def impl_test_inv(self, dtype):
from scipy.linalg import inv as cpu_inv
x = np.asarray(np.random.rand(4, 4), dtype)
x = np.dot(x.T, x)
x_gpu = gpuarray.to_gpu(x)
xinv = cpu_inv(x)
xinv_gpu = linalg.inv(x_gpu)
assert np.allclose(xinv, xinv_gpu.get(), atol=1e-5)
assert xinv_gpu is not x_gpu
xinv_gpu = linalg.inv(x_gpu, overwrite=True)
assert np.allclose(xinv, xinv_gpu.get(), atol=1e-5)
assert xinv_gpu is x_gpu
def impl_test_inv(self, dtype):
from scipy.linalg import inv as cpu_inv
x = np.asarray(np.random.rand(4, 4), dtype)
x = np.dot(x.T, x)
x_gpu = gpuarray.to_gpu(x)
xinv = cpu_inv(x)
xinv_gpu = linalg.inv(x_gpu)
assert np.allclose(xinv, xinv_gpu.get(), atol=1e-5)
assert xinv_gpu is not x_gpu
xinv_gpu = linalg.inv(x_gpu, overwrite=True)
assert np.allclose(xinv, xinv_gpu.get(), atol=1e-5)
assert xinv_gpu is x_gpu
def calculate_H_gpu(X, W, P):
WPW = la.add_diag(P, la.dot(W, W, "t", "n"))
tmp = la.dot(W, la.inv(WPW, overwrite=True))
H = la.dot(X, tmp, "n", "t")
H = gpu.maximum(H, 0)
H = to_unit_variance(H)
return H, tmp
def calculate_H_gpu(X, W, P):
WPW = la.add_diag(P, la.dot(W, W, "t", "n"))
tmp = la.dot(W, la.inv(WPW, overwrite=True))
H = la.dot(X, tmp, "n", "t")
H = gpu.maximum(H, 0)
H = to_unit_variance(H)
return H, tmp
def train_rfn_gpu(X,
n_hidden,
n_iter,
learnrateW,
learnratePsi,
dropout_rate,
input_droput_rate,
minPsi=0.1,
seed=32):
k = n_hidden
n, m = X.shape
W = np.random.normal(scale=0.01, size=(k, m)).astype(np.float32)
P = np.array([0.1] * m, dtype=np.float32)
XXdiag = np.diag(np.dot(X.T, X) /
n).copy() # explicit copy to avoid numpy 1.8 warning
W = gpu.to_gpu(W, allocator=_mempool.allocate)
P = gpu.to_gpu(P, allocator=_mempool.allocate)
X = gpu.to_gpu(X, allocator=_mempool.allocate)
XXdiag = gpu.to_gpu(XXdiag, allocator=_mempool.allocate)
I = la.eye(k, dtype=np.float32)
init_rng(seed)
t0 = time.time()
for cur_iter in range(n_iter):
H, tmp = calculate_H_gpu(X, W, P)
if dropout_rate > 0:
dropout(H, dropout_rate)
Xtmp = X
if input_dropout_rate > 0:
Xtmp = X.copy()
saltpepper_noise(Xtmp, input_dropout_rate)
U = la.dot(Xtmp, H, "t", "n") / n
S = la.dot(H, H, "t", "n") / n
S += I
S -= la.dot(tmp, W, "n", "t")
Cii = la.dot(la.dot(W, S, "t") - 2 * U, W)
Sinv = la.inv(S, overwrite=True)
dW = la.dot(Sinv, U, "n", "t") - W
dP = XXdiag + la.diag(Cii) - P
W += learnrateW * dW
P += learnratePsi * dP
P = gpu.maximum(P, minPsi)
if cur_iter % 25 == 0:
print "iter %3d (elapsed time: %5.2fs)" % (cur_iter,
time.time() - t0)
return W.get(), P.get()
def train_rfn_gpu(X, n_hidden, n_iter, learnrateW, learnratePsi, dropout_rate, input_droput_rate, minPsi=0.1, seed=32):
k = n_hidden
n, m = X.shape
W = np.random.normal(scale=0.01, size=(k, m)).astype(np.float32)
P = np.array([0.1] * m, dtype=np.float32)
XXdiag = np.diag(np.dot(X.T, X) / n).copy() # explicit copy to avoid numpy 1.8 warning
W = gpu.to_gpu(W, allocator=_mempool.allocate)
P = gpu.to_gpu(P, allocator=_mempool.allocate)
X = gpu.to_gpu(X, allocator=_mempool.allocate)
XXdiag = gpu.to_gpu(XXdiag, allocator=_mempool.allocate)
I = la.eye(k, dtype=np.float32)
init_rng(seed)
t0 = time.time()
for cur_iter in range(n_iter):
H, tmp = calculate_H_gpu(X, W, P)
if dropout_rate > 0:
dropout(H, dropout_rate)
Xtmp = X
if input_dropout_rate > 0:
Xtmp = X.copy()
saltpepper_noise(Xtmp, input_dropout_rate)
U = la.dot(Xtmp, H, "t", "n") / n
S = la.dot(H, H, "t", "n") / n
S += I
S -= la.dot(tmp, W, "n", "t")
Cii = la.dot(la.dot(W, S, "t") - 2*U, W)
Sinv = la.inv(S, overwrite=True)
dW = la.dot(Sinv, U, "n", "t") - W
dP = XXdiag + la.diag(Cii) - P
W += learnrateW * dW
P += learnratePsi * dP
P = gpu.maximum(P, minPsi)
if cur_iter % 25 == 0:
print "iter %3d (elapsed time: %5.2fs)" % (cur_iter, time.time() - t0)
return W.get(), P.get()
