Spaces:
Sleeping
Sleeping
| import streamlit as st | |
| from transformers import BlipProcessor, BlipForQuestionAnswering | |
| from PIL import Image | |
| import torch | |
| # Define model names | |
| original_model_name = "Salesforce/blip-vqa-base" | |
| finetuned_model_name = "manik063/blip-mmsd-sarcasm-no-reasoning" | |
| def load_models(): | |
| # Load both processors and models | |
| processor_orig = BlipProcessor.from_pretrained(original_model_name) | |
| model_orig = BlipForQuestionAnswering.from_pretrained(original_model_name) | |
| model_orig = model_orig.to('cuda' if torch.cuda.is_available() else 'cpu') | |
| model_orig.eval() | |
| processor_finetuned = BlipProcessor.from_pretrained(finetuned_model_name) | |
| model_finetuned = BlipForQuestionAnswering.from_pretrained(finetuned_model_name) | |
| model_finetuned = model_finetuned.to('cuda' if torch.cuda.is_available() else 'cpu') | |
| model_finetuned.eval() | |
| return (processor_orig, model_orig), (processor_finetuned, model_finetuned) | |
| # Load models | |
| (orig_processor, orig_model), (finetuned_processor, finetuned_model) = load_models() | |
| st.title("🖼️ Sarcasm Detection: Compare Original vs Fine-tuned BLIP Together") | |
| st.write("Upload an image and caption, and we will check sarcasm with BOTH models.") | |
| uploaded_file = st.file_uploader("Upload Image", type=["png", "jpg", "jpeg"]) | |
| caption = st.text_input("Enter Caption:") | |
| if uploaded_file and caption: | |
| if st.button("Detect Sarcasm 🚀"): | |
| image = Image.open(uploaded_file).convert("RGB") | |
| results = {} | |
| for model_name, processor, model in [ | |
| ("Original BLIP (Salesforce)", orig_processor, orig_model), | |
| ("Fine-tuned BLIP (Manik063)", finetuned_processor, finetuned_model) | |
| ]: | |
| prompt = f"Is this caption sarcastic? Answer strictly 'yes' or 'no': {caption}" | |
| inputs = processor(images=image, text=prompt, return_tensors="pt").to(model.device) | |
| with torch.no_grad(): | |
| outputs = model.generate(**inputs) | |
| generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0].lower() | |
| results[model_name] = generated_text | |
| st.image(image, caption="Uploaded Image", use_column_width=True) | |
| st.write(f"**Caption:** {caption}") | |
| # Display individual model outputs | |
| # for model_name, output in results.items(): | |
| # st.write(f"**{model_name} Output:** {output}") | |
| for model_name, output in results.items(): | |
| if isinstance(output, str): | |
| output = output.lower() | |
| print(output) | |
| print(type(output)) | |
| if "yes" in output or "1" in output: | |
| st.write(f"**{model_name} Output:** 🔴 Sarcasm Detected!") | |
| elif "no" in output or "0" in output: | |
| st.write(f"**{model_name} Output:** 🟢 Not Sarcastic.") | |
| else: | |
| st.write(f"**{model_name} Output:** ⚠️ Could not determine.") | |
| else: | |
| st.write(f"**{model_name} Output:** ⚠️ Invalid output format.") | |
| # Final decision logic | |
| # all_outputs = " ".join(results.values()) | |
| # if "yes" in all_outputs: | |
| # st.success("🔴 Sarcasm Detected by at least one model!") | |
| # elif "no" in all_outputs: | |
| # st.success("🟢 Not Sarcastic according to both models!") | |
| # else: | |
| # st.warning("⚠️ Models could not decide.") | |