new moe model

script.py

@@ -56,8 +56,8 @@ print('Define Model')
 # model = ResNet_LogSpec(sample_rate=24000, return_emb=False).to(device)
 # model_path = './checkpoints/RESNET_LOGSPEC_ALL_DATA_FS_24000.pth'
 
-model = ResNet_MelSpec(sample_rate=24000, return_emb=False).to(device)
-model_path = './checkpoints/RESNET_MELSPEC_ALL_DATA_FS_24000.pth'
+# model = ResNet_MelSpec(sample_rate=24000, return_emb=False).to(device)
+# model_path = './checkpoints/RESNET_MELSPEC_ALL_DATA_FS_24000.pth'
 
 ## LCNN MODEL
 # model = LCNN(return_emb=False, fs=24000).to(device)
@@ -72,9 +72,16 @@ model_path = './checkpoints/RESNET_MELSPEC_ALL_DATA_FS_24000.pth'
 # model_path = './checkpoints/LCNN_ALL_DATA_FS_22050.pth'
 # model_path = './checkpoints/LCNN_ALL_DATA_FS_24000.pth'
 
-model.load_state_dict(torch.load(model_path, map_location=device))
+# model.load_state_dict(torch.load(model_path, map_location=device))
 
 # # MOE MODEL
+expert_1 = LCNN(return_emb=True, fs=24000)
+expert_2 = ResNet_LogSpec(return_emb=True, sample_rate=24000)
+expert_3 = ResNet_MelSpec(return_emb=True, sample_rate=24000)
+
+model = MOE_attention(experts=[expert_1, expert_2, expert_3], device=device)
+model_path = './checkpoints/MOE_TRANSF_3EXP_MODELS_AUG.pth'
+
 # expert_1 = LCNN(return_emb=True, fs=16000).to(device)
 # expert_2 = LCNN(return_emb=True, fs=22050).to(device)
 # expert_3 = LCNN(return_emb=True, fs=24000).to(device)
@@ -100,9 +107,9 @@ model.load_state_dict(torch.load(model_path, map_location=device))
 # # model_path = './checkpoints/MOE_TRANSF_7EXP_AUG_NO_FREEZE.pth'
 # # model_path = './checkpoints/MOE_TRANSF_8EXP_AUG.pth'
 # model_path = './checkpoints/MOE_TRANSF_8EXP_AUG_NO_FREEZE.pth'
-#
-# model = (model).to(device)
-# model.load_state_dict(torch.load(model_path, map_location=device))
+
+model = (model).to(device)
+model.load_state_dict(torch.load(model_path, map_location=device))
 
 model.eval()
 

src/moe_model.py

@@ -62,20 +62,41 @@ class MOE_attention(nn.Module):
     def __init__(self, experts, device, input_dim=128, freezing=False):
         super(MOE_attention, self).__init__()
 
-        self.threshold = 0.25
+        self.threshold = 0.5
         self.temperature = 1.2
 
         self.device = device
         self.experts = nn.ModuleList(experts)
         self.num_experts = len(experts)
 
+        # self.proc_emb = nn.ModuleList([
+        #     nn.Sequential(
+        #         nn.Linear(input_dim, 128),
+        #         nn.BatchNorm1d(128),
+        #         nn.GLU(),
+        #         nn.Linear(64, 32)
+        #     ) for _ in range(self.num_experts)
+        # ])
+
         self.proc_emb = nn.ModuleList([
             nn.Sequential(
-                nn.Linear(input_dim, 128),
+                nn.Linear(128, 128),
                 nn.BatchNorm1d(128),
                 nn.GLU(),
                 nn.Linear(64, 32)
-            ) for _ in range(self.num_experts)
+            ),
+            nn.Sequential(
+                nn.Linear(256, 128),
+                nn.BatchNorm1d(128),
+                nn.GLU(),
+                nn.Linear(64, 32)
+            ),
+            nn.Sequential(
+                nn.Linear(256, 128),
+                nn.BatchNorm1d(128),
+                nn.GLU(),
+                nn.Linear(64, 32)
+            )
         ])
 
         self.TransfEnc = nn.Sequential(