Delete app.py

app.py

@@ -1,201 +0,0 @@
-"""
-Chest X-ray Recommender - HuggingFace Space entry point.
-
-Loads pre-computed CLIP embeddings (embeddings.parquet, built by the
-companion notebook) and serves a Gradio UI that returns the 3 most
-visually similar X-rays for a given text or image query.
-
-Educational demo only. NOT a medical device.
-"""
-from __future__ import annotations
-
-import base64
-import io
-import os
-
-import gradio as gr
-import numpy as np
-import pandas as pd
-import torch
-from PIL import Image
-from transformers import CLIPModel, CLIPProcessor
-
-# ---------------------------------------------------------------------------
-# Config
-# ---------------------------------------------------------------------------
-MODEL_ID        = os.environ.get("CLIP_MODEL_ID", "openai/clip-vit-base-patch32")
-EMBEDDINGS_FILE = os.environ.get("EMBEDDINGS_FILE", "embeddings.parquet")
-TOP_K           = 3
-
-# Optional: a YouTube embed for your walk-through video. Replace the id.
-VIDEO_EMBED_ID  = os.environ.get("VIDEO_EMBED_ID", "")
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-print(f"[startup] device = {device}")
-
-# ---------------------------------------------------------------------------
-# Load model
-# ---------------------------------------------------------------------------
-print(f"[startup] loading CLIP model: {MODEL_ID}")
-clip_model     = CLIPModel.from_pretrained(MODEL_ID).to(device).eval()
-clip_processor = CLIPProcessor.from_pretrained(MODEL_ID)
-
-# ---------------------------------------------------------------------------
-# Load catalog (embeddings + thumbnails + reports)
-# ---------------------------------------------------------------------------
-print(f"[startup] loading catalog: {EMBEDDINGS_FILE}")
-df = pd.read_parquet(EMBEDDINGS_FILE)
-print(f"[startup] catalog rows: {len(df):,}")
-
-EMB_MATRIX = np.vstack(df["embedding"].values).astype("float32")
-# Make sure the catalog vectors are L2-normalised (cheap, idempotent)
-norms = np.linalg.norm(EMB_MATRIX, axis=1, keepdims=True)
-EMB_MATRIX = EMB_MATRIX / np.where(norms == 0, 1, norms)
-
-def _b64_to_pil(b64: str) -> Image.Image:
-    img = Image.open(io.BytesIO(base64.b64decode(b64)))
-    img.load()                   # force-load before BytesIO is garbage-collected
-    return img.convert("RGB")     # RGB renders more reliably in Gradio than "L"
-
-THUMBS  = [_b64_to_pil(b) for b in df["image_b64"]]
-REPORTS = df["report"].fillna("").tolist()
-CLUSTER = df["cluster"].astype(int).tolist() if "cluster" in df.columns else [0] * len(df)
-
-
-# ---------------------------------------------------------------------------
-# Embedding helpers
-# ---------------------------------------------------------------------------
-def _to_tensor(out):
-    """Unwrap a CLIP output that may be a tensor or a HuggingFace ModelOutput."""
-    if torch.is_tensor(out):
-        return out
-    if hasattr(out, "image_embeds"):
-        return out.image_embeds
-    if hasattr(out, "text_embeds"):
-        return out.text_embeds
-    if hasattr(out, "pooler_output"):
-        return out.pooler_output
-    if hasattr(out, "last_hidden_state"):
-        return out.last_hidden_state[:, 0]
-    raise TypeError(f"Cannot unwrap CLIP output of type {type(out)}")
-
-
-@torch.no_grad()
-def _embed_image(pil_img: Image.Image) -> np.ndarray:
-    inputs = clip_processor(images=pil_img.convert("RGB"), return_tensors="pt").to(device)
-    vision_out = clip_model.vision_model(pixel_values=inputs["pixel_values"])
-    pooled = _to_tensor(vision_out)
-    emb = clip_model.visual_projection(pooled)
-    emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
-    return emb.cpu().numpy()[0]
-
-
-@torch.no_grad()
-def _embed_text(text: str) -> np.ndarray:
-    inputs = clip_processor(
-        text=[text], return_tensors="pt",
-        padding=True, truncation=True, max_length=77,
-    ).to(device)
-    text_out = clip_model.text_model(
-        input_ids=inputs["input_ids"],
-        attention_mask=inputs.get("attention_mask"),
-    )
-    pooled = _to_tensor(text_out)
-    emb = clip_model.text_projection(pooled)
-    emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
-    return emb.cpu().numpy()[0]
-
-
-def _top_k(query_vec: np.ndarray, k: int = TOP_K):
-    scores = EMB_MATRIX @ query_vec.astype("float32")
-    idx = np.argsort(-scores)[:k]
-    return [
-        {
-            "index"   : int(i),
-            "score"   : float(scores[i]),
-            "image"   : THUMBS[i],
-            "report"  : REPORTS[i],
-            "cluster" : CLUSTER[i],
-        }
-        for i in idx
-    ]
-
-
-# ---------------------------------------------------------------------------
-# Gradio handler
-# ---------------------------------------------------------------------------
-def recommend(text_query: str, image_query: Image.Image | None):
-    if image_query is not None:
-        q = _embed_image(image_query)
-        used = "image"
-    elif text_query and text_query.strip():
-        q = _embed_text(text_query.strip())
-        used = "text"
-    else:
-        return [], "Please type a description **or** upload an X-ray."
-
-    results = _top_k(q, k=TOP_K)
-    gallery = [(r["image"], f"#{r['index']}  -  score {r['score']:.3f}") for r in results]
-    details = "\n\n".join(
-        f"### Match {n+1}  (score {r['score']:.3f}, cluster {r['cluster']})\n"
-        f"{r['report'][:600]}{'...' if len(r['report']) > 600 else ''}"
-        for n, r in enumerate(results)
-    )
-    details = f"_Query used: **{used}**_\n\n" + details
-    return gallery, details
-
-
-# ---------------------------------------------------------------------------
-# UI
-# ---------------------------------------------------------------------------
-DESCRIPTION = """
-# Chest X-ray Recommender  -  CLIP embeddings
-
-Upload a chest X-ray **or** describe a finding in words, and the app will
-retrieve the 3 most visually similar X-rays from a pre-computed catalog
-drawn from `MLforHealthcare/mimic-cxr`.
-
-> Educational demo only. **Not** a medical device. Do not use for clinical decisions.
-"""
-
-with gr.Blocks(title="Chest X-ray Recommender") as demo:
-    gr.Markdown(DESCRIPTION)
-
-    with gr.Row():
-        with gr.Column(scale=1):
-            text_in = gr.Textbox(
-                lines=2,
-                label="Describe a finding (English)",
-                placeholder='e.g. "bilateral pleural effusion with cardiomegaly"',
-            )
-            image_in = gr.Image(type="pil", label="...or upload an X-ray")
-            btn = gr.Button("Recommend", variant="primary")
-            gr.Examples(
-                examples=[
-                    ["bilateral pleural effusion with cardiomegaly", None],
-                    ["clear lungs, no acute cardiopulmonary process", None],
-                    ["right lower lobe pneumonia", None],
-                    ["pneumothorax", None],
-                ],
-                inputs=[text_in, image_in],
-            )
-        with gr.Column(scale=2):
-            gallery = gr.Gallery(label="Top-3 similar X-rays", columns=3, height=350)
-            details = gr.Markdown()
-
-    btn.click(recommend, inputs=[text_in, image_in], outputs=[gallery, details])
-
-    if VIDEO_EMBED_ID:
-        gr.HTML(
-            f"""
-            <h3>Walk-through video</h3>
-            <iframe width="720" height="405"
-                    src="https://www.youtube.com/embed/{VIDEO_EMBED_ID}"
-                    title="Assignment 3 walk-through" frameborder="0"
-                    allow="autoplay; encrypted-media; picture-in-picture" allowfullscreen>
-            </iframe>
-            """
-        )
-
-if __name__ == "__main__":
-    demo.launch()