fix added

app.py

@@ -1,3 +1,12 @@
+# =========================
+# 0. REPRODUCIBILITY (NEW FIX)
+# =========================
+import torch, numpy as np, random
+
+torch.manual_seed(42)
+np.random.seed(42)
+random.seed(42)
+
 # =========================
 # 1. DOWNLOAD + SELECTIVE EXTRACT
 # =========================
@@ -35,7 +44,7 @@ ALL_REQUIRED_IDS = set(list(train_labels.keys()) + list(dev_labels.keys()))
 print("Total required participants:", len(ALL_REQUIRED_IDS))
 
 # =========================
-# 3. SELECTIVE EXTRACTION (FIXED)
+# 3. SELECTIVE EXTRACTION
 # =========================
 def extract_needed(zip_path):
     with zipfile.ZipFile(zip_path, "r") as zip_ref:
@@ -84,15 +93,8 @@ print("Extracted participants:", len(ALL_PATHS))
 # =========================
 # 6. LIBRARIES
 # =========================
-import numpy as np
 import librosa
-import torch
 import torch.nn as nn
-import random
-
-torch.manual_seed(42)
-np.random.seed(42)
-random.seed(42)
 
 from tqdm import tqdm
 from transformers import AutoTokenizer, AutoModel
@@ -156,11 +158,7 @@ def get_visual_features(folder):
                 feats.append(np.zeros(20))
                 continue
 
-            feat = np.concatenate([
-                df.mean().values,
-                df.std().values
-            ])
-
+            feat = np.concatenate([df.mean().values, df.std().values])
             feats.append(feat)
 
         except:
@@ -192,7 +190,6 @@ Xt_test, Xa_test, Xv_test, y_test = build(dev_labels)
 
 print("Train size:", len(y_train))
 print("Test size:", len(y_test))
-print("Visual shape:", Xv.shape)
 
 # =========================
 # 10. NORMALIZE
@@ -208,7 +205,7 @@ Xa_test = sc_a.transform(Xa_test)
 Xv_test = sc_v.transform(Xv_test)
 
 # =========================
-# 11. BASELINE MODEL
+# 11. MODELS
 # =========================
 class Model(nn.Module):
     def __init__(self, vdim):
@@ -216,59 +213,37 @@ class Model(nn.Module):
         self.t = nn.Sequential(nn.Linear(768,128), nn.ReLU())
         self.a = nn.Sequential(nn.Linear(40,32), nn.ReLU())
         self.v = nn.Sequential(nn.Linear(vdim,64), nn.ReLU())
-
         self.f = nn.Sequential(
             nn.Linear(224,64),
             nn.ReLU(),
             nn.Dropout(0.3),
             nn.Linear(64,1)
         )
-
     def forward(self, t,a,v):
         return self.f(torch.cat([self.t(t), self.a(a), self.v(v)],1))
 
-# =========================
-# 12. ATTENTION MODEL (NEW)
-# =========================
 class AttentionFusionModel(nn.Module):
     def __init__(self, vdim):
         super().__init__()
-
-        self.text_fc = nn.Sequential(nn.Linear(768,128), nn.ReLU())
-        self.audio_fc = nn.Sequential(nn.Linear(40,32), nn.ReLU())
-        self.visual_fc = nn.Sequential(nn.Linear(vdim,64), nn.ReLU())
-
-        self.attention = nn.Sequential(
-            nn.Linear(224,64),
-            nn.Tanh(),
-            nn.Linear(64,3)
-        )
-
-        self.classifier = nn.Sequential(
+        self.t = nn.Sequential(nn.Linear(768,128), nn.ReLU())
+        self.a = nn.Sequential(nn.Linear(40,32), nn.ReLU())
+        self.v = nn.Sequential(nn.Linear(vdim,64), nn.ReLU())
+        self.attn = nn.Sequential(nn.Linear(224,64), nn.Tanh(), nn.Linear(64,3))
+        self.f = nn.Sequential(
             nn.Linear(224,64),
             nn.ReLU(),
             nn.Dropout(0.3),
             nn.Linear(64,1)
         )
-
-    def forward(self, t,a,v):
-        t_feat = self.text_fc(t)
-        a_feat = self.audio_fc(a)
-        v_feat = self.visual_fc(v)
-
-        combined = torch.cat([t_feat, a_feat, v_feat], dim=1)
-        attn = torch.softmax(self.attention(combined), dim=1)
-
-        fused = torch.cat([
-            attn[:,0:1]*t_feat,
-            attn[:,1:2]*a_feat,
-            attn[:,2:3]*v_feat
-        ], dim=1)
-
-        return self.classifier(fused)
+    def forward(self,t,a,v):
+        t,a,v = self.t(t), self.a(a), self.v(v)
+        comb = torch.cat([t,a,v],1)
+        w = torch.softmax(self.attn(comb),1)
+        fused = torch.cat([w[:,0:1]*t, w[:,1:2]*a, w[:,2:3]*v],1)
+        return self.f(fused)
 
 # =========================
-# 13. TRAIN BOTH MODELS
+# 12. TRAIN
 # =========================
 Xt = torch.tensor(Xt, dtype=torch.float32).to(device)
 Xa = torch.tensor(Xa, dtype=torch.float32).to(device)
@@ -279,7 +254,7 @@ Xt_test = torch.tensor(Xt_test, dtype=torch.float32).to(device)
 Xa_test = torch.tensor(Xa_test, dtype=torch.float32).to(device)
 Xv_test = torch.tensor(Xv_test, dtype=torch.float32).to(device)
 
-# ---- Baseline ----
+# Baseline
 baseline_model = Model(Xv.shape[1]).to(device)
 opt1 = torch.optim.Adam(baseline_model.parameters(), lr=1e-4)
 loss_fn = nn.BCEWithLogitsLoss()
@@ -291,14 +266,12 @@ for e in range(10):
     loss.backward()
     opt1.step()
 
-# ---- Attention ----
+# Attention (FIXED)
 attention_model = AttentionFusionModel(Xv.shape[1]).to(device)
-opt2 = torch.optim.AdamW(attention_model.parameters(), lr=3e-5)
-
-pos_weight = torch.tensor([2.0]).to(device)
-loss_fn_attn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
+opt2 = torch.optim.AdamW(attention_model.parameters(), lr=1e-4)
+loss_fn_attn = nn.BCEWithLogitsLoss()   # ✅ FIXED
 
-for e in range(20):
+for e in range(10):
     attention_model.train()
     opt2.zero_grad()
     loss = loss_fn_attn(attention_model(Xt,Xa,Xv).squeeze(), yt)
@@ -306,7 +279,7 @@ for e in range(20):
     opt2.step()
 
 # =========================
-# 14. EVALUATION
+# 13. EVALUATION
 # =========================
 baseline_model.eval()
 attention_model.eval()
@@ -316,11 +289,10 @@ with torch.no_grad():
     pred2 = (torch.sigmoid(attention_model(Xt_test,Xa_test,Xv_test).squeeze())>0.5).int().cpu().numpy()
 
 print("\n===== MODEL COMPARISON =====")
-
 print("\nBaseline Model:")
 print("Accuracy:", accuracy_score(y_test, pred1))
 print("F1 Score:", f1_score(y_test, pred1))
 
-print("\nAttention Fusion Model:")
+print("\nAttention Model:")
 print("Accuracy:", accuracy_score(y_test, pred2))
 print("F1 Score:", f1_score(y_test, pred2))