shithub: opus

Info  •  Files  •  Log  •  Branches

Controlling per-gate sparsity

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -64,12 +64,12 @@

             N = p.shape[0]

             #print("nb = ", nb, ", N = ", N);

             #print(p.shape)

-            density = self.final_density

-            if self.batch < self.t_end:

-                r = 1 - (self.batch-self.t_start)/(self.t_end - self.t_start)

-                density = 1 - (1-self.final_density)*(1 - r*r*r)

             #print ("density = ", density)

             for k in range(nb):

+                density = self.final_density[k]

+                if self.batch < self.t_end:

+                    r = 1 - (self.batch-self.t_start)/(self.t_end - self.t_start)

+                    density = 1 - (1-self.final_density[k])*(1 - r*r*r)

                 A = p[:, k*N:(k+1)*N]

                 A = A - np.diag(np.diag(A))

                 A = np.transpose(A, (1, 0))

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -147,4 +147,4 @@

 #model.load_weights('lpcnet9b_384_10_G16_01.h5')

 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])

-model.fit([in_data, in_exc, features, periods], out_data, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, 0.1)])

+model.fit([in_data, in_exc, features, periods], out_data, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.1, 0.1, 0.1))])

-- 

⑨