app.py

import os
os.system("pip install fairseq2 --extra-index-url https://fair.pkg.atmeta.com/fairseq2/whl/pt2.6.0/cu124 -q")

from huggingface_hub import hf_hub_download
import gradio as gr
import torch
import requests
from PIL import Image
from transformers import SiglipImageProcessor, SiglipVisionModel
from sonar.inference_pipelines.text import TextToEmbeddingModelPipeline
import torch.nn as nn
import torch.nn.functional as F
from io import BytesIO
from transformers.image_utils import load_image

cos = nn.CosineSimilarity()

model_path = hf_hub_download(
    repo_id="Sibgat-Ul/SONAR-Image_enc",
    filename="best_sonar.pth",
    repo_type="model"
)

language_mapping = {
    "English": "eng_Latn",
    "Bengali": "ben_Beng",
    "French": "fra_Latn"
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# -------- Load Image Encoder --------
class SonarImageEnc(nn.Module):
    def __init__(self, path="google/siglip2-base-patch16-384", initial_temperature=0.07):
        super().__init__()
        
        self.model = SiglipVisionModel.from_pretrained(path, torch_dtype="auto")
        
        for param in self.model.parameters():
            param.requires_grad = False
        
        self.projection = nn.Sequential(
            nn.Linear(self.model.config.hidden_size, 2048),
            nn.GELU(),
            nn.Dropout(0.1),
            nn.Linear(2048, 1024), 
            nn.LayerNorm(1024, eps=1e-5),
        )

        for param in self.projection.parameters():
            param.requires_grad = True

        self.temp_s = nn.Parameter(torch.log(torch.tensor(10.0)))
        self.bias = nn.Parameter(torch.tensor(-10.0))
    
        self.logit_scale = nn.Parameter(torch.ones([]) * torch.log(torch.tensor(1.0) / initial_temperature))
        
    def forward(self, pixel_values):
        vision_outputs = self.model(pixel_values=pixel_values)
        pooled_output = vision_outputs.pooler_output
        embeddings = self.projection(pooled_output)
        
        self.logit_scale.data.clamp_(
            min=torch.log(torch.tensor(1.0).to(device) / torch.tensor(0.001).to(device)),
            max=torch.log(torch.tensor(1.0).to(device) / torch.tensor(100.0).to(device))
        )

        return embeddings, torch.exp(self.logit_scale), torch.exp(self.temp_s), self.bias

# Load processor and models
processor = SiglipImageProcessor.from_pretrained("google/siglip2-base-patch16-384")

t2t_model_emb = TextToEmbeddingModelPipeline(
    encoder="text_sonar_basic_encoder",
    tokenizer="text_sonar_basic_encoder",
    device=device,
    dtype=torch.float16,
)

img_encoder = SonarImageEnc().to(device).eval()
img_encoder.load_state_dict(torch.load(model_path, map_location=device))

# -------- Similarity Scoring --------
def compute_similarity(
    image, image_url, 
    option_a, option_b, option_c, option_d,
    lang_opt_a, lang_opt_b, lang_opt_c, lang_opt_d
):
    
    if not image:
        try:
            headers = {
                "User-Agent": "Mozilla/5.0"
            }
            response = requests.get(image_url, headers=headers)
            response.raise_for_status()
            image = Image.open(BytesIO(response.content)).convert("RGB")
        except Exception as e:
            return None, {"Error": f"Image could not be loaded: {str(e)}"}

    # Preprocess image
    inputs = processor(image, return_tensors="pt").to(device)
    
    with torch.no_grad():
        image_emb, _, _, _ = img_encoder(inputs.pixel_values)
        image_emb = image_emb.to(device, torch.float16)

    # Map languages
    lang_codes = [
        language_mapping[lang_opt_a],
        language_mapping[lang_opt_b],
        language_mapping[lang_opt_c],
        language_mapping[lang_opt_d],
    ]
    texts = [option_a, option_b, option_c, option_d]

    # Get embeddings per option with corresponding language
    text_embeddings = []
    for text, lang in zip(texts, lang_codes):
        emb = t2t_model_emb.predict([text], source_lang=lang)
        text_embeddings.append(emb)

    text_embeddings = torch.cat(text_embeddings, dim=0).to(device)

    scores = cos(image_emb, text_embeddings)

    results = {
        f"Option {chr(65+i)}": round(score.item(), 3)
        for i, score in enumerate(scores)
    }

    results = {
        k: f"{round(v * 100, 2)}%"
        for k, v in sorted(results.items(), key=lambda item: item[1], reverse=True)
    }

    return image, results


# -------- Gradio UI --------
with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("## 🔍 SONAR: Image-Text Similarity Scorer")
    gr.Markdown("#### Upload an Image or provide an URL.")
    
    with gr.Row():
        with gr.Column():
            image_url = gr.Textbox(label="Image URL", value="http://images.cocodataset.org/val2017/000000039769.jpg")

            with gr.Row():
                option_a = gr.Textbox(label="Option A", value="Two cats in a bed.")
                lang_opt_a = gr.Dropdown(choices=list(language_mapping.keys()), value="English", label="Language")
                    
            with gr.Row():    
                option_b = gr.Textbox(label="Option B", value="Two cat with two remotes.")
                lang_opt_b = gr.Dropdown(choices=list(language_mapping.keys()), value="English", label="Language")
                
            with gr.Row():
                option_c = gr.Textbox(label="Option C", value="Two remotes.")
                lang_opt_c = gr.Dropdown(choices=list(language_mapping.keys()), value="English", label="Language")
                
            with gr.Row():
                option_d = gr.Textbox(label="Option D", value="Two cats.")
                lang_opt_d = gr.Dropdown(choices=list(language_mapping.keys()), value="English", label="Language")
                
                # language = gr.Dropdown(choices=list(language_mapping.keys()), value="English", label="Select Language")

        with gr.Column():
            image_input = gr.Image(label="Upload an image", type="pil")
            btn = gr.Button("Done")

            with gr.Row():
                img_output = gr.Image(label="Input Image", type="pil", width=300, height=300)
                result_output = gr.JSON(label="Similarity Scores")

    btn.click(
        fn=compute_similarity,
        inputs=[
            image_input, image_url,
            option_a, option_b, option_c, option_d,
            lang_opt_a, lang_opt_b, lang_opt_c, lang_opt_d
        ],
        outputs=[img_output, result_output]
    )


demo.launch()