eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-05-17 08:58:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

Managing to actually use sparse matrices

Browse source 
Now 2x real-time!
This commit is contained in:
Jean-Marc Valin 2018-11-28 20:20:17 -05:00
parent 4de3e53a73
commit d961d009a0
3 changed files with 14 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

10
dnn/dump_lpcnet.py
View file

@ -67,11 +67,19 @@ def printSparseVector(f, A, name):
A[:,N:2*N] = A[:,N:2*N] - np.diag(np.diag(A[:,N:2*N]))
A[:,2*N:] = A[:,2*N:] - np.diag(np.diag(A[:,2*N:]))
printVector(f, diag, name + '_diag')
idx = np.zeros((0,), dtype='int')
for i in range(3*N//16):
pos = idx.shape[0]
idx = np.append(idx, -1)
nb_nonzero = 0
for j in range(N):
if np.sum(np.abs(A[j, i*16:(i+1)*16])) > 1e-10:
nb_nonzero = nb_nonzero + 1
idx = np.append(idx, j)
W = np.concatenate([W, A[j, i*16:(i+1)*16]])
idx[pos] = nb_nonzero
printVector(f, W, name)
idx = np.tile(np.concatenate([np.array([N]), np.arange(N)]), 3*N//16)
#idx = np.tile(np.concatenate([np.array([N]), np.arange(N)]), 3*N//16)
printVector(f, idx, name + '_idx', dtype='int')
return;

4
dnn/lpcnet.py
View file

@ -57,7 +57,7 @@ class Sparsify(Callback):
pass
else:
#print("constrain");
layer = self.model.get_layer('cu_dnngru_1')
layer = self.model.get_layer('gru_a')
w = layer.get_weights()
p = w[1]
nb = p.shape[1]//p.shape[0]
@ -72,6 +72,7 @@ class Sparsify(Callback):
for k in range(nb):
A = p[:, k*N:(k+1)*N]
A = A - np.diag(np.diag(A))
A = np.transpose(A, (1, 0))
L=np.reshape(A, (N, N//16, 16))
S=np.sum(L*L, axis=-1)
SS=np.sort(np.reshape(S, (-1,)))
@ -79,6 +80,7 @@ class Sparsify(Callback):
mask = (S>=thresh).astype('float32');
mask = np.repeat(mask, 16, axis=1)
mask = np.minimum(1, mask + np.diag(np.ones((N,))))
mask = np.transpose(mask, (1, 0))
p[:, k*N:(k+1)*N] = p[:, k*N:(k+1)*N]*mask
#print(thresh, np.mean(mask))
w[1] = p

2
dnn/train_lpcnet.py
View file

@ -139,7 +139,7 @@ periods = (50*features[:,:,36:37]+100).astype('int16')
in_data = np.concatenate([in_data, pred], axis=-1)
# dump models to disk as we go
checkpoint = ModelCheckpoint('lpcnet9_384_10_G16_{epoch:02d}.h5')
checkpoint = ModelCheckpoint('lpcnet9b_384_10_G16_{epoch:02d}.h5')
#model.load_weights('wavenet4f2_30.h5')
model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])