guided attn #10

train.py

@@ -106,10 +106,8 @@ def train(model, criterion, data_loader, optimizer, epoch):
             
         # create attention mask
         # TODO: vectorize
-        print(text_input_var.shape)
-        print(mel_spec_var.shape)
         N = text_input_var.shape[1]
-        T = mel_spec_var.shape[1]
+        T = mel_spec_var.shape[1] / c.r
         M = np.zeros([N, T])
         for t in range(T):
             for n in range(N):
@@ -117,7 +115,7 @@ def train(model, criterion, data_loader, optimizer, epoch):
                 M[n, t] = val
         e_x = np.exp(M - np.max(M))
         M = e_x / e_x.sum(axis=0) # only difference
-        M = Variable(torch.FloatTensor(M)).cuda()
+        M = Variable(torch.FloatTensor(M).t()).cuda()
         M = torch.stack([M]*32)
 
         # forward pass