app.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import os
import torch
import torch.nn.functional as F
import gradio as gr
import numpy as np
import torchvision.transforms as T
from PIL import Image

from lib.framework import create_model
from lib.options import ParamSet, _retrieve_parameter, _dispatch_by_group
from lib.dataloader import ImageMixin

# ===========================================
# 1) パス設定
# ===========================================
WEIGHT_PATH    = "./cxp_projection_rotation.pt"
PARAMETER_JSON = "./parameters.json"

# ===========================================
# 2) クラスラベル定義
# ===========================================
LABEL_APorPA = ["AP", "PA", "Lateral"]
LABEL_ROUND  = ["Upright", "Inverted", "Left-rotation", "Right-rotation"]

# ===========================================
# 3) 前処理クラス
# ===========================================
class ImageHandler(ImageMixin):
    def __init__(self, params):
        self.params = params
        self.transform = T.Compose([
            T.ToTensor(),
        ])

    def set_image(self, image: Image.Image):
        tensor = self.transform(image)            # [C,H,W], float32 in [0,1]
        return {"image": tensor.unsqueeze(0)}     # バッチ次元追加

# ===========================================
# 4) パラメータロード
# ===========================================
def load_parameter(parameter_path):
    _args = ParamSet()
    params = _retrieve_parameter(parameter_path)
    for k, v in params.items():
        setattr(_args, k, v)
    # 推論用に上書き
    _args.augmentation = "no"
    _args.sampler      = "no"
    _args.pretrained   = False
    _args.mlp          = None
    _args.net          = _args.model
    _args.device       = torch.device("cpu")
    return (
        _dispatch_by_group(_args, "model"),
        _dispatch_by_group(_args, "dataloader"),
    )

args_model, args_dataloader = load_parameter(PARAMETER_JSON)

# ===========================================
# 5) モデル作成＆重みロード
# ===========================================
model = create_model(args_model)
print(f"Loading weights from {WEIGHT_PATH}")
model.load_weight(WEIGHT_PATH)
model.eval()

# ===========================================
# 6) 推論＋HTML生成
# ===========================================
def predict_html(image_path: str) -> str:
    # 画像読み込み
    img = Image.open(image_path).convert("L")
    handler = ImageHandler(args_dataloader)
    batch = handler.set_image(img)

    with torch.no_grad():
        outputs = model(batch)
        logits_proj = outputs.get("label_APorPA")
        logits_rot  = outputs.get("label_round")

        # softmax で確率に変換
        probs_proj = F.softmax(logits_proj, dim=1)[0].cpu().numpy()
        probs_rot  = F.softmax(logits_rot,  dim=1)[0].cpu().numpy()

    # argmax でラベル選択
    idx_proj = int(probs_proj.argmax())
    idx_rot  = int(probs_rot.argmax())
    pred_proj = LABEL_APorPA[idx_proj]
    pred_rot  = LABEL_ROUND[idx_rot]
    conf_proj = float(probs_proj[idx_proj])
    conf_rot  = float(probs_rot[idx_rot])

    # ファイル名から元ラベル取得（例: "1_AP_Upright.png" → orig_proj="AP", orig_rot="Upright"）
    base = os.path.splitext(os.path.basename(image_path))[0]
    parts = base.split("_")
    if len(parts) >= 3:
        orig_proj = parts[1]
        orig_rot  = parts[2]
    else:
        orig_proj = orig_rot = None

    # 警告HTML作成用ヘルパー
    def make_warning(kind, orig, pred, conf):
        # kind: "projection" or "rotation"
        high_thr = 0.8
        med_thr  = 0.5
        if orig and orig != pred:
            if conf >= high_thr:
                return (
                    f"<p style='color:red'>⚠ Potentially mislabeled {kind}: "
                    f"filename says {orig}, model predicts {pred} (confidence {conf:.2f})</p>"
                )
            elif conf >= med_thr:
                return (
                    f"<p style='color:orange'>⚠ There is a possibility of mislabeled {kind}: "
                    f"model predicts {pred} with moderate confidence ({conf:.2f})</p>"
                )
        if conf < med_thr:
            return (
                f"<p style='color:orange'>⚠ Low confidence for {kind} ({conf:.2f}); "
                f"please check image quality or framing.</p>"
            )
        return ""

    # 警告HTML
    warn_html = ""
    warn_html += make_warning("projection", orig_proj, pred_proj, conf_proj)
    warn_html += make_warning("rotation",   orig_rot,  pred_rot,  conf_rot)

    # クラスごとのスコア表示用HTML
    scores_proj = ", ".join(
        f"{LABEL_APorPA[i]}: {p:.2f}" for i, p in enumerate(probs_proj)
    )
    scores_rot = ", ".join(
        f"{LABEL_ROUND[i]}: {p:.2f}" for i, p in enumerate(probs_rot)
    )

    # 結果表示用HTML
    html = (
        f"<p><strong>Projection :</strong> {pred_proj} "
        f"<small>({scores_proj})</small></p>"
        f"<p><strong>Rotation   :</strong> {pred_rot} "
        f"<small>({scores_rot})</small></p>"
        f"{warn_html}"
    )
    return html

# ===========================================
# 7) Gradio UI
# ===========================================
html_header = """
<div style="padding:10px;border:1px solid #ddd;border-radius:5px">
  <h2>Chest X‑ray Projection & Rotation Classification</h2>
  <p>Upload a 256×256 grayscale PNG. The model predicts projection (AP/PA/Lateral)
     and rotation (Upright/Inverted/Left/Right) and shows softmax confidences. 
     It warns if filename label differs or if confidence is low. Please name the files using the format: [Number]_projection_rotation.png.
     For the projection part of the filename, please use one of the following three terms: AP/PA/Lateral
     For the rotation part of the filename, please use one of the following four terms: Upright/Inverted/Left90/Right90
     As samples, We have prepared two sets of images: A PA view in the Upright position. An AP view with Left rotation. For each set, we have created two versions, one of which includes a mislabel in either its projection or rotation tag.</p>
</div>
"""

with gr.Blocks(title="CXR Projection & Rotation") as demo:
    gr.HTML(html_header)

    with gr.Row():
        input_image = gr.Image(
            label="Upload PNG (256×256)", 
            type="filepath", 
            image_mode="L"
        )
        output_html = gr.HTML()

    send_btn = gr.Button("Run Inference")
    send_btn.click(
        fn=predict_html,
        inputs=input_image,
        outputs=output_html
    )

    # サンプル例
    gr.Examples(
        examples=[
            "./sample/1_AP_Upright.png",
            "./sample/1_PA_Inverted.png",
            "./sample/2_AP_Right-rotation.png",
            "./sample/2_Lateral_Left-rotation.png",
        ],
        inputs=input_image
    )

    # サンプルのファイル名を一覧で表示
    gr.Markdown(
        "**Sample filenames:**  𝚮\n"
        "- 1_AP_Upright.png  \n"
        "- 1_PA_Inverted.png  \n"
        "- 2_AP_Right90.png  \n"
        "- 2_Lateral_Left90.png"
    )

if __name__ == "__main__":
    demo.launch(debug=True)