app.py

#!/usr/bin/env python3
# from doctest import OutputChecker
# import sys
# import torch
# import re
# import os
# import gradio as gr
# import requests
# 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
# from huggingface_hub import login

#!/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 torch.nn.functional import softmax
import numpy as np

from transformers import AutoTokenizer, AutoModelForCausalLM
#from torch.nn.functional import softmax

from huggingface_hub import login


#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 get_sim(x):
    x =  str(x)[1:-1]
    x =  str(x)[1:-1]
    return x
     



import os
#print(os.getenv('HF_token'))
hf_api_token = os.getenv("HF_token")  # For sensitive secrets
#app_mode = os.getenv("APP_MODE")  # For public variables


access_token = hf_api_token
#print(login(token = access_token))


tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")






#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_G, caption_B, caption_VR, visual_context_label, visual_context_prob):
    caption_G = caption_G
    caption_B = caption_B
    caption_VR = caption_VR
    visual_context_label= visual_context_label
    visual_context_prob = visual_context_prob
    caption_emb_G = model_sts.encode(caption_G, convert_to_tensor=True)
    caption_emb_B = model_sts.encode(caption_B, convert_to_tensor=True)
    caption_emb_VR = model_sts.encode(caption_VR, convert_to_tensor=True)

    visual_context_label_emb = model_sts.encode(visual_context_label, convert_to_tensor=True)


    sim_1 =  cosine_scores = util.pytorch_cos_sim(caption_emb_G, visual_context_label_emb)
    sim_1 = sim_1.cpu().numpy()
    sim_1 = get_sim(sim_1)

    sim_2 = cosine_scores = util.pytorch_cos_sim(caption_emb_B, visual_context_label_emb)
    sim_2 = sim_2.cpu().numpy()
    sim_2 = get_sim(sim_2)

    sim_3 = cosine_scores = util.pytorch_cos_sim(caption_emb_VR, visual_context_label_emb)
    sim_3 = sim_3.cpu().numpy()
    sim_3 = get_sim(sim_3)
 

    LM_1 = sentence_prob_mean(caption_G)
    LM_2 = sentence_prob_mean(caption_B)
    LM_3 = sentence_prob_mean(caption_VR)

    #LM  = scorer.sentence_score(caption, reduce="mean")
    score_1 = pow(float(LM_1),pow((1-float(sim_1))/(1+ float(sim_1)),1-float(visual_context_prob)))
    score_2 = pow(float(LM_2),pow((1-float(sim_2))/(1+ float(sim_2)),1-float(visual_context_prob)))
    score_3 = pow(float(LM_3),pow((1-float(sim_3))/(1+ float(sim_3)),1-float(visual_context_prob)))

    #return {"LM": float(LM)/1, "sim": float(sim)/1, "score": float(score)/1 }
    return {"Greedy": float(score_1)/1, "Best-Beam-5": float(score_2)/1, "Visual_re-Ranker": float(score_3)/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",
    description="Demo for 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")],
    # a baby is eating in front of a birthday cake /a baby sitting in front of a giant cake 
    inputs=[gr.Textbox(value="baby is eating in front of a birthday cake") , gr.Textbox(value="a baby sitting in front of a cake"), gr.Textbox(value="a baby sitting in front of a birthday cake"), gr.Textbox(value="candle wax light"),  gr.Textbox(value="0.958")],
    #outputs=[gr.Textbox(value="Language Model Score") , gr.Textbox(value="Semantic Similarity Score"),  gr.Textbox(value="Belief revision score via visual context")],
    outputs="label",
)

demo.launch()