app.py

import os

# ✅ Use /tmp for all cache & runtime folders (Hugging Face safe)
os.environ["STREAMLIT_HOME"] = "/tmp"
os.environ["STREAMLIT_RUNTIME_METRICS_ENABLED"] = "false"
os.environ["STREAMLIT_WATCHED_MODULES"] = ""
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
os.environ["HF_HOME"] = "/tmp/huggingface"

import streamlit as st
import torch
import joblib
import numpy as np
import random
from PIL import Image
from transformers import AutoTokenizer, AutoModel, ViTModel, ViTImageProcessor

# Use CPU only
device = torch.device("cpu")

# === Define Swahili VQA Model ===
class SwahiliVQAModel(torch.nn.Module):
    def __init__(self, num_answers):
        super().__init__()
        self.vision_encoder = ViTModel.from_pretrained('google/vit-base-patch16-224-in21k')
        self.text_encoder = AutoModel.from_pretrained("benjamin/roberta-base-wechsel-swahili")
        self.fusion = torch.nn.Sequential(
            torch.nn.Linear(768 + 768, 512),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.3),
            torch.nn.LayerNorm(512)
        )
        self.classifier = torch.nn.Linear(512, num_answers)

    def forward(self, image, input_ids, attention_mask):
        vision_outputs = self.vision_encoder(pixel_values=image)
        image_feats = vision_outputs.last_hidden_state[:, 0, :]
        text_outputs = self.text_encoder(input_ids=input_ids, attention_mask=attention_mask)
        text_feats = text_outputs.last_hidden_state[:, 0, :]
        combined = torch.cat([image_feats, text_feats], dim=1)
        fused = self.fusion(combined)
        return self.classifier(fused)

# === Load model and encoders ===
le = joblib.load("Vit_3895_label_encoder_best.pkl")
model = SwahiliVQAModel(num_answers=len(le.classes_)).to(device)

# Load full state dict (already trained classifier)
state_dict = torch.load("Vit_3895_best_model_epoch25.pth", map_location=device)
model.load_state_dict(state_dict)
model.eval()

tokenizer = AutoTokenizer.from_pretrained("benjamin/roberta-base-wechsel-swahili")
vit_processor = ViTImageProcessor.from_pretrained('google/vit-base-patch16-224-in21k')

# === Streamlit App ===
st.set_page_config(page_title="Swahili VQA", layout="wide")
st.title("🦜 Swahili Visual Question Answering App")
#st.markdown("**Pakia picha na uliza swali kwa Kiswahili**")
st.info("📥 Pakia picha na andika swali kisha bonyeza Tuma ili kupata jibu.")

uploaded_image = st.file_uploader("📂 Pakia picha hapa:", type=["jpg", "jpeg", "png"])

def generate_random_color():
    return f"rgb({random.randint(150, 255)}, {random.randint(80, 200)}, {random.randint(80, 200)})"

col1, col2 = st.columns([1, 2], gap="large")

with col1:
    if uploaded_image:
        st.image(uploaded_image, caption="Picha Iliyopakiwa")

with col2:
    st.markdown("<div style='padding-top: 15px;'>", unsafe_allow_html=True)
    question = st.text_input("💬 Andika swali lako hapa:")
    submit_button = st.button("📩 Tuma")
    st.markdown("</div>", unsafe_allow_html=True)

    if submit_button and uploaded_image and question:
        with st.spinner("🔍 Inachakata jibu..."):
            image = Image.open(uploaded_image).convert("RGB")
            image_tensor = vit_processor(images=image, return_tensors="pt")["pixel_values"]

            inputs = tokenizer(question, max_length=128, padding="max_length", truncation=True, return_tensors="pt")
            input_ids = inputs["input_ids"]
            attention_mask = inputs["attention_mask"]

            with torch.no_grad():
                logits = model(image_tensor, input_ids, attention_mask)
                probs = torch.softmax(logits, dim=1)

            top_probs, top_indices = torch.topk(probs, 5)
            decoded_answers = le.inverse_transform(top_indices.cpu().numpy()[0])

            results = [
                {"answer": ans, "confidence": round(prob * 100, 2)}
                for ans, prob in zip(decoded_answers, top_probs[0].tolist())
            ]

            results = sorted(results, key=lambda x: x["confidence"], reverse=True)
            st.subheader("🔎 Majibu Yanayowezekana:")

            max_confidence = max(result["confidence"] for result in results)
            for i, pred in enumerate(results):
                bar_width = (pred["confidence"] / max_confidence) * 70
                color = generate_random_color()
                st.markdown(
                    f"""
                    <div style="margin: 4px 0; padding: 2px 0; {'border-bottom: 1px solid rgba(150, 150, 150, 0.1);' if i < len(results)-1 else ''}">
                        <div style="font-size: 14px; font-weight: bold; margin-bottom: 2px;">
                            {pred['answer']}
                        </div>
                        <div style="display: flex; align-items: center; gap: 6px;">
                            <div style="width: {bar_width}%; height: 8px; border-radius: 3px; background: {color};"></div>
                            <div style="font-size: 13px; min-width: 45px;">
                                {pred['confidence']}%
                            </div>
                        </div>
                    </div>
                    """,
                    unsafe_allow_html=True
                )
                
    else:
        pass  # or just remove the else block entirely

# === Sidebar Footer Description (ALWAYS VISIBLE) ===
st.sidebar.markdown("""
---
## Swahili VQA App

This app allows users to ask questions about images in **Swahili**. Powered by a multimodal AI model trained on visual and textual data.

## 🛠 How to Use

1. 📤 Upload an image.  
2. 💬 Type a question in Swahili.  
3. 📩 Click **Tuma**.  
4. 🤖 The model will predict top 5 possible answers with confidence score.

## 📌 Note
🎓 Designed for educational and research purposes
""")