Update model.py

model.py

@@ -1,49 +1,95 @@
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
+import math
 
-class DualStreamTransformer(nn.Module):
-    def __init__(self, n_feat1=24, n_feat2=10, d_model=32, num_classes=2):
-        super(DualStreamTransformer, self).__init__()
-        
-        dim_ff = 128 
-        
-        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, dim_feedforward=dim_ff, batch_first=True)
-        self.encoder_1 = nn.TransformerEncoder(encoder_layer_1, num_layers=2)
-        
-        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, dim_feedforward=dim_ff, batch_first=True)
-        self.encoder_2 = nn.TransformerEncoder(encoder_layer_2, num_layers=2)
-        
-    
-        self.fusion = nn.Sequential(
-            nn.Linear(d_model * 2, d_model), 
-            nn.ReLU()                        
-        )
-
-    def forward(self, x1, x2):
-        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, :]
-        
-        
-        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, :]
-        
-        combined = torch.cat((feat1, feat2), dim=1)
-        return self.fusion(combined) 
+
+class FocalLoss(nn.Module):
+    def __init__(self, alpha=1, gamma=2, reduction='mean'):
+        super(FocalLoss, self).__init__()
+        self.alpha = alpha
+        self.gamma = gamma
+        self.reduction = reduction
+
+    def forward(self, inputs, targets):
+        ce_loss = F.cross_entropy(inputs, targets, reduction='none')
+        pt = torch.exp(-ce_loss)
+        focal_loss = self.alpha * (1 - pt) ** self.gamma * ce_loss
+        if self.reduction == 'mean': return focal_loss.mean()
+        return focal_loss.sum()
 
 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):
-        cosine = torch.matmul(nn.functional.normalize(x), nn.functional.normalize(self.weight).t())
-        return cosine
+    def __init__(self, in_features, out_features, s=30.0, m=0.50, easy_margin=False):
+        super(ArcMarginProduct, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.s = s
+        self.m = m
+        self.weight = nn.Parameter(torch.FloatTensor(out_features, in_features))
+        nn.init.xavier_uniform_(self.weight)
+
+        self.easy_margin = easy_margin
+        self.cos_m = math.cos(m)
+        self.sin_m = math.sin(m)
+        self.th = math.cos(math.pi - m)
+        self.mm = math.sin(math.pi - m) * m
+
+    def forward(self, input, label):
+        cosine = F.linear(F.normalize(input), F.normalize(self.weight))
+        sine = torch.sqrt((1.0 - torch.pow(cosine, 2)).clamp(0, 1))
+        phi = cosine * self.cos_m - sine * self.sin_m
+        if self.easy_margin:
+            phi = torch.where(cosine > 0, phi, cosine)
+        else:
+            phi = torch.where(cosine > self.th, phi, cosine - self.mm)
+        
+        one_hot = torch.zeros(cosine.size(), device=input.device)
+        one_hot.scatter_(1, label.view(-1, 1).long(), 1)
+        output = (one_hot * phi) + ((1.0 - one_hot) * cosine)
+        output *= self.s
+        return output
+
+    def predict(self, input):
+        return F.linear(F.normalize(input), F.normalize(self.weight)) * self.s
+
+
+class DualStreamTransformer(nn.Module):
+    def __init__(self, feat_num_1, feat_num_2, d_model=64, num_classes=3, dropout=0.3):
+        super().__init__()
+        # Stream 1: CCMQ
+        self.feat_tokenizers_1 = nn.ModuleList([nn.Linear(1, d_model) for _ in range(feat_num_1)])
+        enc_layer_1 = nn.TransformerEncoderLayer(d_model=d_model, nhead=4, dim_feedforward=128, dropout=dropout, batch_first=True)
+        self.encoder_1 = nn.TransformerEncoder(enc_layer_1, num_layers=2)
+        self.cls_token_1 = nn.Parameter(torch.zeros(1, 1, d_model))
+        
+        # Stream 2: OSDI
+        self.feat_tokenizers_2 = nn.ModuleList([nn.Linear(1, d_model) for _ in range(feat_num_2)])
+        enc_layer_2 = nn.TransformerEncoderLayer(d_model=d_model, nhead=4, dim_feedforward=128, dropout=dropout, batch_first=True)
+        self.encoder_2 = nn.TransformerEncoder(enc_layer_2, num_layers=2)
+        self.cls_token_2 = nn.Parameter(torch.zeros(1, 1, d_model))
+
+        # Fusion
+        self.fusion = nn.Sequential(
+            nn.Linear(d_model * 2, d_model),
+            nn.LayerNorm(d_model),
+            nn.ReLU(),
+            nn.Dropout(dropout)
+        )
+
+    def forward_stream(self, x, tokenizers, encoder, cls_token):
+        batch_size = x.size(0)
+        tokens = []
+        for i, tokenizer in enumerate(tokenizers):
+            val = x[:, i].unsqueeze(1)
+            tokens.append(tokenizer(val))
+        x_emb = torch.stack(tokens, dim=1)
+        cls_tokens = cls_token.expand(batch_size, -1, -1)
+        x_emb = torch.cat((cls_tokens, x_emb), dim=1)
+        x_out = encoder(x_emb)
+        return x_out[:, 0, :] 
+
+    def forward(self, x1, x2):
+        feat_1 = self.forward_stream(x1, self.feat_tokenizers_1, self.encoder_1, self.cls_token_1)
+        feat_2 = self.forward_stream(x2, self.feat_tokenizers_2, self.encoder_2, self.cls_token_2)
+        combined = torch.cat((feat_1, feat_2), dim=1)
+        return self.fusion(combined)