app.py

#!/usr/bin/env python3
from doctest import OutputChecker
import sys
import torch
import re
import os
import gradio as gr
import requests
import torch
from transformers import GPT2Tokenizer, GPT2LMHeadModel
from torch.nn.functional import softmax
import numpy as np

#url = "https://github.com/simonepri/lm-scorer/tree/master/lm_scorer/models"
#resp = requests.get(url)

from sentence_transformers import SentenceTransformer, util
#from sentence_transformers import SentenceTransformer, util
#from sklearn.metrics.pairwise import cosine_similarity
#from lm_scorer.models.auto import AutoLMScorer as LMScorer
#from sentence_transformers import SentenceTransformer, util
#from sklearn.metrics.pairwise import cosine_similarity

#device = "cuda:0" if torch.cuda.is_available() else "cpu"
#model_sts = gr.Interface.load('huggingface/sentence-transformers/stsb-distilbert-base') 

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


	
# Load pre-trained model 

# model = GPT2LMHeadModel.from_pretrained('distilgpt2', 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')

tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = GPT2LMHeadModel.from_pretrained('gpt2')


def sentence_prob_mean(text):
    # Tokenize the input text and add special tokens
    input_ids = tokenizer.encode(text, return_tensors='pt')

    # Obtain model outputs
    with torch.no_grad():
        outputs = model(input_ids, labels=input_ids)
        logits = outputs.logits  # logits are the model outputs before applying softmax

    # Shift logits and labels so that tokens are aligned:
    shift_logits = logits[..., :-1, :].contiguous()
    shift_labels = input_ids[..., 1:].contiguous()

    # Calculate the softmax probabilities
    probs = softmax(shift_logits, dim=-1)

    # Gather the probabilities of the actual token IDs
    gathered_probs = torch.gather(probs, 2, shift_labels.unsqueeze(-1)).squeeze(-1)

    # Compute the mean probability across the tokens
    mean_prob = torch.mean(gathered_probs).item()

    return mean_prob




def cos_sim(a, b):
    return np.inner(a, b) / (np.linalg.norm(a) * (np.linalg.norm(b)))


  
def Visual_re_ranker(caption, visual_context_label, visual_context_prob):
    caption = caption 
    visual_context_label= visual_context_label
    visual_context_prob = visual_context_prob
    caption_emb = model_sts.encode(caption, convert_to_tensor=True)
    visual_context_label_emb = model_sts.encode(visual_context_label, convert_to_tensor=True)


    sim =  cosine_scores = util.pytorch_cos_sim(caption_emb, visual_context_label_emb)
    sim = sim.cpu().numpy()
    sim = str(sim)[1:-1]
    sim = str(sim)[1:-1] 

   # LM = cloze_prob(caption)
    LM = sentence_prob_mean(caption)
    #LM  = scorer.sentence_score(caption, reduce="mean")
    score = pow(float(LM),pow((1-float(sim))/(1+ float(sim)),1-float(visual_context_prob)))
    

    #return {"LM": float(LM)/1, "sim": float(sim)/1, "score": float(score)/1 }
    return {"init hypothesis": float(LM)/1, "Visual Belief Revision": float(score)/1 }
    #return LM, sim, score 



demo = gr.Interface(
    fn=Visual_re_ranker,
    description="Demo for Belief Revision based Caption Re-ranker with Visual Semantic Information",
    inputs=[gr.Textbox(value="a city street filled with traffic at night") , gr.Textbox(value="traffic"),  gr.Textbox(value="0.7458009")],
    #outputs=[gr.Textbox(value="Language Model Score") , gr.Textbox(value="Semantic Similarity Score"),  gr.Textbox(value="Belief revision score via visual context")],
    outputs="label",
    share=True,
)
demo.launch()