Upload 4 files

app.py

@@ -0,0 +1,160 @@
+import torch
+import pandas as pd
+import numpy as np
+import matplotlib.pyplot as plt
+import seaborn as sns
+import gradio as gr
+import os
+
+# 從 model.py 匯入架構
+from model import DualStreamTransformer, ArcMarginProduct
+
+css = """
+.scroll-box {
+    height: 300px;
+    overflow-y: auto !important;
+    overflow-x: hidden !important;
+    display: block !important;
+    width: 100% !important;
+    max-width: 100% !important;
+}
+.scroll-box * {
+    max-width: 100% !important;
+    box-sizing: border-box !important;
+}
+.vertical-radio {
+    display: block !important;
+    width: 100% !important;
+}
+.vertical-radio .wrap {
+    display: flex !important;
+    flex-direction: column !important;
+    width: 100% !important;
+    min-width: 0 !important;
+}
+.vertical-radio .gradio-radio-item {
+    width: 100% !important;
+    white-space: normal !important;
+    word-break: break-all !important;
+}
+"""
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+MODEL_PATH = "./best_model_fold_5.pt"
+
+model = DualStreamTransformer(n_feat1=25, n_feat2=12, d_model=32).to(DEVICE)
+metric_fc = ArcMarginProduct(32, 2).to(DEVICE)
+
+if os.path.exists(MODEL_PATH):
+    checkpoint = torch.load(MODEL_PATH, map_location=DEVICE)
+    if isinstance(checkpoint, dict) and 'model' in checkpoint:
+        model.load_state_dict(checkpoint['model'])
+        metric_fc.load_state_dict(checkpoint['fc'])
+    else:
+        model.load_state_dict(checkpoint)
+    model.eval()
+    print("模型載入成功！")
+
+# ==========================================
+# 邏輯函式
+# ==========================================
+def analyze_and_predict(*all_answers):
+    if any(a is None for a in all_answers):
+        raise gr.Error("請完整填寫所有問卷題目！")
+
+    ccmq_map = {"總是": 5, "經常": 4, "有時": 3, "很少": 2, "沒有": 1}
+    osdi_map = {"總是": 4, "經常": 3, "一半一半": 2, "偶而": 1, "完全不曾": 0}
+    
+    # 資料處理
+    x1 = torch.tensor([[ccmq_map[a] for a in all_answers[:25]]], dtype=torch.float32).to(DEVICE)
+    x2 = torch.tensor([[osdi_map[a] for a in all_answers[25:]]], dtype=torch.float32).to(DEVICE)
+
+    with torch.no_grad():
+        feats = model(x1, x2)
+        logits = metric_fc.predict(feats)
+        probs = torch.softmax(logits, dim=1)
+        pred_idx = torch.argmax(probs, dim=1).item()
+        conf = probs[0, pred_idx].item()
+
+    # 繪圖展示 (研討會風格)
+    plt.rcParams['font.sans-serif'] = ['Microsoft JhengHei', 'DejaVu Sans']
+    fig, ax = plt.subplots(figsize=(6, 4))
+    sns.barplot(x=[conf, 1-conf], y=["預測類別", "其他"], palette="viridis", ax=ax)
+    ax.set_title(f"AI 診斷信心度: {conf:.2%}")
+
+    # 表格數據
+    table_data = [] # 此處可根據需求填充
+    
+    res_label = "🔴 乾眼風險 (SJS/DES)" if pred_idx == 1 else "🟢 正常/健康"
+    return (
+        gr.update(visible=False), 
+        gr.update(visible=True),
+        f"### 診斷結果：{res_label}",
+        "根據 FT-Transformer 的注意力機制分析，您的特徵與臨床乾眼指標有顯著關連。",
+        {"風險機率": conf if pred_idx==1 else 1-conf, "健康程度": 1 - (conf if pred_idx==1 else 1-conf)},
+        table_data,
+        fig,
+        fig # Demo 用，可替換為關聯圖
+    )
+
+def reset_system():
+    return [gr.update(visible=True), gr.update(visible=False), gr.update(selected=0)] + [None] * 37
+with gr.Blocks(theme=gr.themes.Soft(), css=".scroll-box { height: 450px; overflow-y: auto; }") as demo:
+    gr.Markdown("# 舌象與眼疾中西醫 AI 診斷系統")
+
+    with gr.Column(visible=True) as stage_1:
+        with gr.Tabs() as survey_tabs:
+            with gr.Tab("CCMQ 體質評估", id=0):
+                with gr.Group(elem_classes="scroll-box"):
+                    ccmq_labels = ["惡寒惡風", "自汗", "胸悶腹脹","咽喉痰梗感","多愁善感","易受驚","面部暗沉","褐班","黑眼圈","健忘","唇色暗","身熱、面熱","膚乾口乾","唇紅","便祕","兩顴紅","眼乾澀","四肢冷","惡寒、腰膝冷","飲冷腹瀉","口苦口臭","帶下色黃/下陰潮濕","鼻塞流涕","變天咳喘","過敏"]
+                    all_ccmq = [gr.Radio(["總是", "經常", "有時", "很少", "沒有"], label=f"{i+1}. {txt}") for i, txt in enumerate(ccmq_labels)]
+                btn_next = gr.Button("下一步", variant="primary")
+
+            with gr.Tab("OSDI 症狀評估", id=1):
+                with gr.Group(elem_classes="scroll-box"):
+                    gr.Markdown("#### A. 眼睛症狀")
+                    gr.Markdown("#### 在過去一週中，您是否出現下列任一症狀？")
+                    o1 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="1. 眼睛對光敏感？")
+                    o2 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="2. 眼睛有異物感？")
+                    o3 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="3. 眼睛疼痛？")
+                    o4 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="4. 視線模糊？")
+                    o5 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="5. 視力減退？")
+
+                    gr.Markdown("---")
+                    gr.Markdown("#### B. 日常活動限制")
+                    gr.Markdown("#### 下列活動，是否因眼睛問題而受到限制？")
+                    o6 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="6. 閱讀？")
+                    o7 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="7. 夜間駕駛？")
+                    o8 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="8. 操作電腦與提款機？")
+                    o9 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="9. 觀看電視？")
+
+                    gr.Markdown("---")
+                    gr.Markdown("#### C. 環境因素不適感")
+                    gr.Markdown("#### 在過去一週中遇到任一狀況時，您的眼睛是否曾感覺不適？")
+                    o10 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="10. 刮風的狀況？")
+                    o11 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="11. 濕度較低？")
+                    o12 = gr.Radio(["總是", "經常","一半一半","偶而","完全不曾"], label="12. 區域使用空調？")
+
+                    all_osdi = [o1, o2, o3, o4, o5, o6, o7, o8, o9, o10, o11, o12]
+                submit_btn = gr.Button("生成診斷報告", variant="primary")
+
+    with gr.Column(visible=False) as stage_2:
+        gr.Markdown("## 診斷分析報告")
+        with gr.Row():
+            res_table = gr.Dataframe(headers=["項目", "回答", "分值"], interactive=False)
+            with gr.Column():
+                res_prob = gr.Label(label="預測機率")
+                res_title = gr.Markdown("### 診斷結果")
+                res_desc = gr.Markdown("詳細說明...")
+                plot_1 = gr.Plot()
+                plot_2 = gr.Plot()
+        finish_btn = gr.Button("結束並重新開始", size="lg")
+
+    # 邏輯綁定
+    all_inputs = all_ccmq + all_osdi
+    btn_next.click(fn=lambda: gr.Tabs(selected=1), outputs=survey_tabs)
+    submit_btn.click(fn=analyze_and_predict, inputs=all_inputs, outputs=[stage_1, stage_2, res_title, res_desc, res_prob, res_table, plot_1, plot_2])
+    finish_btn.click(fn=reset_system, outputs=[stage_1, stage_2, survey_tabs] + all_inputs)
+
+if __name__ == "__main__":
+    demo.launch()

best_model_fold_5.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1cdb8693e41015708044318fbe6b50bd046463d9aea7d5a4d4d104f65ad711a
+size 265683

model.py

@@ -0,0 +1,53 @@
+import torch
+import torch.nn as nn
+
+class DualStreamTransformer(nn.Module):
+    def __init__(self, n_feat1=25, n_feat2=12, d_model=32, num_classes=2):
+        super(DualStreamTransformer, self).__init__()
+        
+        # Stream 1: CCMQ Tokenizer & Encoder
+        self.feat_tokenizers_1 = nn.ModuleList([nn.Linear(1, d_model) for _ in range(n_feat1)])
+        self.cls_token_1 = nn.Parameter(torch.zeros(1, 1, d_model))
+        encoder_layer_1 = nn.TransformerEncoderLayer(d_model=d_model, nhead=4, batch_first=True)
+        self.encoder_1 = nn.TransformerEncoder(encoder_layer_1, num_layers=3)
+        
+        # Stream 2: OSDI Tokenizer & Encoder
+        self.feat_tokenizers_2 = nn.ModuleList([nn.Linear(1, d_model) for _ in range(n_feat2)])
+        self.cls_token_2 = nn.Parameter(torch.zeros(1, 1, d_model))
+        encoder_layer_2 = nn.TransformerEncoderLayer(d_model=d_model, nhead=4, batch_first=True)
+        self.encoder_2 = nn.TransformerEncoder(encoder_layer_2, num_layers=3)
+        
+        # Fusion 層
+        self.mlp_head = nn.Sequential(
+            nn.Linear(d_model * 2, d_model),
+            nn.ReLU(),
+            nn.Linear(d_model, d_model) # 輸出 Embedding 給 ArcMargin
+        )
+
+    def forward(self, x1, x2):
+        # Stream 1 推論
+        tokens1 = [layer(x1[:, i].unsqueeze(1)) for i, layer in enumerate(self.feat_tokenizers_1)]
+        x1_emb = torch.stack(tokens1, dim=1)
+        x1_emb = torch.cat((self.cls_token_1.expand(x1.size(0), -1, -1), x1_emb), dim=1)
+        feat1 = self.encoder_1(x1_emb)[:, 0, :] # 取 CLS token
+        
+        # Stream 2 推論
+        tokens2 = [layer(x2[:, i].unsqueeze(1)) for i, layer in enumerate(self.feat_tokenizers_2)]
+        x2_emb = torch.stack(tokens2, dim=1)
+        x2_emb = torch.cat((self.cls_token_2.expand(x2.size(0), -1, -1), x2_emb), dim=1)
+        feat2 = self.encoder_2(x2_emb)[:, 0, :] # 取 CLS token
+        
+        # 特徵融合
+        combined = torch.cat((feat1, feat2), dim=1)
+        return self.mlp_head(combined)
+
+class ArcMarginProduct(nn.Module):
+    def __init__(self, in_features, out_features, s=30.0, m=0.5):
+        super(ArcMarginProduct, self).__init__()
+        self.weight = nn.Parameter(torch.FloatTensor(out_features, in_features))
+        nn.init.xavier_uniform_(self.weight)
+        
+    def predict(self, x):
+        # 推論時直接做線性映射或餘弦相似度
+        cosine = torch.matmul(nn.functional.normalize(x), nn.functional.normalize(self.weight).t())
+        return cosine

requirements.txt

@@ -0,0 +1,18 @@
+# 核心 Web 介面
+gradio
+
+# 深度學習框架
+torch
+torchvision
+
+# 資料處理與數值運算
+pandas
+numpy
+openpyxl
+
+# 資料視覺化 (研討會展示圖表用)
+matplotlib
+seaborn
+
+# 科學運算
+scikit-learn