added separate encodings h and q

Author: aripakman

ncp.py

@@ -90,8 +90,10 @@ def __init__(self, params):
 
         if self.params['model'] == 'Gauss2D':
             self.h = Mixture_Gaussian_encoder(params)         
+            self.q = Mixture_Gaussian_encoder(params)         
         elif self.params['model'] == 'MNIST':
             self.h = MNIST_encoder(params)         
+            self.q = MNIST_encoder(params)         
         else:
             raise NameError('Unknown model '+ self.params['model'])
 
@@ -111,7 +113,7 @@ def __init__(self, params):
                 )
 
         self.f = torch.nn.Sequential(
-                torch.nn.Linear(self.g_dim +2*self.h_dim, H),
+                torch.nn.Linear(self.g_dim +self.h_dim, H),
                 torch.nn.PReLU(),    
                 torch.nn.Linear(H, H),                
                 torch.nn.PReLU(),    
@@ -136,8 +138,7 @@ def forward(self,data, cs, n):
         assert(n == self.previous_n+1)
         self.previous_n = self.previous_n + 1 
 
-        K = len(set(cs[:n]))  #num of already _assigned_clusters
-        # K is the number of distinct classes in [0:n]          
+        K = len(set(cs[:n]))  # num of already created clusters
 
         if n==1:
 
@@ -159,11 +160,13 @@ def forward(self,data, cs, n):
 
 
             self.hs = self.h(data).view([self.batch_size,self.N, self.h_dim])            
-            self.Q = self.hs[:,2:,].sum(dim=1)     #[batch_size,h_dim]
-            
             self.Hs = torch.zeros([self.batch_size, 1, self.h_dim]).to(self.device)
             self.Hs[:,0,:] = self.hs[:,0,:]
 
+            self.qs = self.q(data).view([self.batch_size,self.N, self.h_dim])            
+            self.Q = self.qs[:,2:,].sum(dim=1)     #[batch_size,h_dim]
+            
+            
 
         else:            
             if K == self.previous_K:            
@@ -177,7 +180,7 @@ def forward(self,data, cs, n):
                 self.previous_n = 0
 
             else:
-                self.Q -= self.hs[:,n,]
+                self.Q -= self.qs[:,n,]
 
 
         self.previous_K = K
@@ -196,7 +199,7 @@ def forward(self,data, cs, n):
             gs  = self.g(Hs2).view([self.batch_size, K, self.g_dim])
             Gk = gs.sum(dim=1)   #[batch_size,g_dim]
 
-            uu = torch.cat((Gk,self.Q,self.hs[:,n,:]), dim=1)  #prepare argument for the call to f()
+            uu = torch.cat((Gk,self.Q), dim=1)  #prepare argument for the call to f()
             logprobs[:,k] = torch.squeeze(self.f(uu))    
 
 
@@ -208,7 +211,7 @@ def forward(self,data, cs, n):
 
         Gk = gs.sum(dim=1)
 
-        uu = torch.cat((Gk,self.Q,self.hs[:,n,:]), dim=1)   #prepare argument for the call to f()
+        uu = torch.cat((Gk,self.Q), dim=1)   #prepare argument for the call to f()
         logprobs[:,K] = torch.squeeze(self.f(uu))    
 
 

ncp_sampler.py

@@ -30,7 +30,7 @@ def __init__(self, model, data):
             data = data.view([self.N, 28,28])
 
         self.hs = model.h(data)            
-        self.qs = self.hs
+        self.qs = model.q(data)            
 
 
         self.f = model.f
@@ -91,7 +91,7 @@ def sample(self, S):
 
                 logprobs = torch.zeros([S, maxK+1]).to(self.device)
                 rQ = self.Q.repeat(S,1)
-                rhn = self.hs[n,:].unsqueeze(0).repeat(S,1)
+
 
 
                 for k in range(maxK+1):
@@ -110,7 +110,7 @@ def sample(self, S):
 
                     Gk = gs.sum(dim=1)
 
-                    uu = torch.cat((Gk,rQ,rhn), dim=1)
+                    uu = torch.cat((Gk,rQ), dim=1)
                     logprobs[:,k] = torch.squeeze(self.f(uu))