Upload 4 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+NanumGothic.ttf filter=lfs diff=lfs merge=lfs -text

NanumGothic.ttf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48a28e97b34fc8e5b157657633670cd1b7de126cfc414da65ce9c3d5bc8be733
+size 4691820

app.py

@@ -0,0 +1,16 @@
+import gradio as gr
+from emotion_predictor import predict_and_plot
+
+def analyze_dialogue(text):
+    return predict_and_plot(text)
+
+iface = gr.Interface(
+    fn=analyze_dialogue,
+    inputs=gr.Textbox(lines=15, label="대화 입력 (형식: 화자: 발화문)"),
+    outputs="html",
+    title="KOTE 감정 예측 및 시각화",
+    description="형식에 맞는 대화를 입력하면, 화자별 부정 감정 예측과 시각화 결과를 보여줍니다."
+)
+
+iface.launch()
+#11

emotion_predictor.py

@@ -0,0 +1,132 @@
+import re
+import math
+import matplotlib.pyplot as plt
+import matplotlib.font_manager as fm
+import torch
+import torch.nn as nn
+from transformers import ElectraModel, AutoTokenizer
+import numpy as np
+from sklearn.linear_model import LinearRegression
+from collections import defaultdict
+import base64
+from io import BytesIO
+
+# 폰트 설정
+font_path = './NanumGothic.ttf'
+fm.fontManager.addfont(font_path)
+plt.rcParams['font.family'] = fm.FontProperties(fname=font_path).get_name()
+plt.rcParams['axes.unicode_minus'] = False
+
+# 라벨 정의
+LABELS = [ ... ]  # 생략 없이 LABEL 전체 리스트 삽입
+NEGATIVE_EMOTIONS = [ ... ]  # 생략 없이 NEGATIVE 전체 리스트 삽입
+
+# 디바이스
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# 모델 정의
+class KOTEtagger(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.electra = ElectraModel.from_pretrained("beomi/KcELECTRA-base", revision='v2021').to(device)
+        self.tokenizer = AutoTokenizer.from_pretrained("beomi/KcELECTRA-base", revision='v2021')
+        self.classifier = nn.Linear(self.electra.config.hidden_size, 44).to(device)
+    
+    def forward(self, text):
+        encoding = self.tokenizer.encode_plus(
+            text,
+            add_special_tokens=True,
+            max_length=512,
+            return_token_type_ids=False,
+            padding="max_length",
+            return_attention_mask=True,
+            return_tensors='pt',
+        ).to(device)
+        output = self.electra(encoding["input_ids"], attention_mask=encoding["attention_mask"])
+        output = output.last_hidden_state[:, 0, :]
+        output = self.classifier(output)
+        return torch.sigmoid(output)
+
+# 모델 로드
+trained_model = KOTEtagger()
+trained_model.load_state_dict(torch.load("kote_pytorch_lightning.bin", map_location=device), strict=False)
+trained_model.eval()
+
+# 함수들
+def parse_dialogue(text):
+    lines = text.strip().split('\n')
+    return [
+        (match.group(1).strip(), match.group(2).strip())
+        for line in lines
+        if (match := re.match(r"([^:]+):(.+)", line.strip()))
+    ]
+
+def adjusted_score(raw_score, k=5):
+    return 100 / (1 + math.exp(-k * (raw_score - 0.5)))
+
+def apply_ema(scores, alpha=0.4):
+    if not scores:
+        return []
+    smoothed = [scores[0]]
+    for s in scores[1:]:
+        smoothed.append(alpha * s + (1 - alpha) * smoothed[-1])
+    return smoothed
+
+# 메인 처리 함수
+def predict_and_plot(raw_text):
+    dialogue = parse_dialogue(raw_text)
+    emotion_scores = defaultdict(lambda: defaultdict(list))
+
+    # 예측
+    for speaker, sentence in dialogue:
+        preds = trained_model(sentence)[0]
+        for label, score in zip(LABELS, preds):
+            if label in NEGATIVE_EMOTIONS:
+                adjusted = adjusted_score(score.item())
+                emotion_scores[speaker][label].append(adjusted)
+
+    html_output = ""
+    for speaker in emotion_scores:
+        html_output += f"<h3>{speaker} 감정 예측 결과:</h3>"
+        fig, ax = plt.subplots(figsize=(10, 4))
+        max_y = 0
+        plotted = False
+        predicted_scores = {}
+
+        for label in NEGATIVE_EMOTIONS:
+            raw_scores = emotion_scores[speaker].get(label, [])
+            scores = apply_ema(raw_scores)
+            if len(scores) >= 2 and max(scores) >= 40:
+                X = np.arange(len(scores)).reshape(-1, 1)
+                y = np.array(scores)
+                model = LinearRegression().fit(X, y)
+                predicted = model.predict([[len(scores)]])[0]
+                predicted_scores[label] = predicted
+                line, = ax.plot(scores, label=label)
+                color = line.get_color()
+                ax.plot([len(scores)-1, len(scores)], [scores[-1], predicted], linestyle='--', color=color)
+                plotted = True
+                max_y = max(max_y, predicted, *scores)
+                html_output += f"<p>- {label}: 예측 점수 {predicted:.2f}"
+                if predicted >= 80:
+                    html_output += f" <b style='color:red'>⚠️ 경고!</b>"
+                html_output += "</p>"
+
+        if plotted:
+            ax.set_title(f"{speaker}의 부정 감정 변화 및 예측")
+            ax.set_xlabel("발화 순서")
+            ax.set_ylabel("감정 점수")
+            ax.set_ylim(0, max(100, max_y + 10))
+            ax.legend()
+            ax.grid(True)
+            buf = BytesIO()
+            plt.tight_layout()
+            plt.savefig(buf, format='png')
+            plt.close(fig)
+            img_base64 = base64.b64encode(buf.getvalue()).decode('utf-8')
+            html_output += f"<img src='data:image/png;base64,{img_base64}'/><hr/>"
+        else:
+            html_output += "<p>⚠️ 시각화할 수 있는 감정이 없습니다.</p><hr/>"
+
+    return html_output
+#22

requirements.txt

@@ -0,0 +1,5 @@
+torch
+transformers
+matplotlib
+scikit-learn
+gradio