scorer.models.auto import AutoLMScorer as LMScorer
|
17 |
+
#from sentence_transformers import SentenceTransformer, util
|
18 |
+
#from sklearn.metrics.pairwise import cosine_similarity
|
19 |
+
|
20 |
+
|
21 |
+
#model_sts = gr.Interface.load('huggingface/sentence-transformers/stsb-distilbert-base')
|
22 |
+
|
23 |
+
model_sts = SentenceTransformer('stsb-distilbert-base')
|
24 |
+
|
25 |
+
#batch_size = 1
|
26 |
+
#scorer = LMScorer.from_pretrained('gpt2' , device=device, batch_size=batch_size)
|
27 |
+
|
28 |
+
#import torch
|
29 |
+
from transformers import GPT2Tokenizer, GPT2LMHeadModel
|
30 |
+
import numpy as np
|
31 |
+
import re
|
32 |
+
|
33 |
+
|
34 |
+
|
35 |
+
|
36 |
+
def Sort_Tuple(tup):
|
37 |
+
|
38 |
+
# (Sorts in descending order)
|
39 |
+
tup.sort(key = lambda x: x[1])
|
40 |
+
return tup[::-1]
|
41 |
+
|
42 |
+
|
43 |
+
def softmax(x):
|
44 |
+
exps = np.exp(x)
|
45 |
+
return np.divide(exps, np.sum(exps))
|
46 |
+
|
47 |
+
|
48 |
+
def get_sim(x):
|
49 |
+
x = str(x)[1:-1]
|
50 |
+
x = str(x)[1:-1]
|
51 |
+
return x
|
52 |
+
|
53 |
+
|
54 |
+
# Load pre-trained model
|
55 |
+
|
56 |
+
model = GPT2LMHeadModel.from_pretrained('distilgpt2', output_hidden_states = True, output_attentions = True)
|
57 |
+
|
58 |
+
#model = gr.Interface.load('huggingface/distilgpt2', output_hidden_states = True, output_attentions = True)
|
59 |
+
|
60 |
+
#model.eval()
|
61 |
+
#tokenizer = gr.Interface.load('huggingface/distilgpt2')
|
62 |
+
|
63 |
+
tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
|
64 |
+
#tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
|
65 |
+
|
66 |
+
|
67 |
+
def cloze_prob(text):
|
68 |
+
|
69 |
+
whole_text_encoding = tokenizer.encode(text)
|
70 |
+
# Parse out the stem of the whole sentence (i.e., the part leading up to but not including the critical word)
|
71 |
+
text_list = text.split()
|
72 |
+
stem = ' '.join(text_list[:-1])
|
73 |
+
stem_encoding = tokenizer.encode(stem)
|
74 |
+
# cw_encoding is just the difference between whole_text_encoding and stem_encoding
|
75 |
+
# note: this might not correspond exactly to the word itself
|
76 |
+
cw_encoding = whole_text_encoding[len(stem_encoding):]
|
77 |
+
# Run the entire sentence through the model. Then go "back in time" to look at what the model predicted for each token, starting at the stem.
|
78 |
+
# Put the whole text encoding into a tensor, and get the model's comprehensive output
|
79 |
+
tokens_tensor = torch.tensor([whole_text_encoding])
|
80 |
+
|
81 |
+
with torch.no_grad():
|
82 |
+
outputs = model(tokens_tensor)
|
83 |
+
predictions = outputs[0]
|
84 |
+
|
85 |
+
logprobs = []
|
86 |
+
# start at the stem and get downstream probabilities incrementally from the model(see above)
|
87 |
+
start = -1-len(cw_encoding)
|
88 |
+
for j in range(start,-1,1):
|
89 |
+
raw_output = []
|
90 |
+
for i in predictions[-1][j]:
|
91 |
+
raw_output.append(i.item())
|
92 |
+
|
93 |
+
logprobs.append(np.log(softmax(raw_output)))
|
94 |
+
|
95 |
+
# if the critical word is three tokens long, the raw_probabilities should look something like this:
|
96 |
+
# [ [0.412, 0.001, ... ] ,[0.213, 0.004, ...], [0.002,0.001, 0.93 ...]]
|
97 |
+
# Then for the i'th token we want to find its associated probability
|
98 |
+
# this is just: raw_probabilities[i][token_index]
|
99 |
+
conditional_probs = []
|
100 |
+
for cw,prob in zip(cw_encoding,logprobs):
|
101 |
+
conditional_probs.append(prob[cw])
|
102 |
+
# now that you have all the relevant probabilities, return their product.
|
103 |
+
# This is the probability of the critical word given the context before it.
|
104 |
+
|
105 |
+
return np.exp(np.sum(conditional_probs))
|
106 |
+
|
107 |
+
|
108 |
+
|
109 |
+
|
110 |
+
|
111 |
+
def cos_sim(a, b):
|
112 |
+
return np.inner(a, b) / (np.linalg.norm(a) * (np.linalg.norm(b)))
|
113 |
+
|
114 |
+
|
115 |
+
|
116 |
+
#def Visual_re_ranker(caption, visual_context_label, visual_context_prob):
|
117 |
+
def Visual_re_ranker(caption_man, caption_woman, visual_context_label, visual_context_prob):
|
118 |
+
caption_man = caption_man
|
119 |
+
caption_woman = caption_woman
|
120 |
+
visual_context_label= visual_context_label
|
121 |
+
visual_context_prob = visual_context_prob
|
122 |
+
caption_emb_man = model_sts.encode(caption_man, convert_to_tensor=True)
|
123 |
+
caption_emb_woman = model_sts.encode(caption_woman, convert_to_tensor=True)
|
124 |
+
visual_context_label_emb = model_sts.encode(visual_context_label, convert_to_tensor=True)
|
125 |
+
|
126 |
+
sim_m = cosine_scores = util.pytorch_cos_sim(caption_emb_man, visual_context_label_emb)
|
127 |
+
sim_m = sim_m.cpu().numpy()
|
128 |
+
sim_m = get_sim(sim_m)
|
129 |
+
|
130 |
+
sim_w = cosine_scores = util.pytorch_cos_sim(caption_emb_woman, visual_context_label_emb)
|
131 |
+
sim_w = sim_w.cpu().numpy()
|
132 |
+
sim_w = get_sim(sim_w)
|
133 |
+
|
134 |
+
|
135 |
+
LM_man = cloze_prob(caption_man)
|
136 |
+
LM_woman = cloze_prob(caption_woman)
|
137 |
+
#LM = scorer.sentence_score(caption, reduce="mean")
|
138 |
+
score_man = pow(float(LM_man),pow((1-float(sim_m))/(1+ float(sim_m)),1-float(visual_context_prob)))
|
139 |
+
score_woman = pow(float(LM_woman),pow((1-float(sim_w))/(1+ float(sim_w)),1-float(visual_context_prob)))
|
140 |
+
|
141 |
+
|
142 |
+
|
143 |
+
|
144 |
+
#return {"LM": float(LM)/1, "sim": float(sim)/1, "score": float(score)/1 }
|
145 |
+
return {"Man": float(score_man)/1, "Woman": float(score_woman)/1}
|
146 |
+
#return LM, sim, score
|
147 |
+
|
148 |
+
|
149 |
+
|
150 |
+
|
151 |
+
|
152 |
+
demo = gr.Interface(
|
153 |
+
fn=Visual_re_ranker,
|
154 |
+
description="Demo for Women Wearing Lipstick: Measuring the Bias Between Object and Its Related Gender",
|
155 |
+
inputs=[gr.Textbox(value="a man sitting on a surfboard in the ocean") , gr.Textbox(value="a man sitting on a surfboard in the ocean"), gr.Textbox(value="paddle"), gr.Textbox(value="0.5283")],
|
156 |
+
#outputs=[gr.Textbox(value="Language Model Score") , gr.Textbox(value="Semantic Similarity Score"), gr.Textbox(value="Belief revision score via visual context")],
|
157 |
+
outputs="label",
|
158 |
+
)
|
159 |
+
demo.launch()
|
160 |
+
|