def update(wts, grad, incr, epsilon, momentum, decay, batch_size):
#assert weight.incr.get().mean() < 1
#a, b, c, = weight.incr.get().mean(), weight.wt.get().mean(), (grad.get() * weight.epsilon / batch_size).mean()
#util.log_info('%s %s %s %s %s', weight.name, a, b, c, grad.get().mean())
Assert.eq(grad.shape, wts.shape)
assert (grad.dtype == np.float32)
assert (wts.dtype == np.float32)
assert (incr.dtype == np.float32)
if momentum > 0.0:
matrix_add(incr,
grad,
alpha=momentum,
beta=np.float32(epsilon / batch_size))
matrix_add(incr, wts, alpha=1, beta=np.float32(-decay * epsilon))
matrix_add(wts, incr)
else:
matrix_add(wts, grad, alpha=1, beta=np.float32(epsilon / batch_size))
def update(wts, grad, incr, epsilon, momentum, decay, batch_size):
#assert weight.incr.get().mean() < 1
#a, b, c, = weight.incr.get().mean(), weight.wt.get().mean(), (grad.get() * weight.epsilon / batch_size).mean()
#util.log_info('%s %s %s %s %s', weight.name, a, b, c, grad.get().mean())
Assert.eq(grad.shape, wts.shape)
assert(grad.dtype == np.float32)
assert(wts.dtype == np.float32)
assert(incr.dtype == np.float32)
if momentum > 0.0:
matrix_add(incr, grad,
alpha=momentum,
beta=np.float32(epsilon / batch_size))
matrix_add(incr, wts,
alpha=1,
beta=np.float32(-decay * epsilon))
matrix_add(wts, incr)
else:
matrix_add(wts, grad,
alpha=1,
beta=np.float32(epsilon / batch_size))
def wt(self):
if self._wt is not None:
Assert.eq(self._wt.shape, self.shape)
return self._wt
def set_weight(self, w):
if self.shape is None:
self.shape = w.shape
Assert.eq(w.shape, self.shape)
self._wt = to_gpu(w)
def wt(self):
if self._wt is not None:
Assert.eq(self._wt.shape, self.shape)
return self._wt
def set_weight(self, w):
if self.shape is None:
self.shape = w.shape
Assert.eq(w.shape, self.shape)
self._wt = to_gpu(w)