AhmedSSabir's picture
Update app.py
959c369
raw
history blame
4.06 kB
#!/usr/bin/env python3
from doctest import OutputChecker
import sys
import torch
import re
import os
import gradio as gr
import requests
# just for the sake of this demo, we use cloze prob to initialize the hypothesis
#url = "https://github.com/simonepri/lm-scorer/tree/master/lm_scorer/models"
#resp = requests.get(url)
from sentence_transformers import SentenceTransformer, util
model_sts = SentenceTransformer('stsb-distilbert-base')
# model_sts = SentenceTransformer('roberta-large-nli-stsb-mean-tokens')
#batch_size = 1
#scorer = LMScorer.from_pretrained('gpt2' , device=device, batch_size=batch_size)
#import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel
import numpy as np
import re
def Sort_Tuple(tup):
# (Sorts in descending order)
tup.sort(key = lambda x: x[1])
return tup[::-1]
def softmax(x):
exps = np.exp(x)
return np.divide(exps, np.sum(exps))
def get_sim(x):
x = str(x)[1:-1]
x = str(x)[1:-1]
return x
# Load pre-trained model
#model = GPT2LMHeadModel.from_pretrained('distilgpt2', output_hidden_states = True, output_attentions = True)
model = GPT2LMHeadModel.from_pretrained('gpt2', output_hidden_states = True, output_attentions = True)
#model = gr.Interface.load('huggingface/distilgpt2', output_hidden_states = True, output_attentions = True)
#model.eval()
#tokenizer = gr.Interface.load('huggingface/distilgpt2')
#tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
#tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
def cloze_prob(text):
whole_text_encoding = tokenizer.encode(text)
text_list = text.split()
stem = ' '.join(text_list[:-1])
stem_encoding = tokenizer.encode(stem)
cw_encoding = whole_text_encoding[len(stem_encoding):]
tokens_tensor = torch.tensor([whole_text_encoding])
with torch.no_grad():
outputs = model(tokens_tensor)
predictions = outputs[0]
logprobs = []
start = -1-len(cw_encoding)
for j in range(start,-1,1):
raw_output = []
for i in predictions[-1][j]:
raw_output.append(i.item())
logprobs.append(np.log(softmax(raw_output)))
conditional_probs = []
for cw,prob in zip(cw_encoding,logprobs):
conditional_probs.append(prob[cw])
return np.exp(np.sum(conditional_probs))
def cos_sim(a, b):
return np.inner(a, b) / (np.linalg.norm(a) * (np.linalg.norm(b)))
def Visual_re_ranker(caption_man, caption_woman, context_label, context_prob):
caption_man = caption_man
caption_woman = caption_woman
context_label= context_label
context_prob = context_prob
caption_emb_man = model_sts.encode(caption_man, convert_to_tensor=True)
caption_emb_woman = model_sts.encode(caption_woman, convert_to_tensor=True)
context_label_emb = model_sts.encode(context_label, convert_to_tensor=True)
sim_m = cosine_scores = util.pytorch_cos_sim(caption_emb_man, context_label_emb)
sim_m = sim_m.cpu().numpy()
sim_m = get_sim(sim_m)
sim_w = cosine_scores = util.pytorch_cos_sim(caption_emb_woman, context_label_emb)
sim_w = sim_w.cpu().numpy()
sim_w = get_sim(sim_w)
LM_man = cloze_prob(caption_man)
LM_woman = cloze_prob(caption_woman)
#LM = scorer.sentence_score(caption, reduce="mean")
score_man = pow(float(LM_man),pow((1-float(sim_m))/(1+ float(sim_m)),1-float(context_prob)))
score_woman = pow(float(LM_woman),pow((1-float(sim_w))/(1+ float(sim_w)),1-float(context_prob)))
#return {"LM": float(LM)/1, "sim": float(sim)/1, "score": float(score)/1 }
return {"Man": float(score_man)/1, "Woman": float(score_woman)/1}
#return LM, sim, score
demo = gr.Interface(
fn=Visual_re_ranker,
description="Demo for Women Wearing Lipstick: Measuring the Bias Between Object and Its Related Gender",
inputs=[gr.Textbox(value="a man riding a motorcycle on a road") , gr.Textbox(value="a woman riding a motorcycle on a road"), gr.Textbox(value="motor scooter"), gr.Textbox(value="0.2183")],
outputs="label",
)
demo.launch()