src/streamlit_app.py

# src/streamlit_app.py

# ─── 1) Force all HF/Streamlit temp dirs into /tmp ─────────────────────────────
import os
for k,v in [
    ("HOME",               "/tmp"),
    ("XDG_CONFIG_HOME",    "/tmp"),
    ("STREAMLIT_HOME",     "/tmp"),
    ("XDG_CACHE_HOME",     "/tmp"),
    ("HF_HOME",            "/tmp/hf"),
    ("TRANSFORMERS_CACHE", "/tmp/hf/transformers"),
]:
    os.environ[k] = v

# create those dirs
for d in ["/tmp/streamlit", "/tmp/hf/transformers"]:
    os.makedirs(d, exist_ok=True)


# ─── 2) Imports ────────────────────────────────────────────────────────────────
import streamlit as st
from PIL import Image
import torch
import torch.nn as nn
import torchvision.transforms as T
from transformers import ViTModel, T5ForConditionalGeneration, T5Tokenizer


# ─── 3) Your model repo & subfolders ───────────────────────────────────────────
MODEL_ID = "RakeshNJ12345/Automated-Chest-XRay-Report"


@st.cache_resource(show_spinner="Loading fine-tuned ViT & T5…")
def load_models():
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # 1) load fine-tuned ViT from vit/
    vit = ViTModel.from_pretrained(
        MODEL_ID,
        subfolder="vit",
        ignore_mismatched_sizes=True
    ).to(device)

    # 2) load fine-tuned T5 + tokenizer from T5/
    t5    = T5ForConditionalGeneration.from_pretrained(
                MODEL_ID, subfolder="T5"
            ).to(device)
    tok   = T5Tokenizer.from_pretrained(
                MODEL_ID, subfolder="T5"
            )

    # 3) build a single Linear projection so we reuse it each call
    proj = nn.Linear(vit.config.hidden_size, t5.config.d_model).to(device)

    return device, vit, proj, t5, tok


# ─── 4) Kick off model loading ─────────────────────────────────────────────────
device, vit, proj, t5, tokenizer = load_models()


# ─── 5) Image preprocessing ────────────────────────────────────────────────────
transform = T.Compose([
    T.Resize((224, 224)),
    T.ToTensor(),
    T.Normalize(mean=0.5, std=0.5),
])


# ─── 6) Streamlit UI ───────────────────────────────────────────────────────────
st.set_page_config(page_title="Radiology Report Analysis", layout="wide")

st.markdown(
    "<h1 style='text-align:center;'>🩺 Radiology Report Analysis</h1>",
    unsafe_allow_html=True
)
st.markdown(
    "<p style='text-align:center;'>Upload a chest X-ray and click <b>Generate Report</b>.</p>",
    unsafe_allow_html=True
)

if "stage" not in st.session_state:
    st.session_state.stage = "upload"


# ─── 7) UPLOAD SCREEN ──────────────────────────────────────────────────────────
if st.session_state.stage == "upload":
    uploaded = st.file_uploader(
        "📤 Upload Chest X-ray (PNG/JPG)", 
        type=["png","jpg","jpeg"],
        label_visibility="visible"
    )
    if uploaded:
        st.image(uploaded, width=350,
                 caption=f"{uploaded.name} — {uploaded.size/1e6:.2f} MB")
        if st.button("▶️ Generate Report"):
            st.session_state.uploaded = uploaded
            st.session_state.stage    = "report"
            st.experimental_rerun()


# ─── 8) REPORT SCREEN ──────────────────────────────────────────────────────────
elif st.session_state.stage == "report":
    img_file = st.session_state.uploaded
    img      = Image.open(img_file).convert("RGB")

    with st.spinner("🔍 Analyzing…"):
        # 1) ViT features
        x       = transform(img).unsqueeze(0).to(device)      # [1,3,224,224]
        vit_out = vit(pixel_values=x).pooler_output           # [1,768]

        # 2) project → prefix
        vision_pref = proj(vit_out).unsqueeze(1)              # [1,1,d_model]

        # 3) prepare prompt tokens
        enc     = tokenizer("report:", return_tensors="pt").to(device)
        txt_emb = t5.encoder.embed_tokens(enc.input_ids)     # [1,L,d_model]

        # 4) concat embeddings + mask
        enc_emb  = torch.cat([vision_pref, txt_emb], dim=1)   # [1,1+L,d]
        enc_mask = torch.cat([
            torch.ones(1,1,device=device,dtype=torch.long),
            enc.attention_mask
        ], dim=1)                                            # [1,1+L]

        # 5) feed into T5 encoder
        enc_out = t5.encoder(
            inputs_embeds   = enc_emb,
            attention_mask  = enc_mask
        )

        # 6) generate **without** encoder_attention_mask kwarg
        out_ids = t5.generate(
            encoder_outputs=enc_out,
            max_length     = 64,
            num_beams      = 2,
            do_sample      = False,
            eos_token_id   = tokenizer.eos_token_id,
        )
        report = tokenizer.decode(out_ids[0], skip_special_tokens=True)

    # ── display side by side ────────────────────────────────────────────────
    c1, c2 = st.columns(2)
    with c1:
        st.subheader("📤 Uploaded X-ray")
        st.image(img, use_container_width=True)
        st.markdown(f"**Filename:** {img_file.name}")
        st.markdown(f"**Size:** {img_file.size/1e6:.2f} MB")
    with c2:
        st.subheader("📝 AI Diagnosis & Report")
        st.markdown(
            f"<div style='background:#e0f7fa;padding:12px;border-radius:6px;'>{report}</div>",
            unsafe_allow_html=True
        )
        if st.button("⬅️ Upload Another"):
            del st.session_state.uploaded
            st.session_state.stage = "upload"
            st.experimental_rerun()


# ─── Footer ────────────────────────────────────────────────────────────────────
st.markdown("""
<hr>
<p style='text-align:center;color:gray;font-size:0.85em;'>
  ⚙️ Powered by fine-tuned ViT + T5 · Built with Streamlit · Hosted on HF Spaces
</p>
""", unsafe_allow_html=True)