notwa/optim
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity

remove biasing from Conv1Dper in favor of Bias layer

Browse source
This commit is contained in:
Connor Olding 2017-08-02 11:30:08 +00:00
parent 4ee2181691
commit 049d966710
1 changed files with 2 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
onn.py
View file

@ -494,16 +494,13 @@ class Conv1Dper(Layer):
serialize = { serialize = {
'W': 'coeffs', 'W': 'coeffs',
'b': 'biases',
} }
def __init__(self, kernel_size, bias=True, pos=None, def __init__(self, kernel_size, pos=None,
init=init_glorot_uniform, reg_w=None, reg_b=None): init=init_glorot_uniform, reg_w=None):
super().__init__() super().__init__()
self.kernel_size = int(kernel_size) self.kernel_size = int(kernel_size)
self.bias = bool(bias)
self.coeffs = self._new_weights('coeffs', init=init, regularizer=reg_w) self.coeffs = self._new_weights('coeffs', init=init, regularizer=reg_w)
self.biases = self._new_weights('biases', init=init_zeros, regularizer=reg_b)
if pos is None: if pos is None:
self.wrap0 = (self.kernel_size - 0) // 2 self.wrap0 = (self.kernel_size - 0) // 2
self.wrap1 = (self.kernel_size - 1) // 2 self.wrap1 = (self.kernel_size - 1) // 2
@ -525,7 +522,6 @@ class Conv1Dper(Layer):
assert len(shape) == 1, shape assert len(shape) == 1, shape
self.output_shape = shape self.output_shape = shape
self.coeffs.shape = (1, self.kernel_size) self.coeffs.shape = (1, self.kernel_size)
self.biases.shape = (1, shape[0])
def forward(self, X): def forward(self, X):
if self.wrap0 == 0: if self.wrap0 == 0:
@ -536,14 +532,10 @@ class Conv1Dper(Layer):
Xper = np.hstack((X[:,-self.wrap0:],X,X[:,:self.wrap1])) Xper = np.hstack((X[:,-self.wrap0:],X,X[:,:self.wrap1]))
self.cols = rolling_batch(Xper, self.kernel_size) self.cols = rolling_batch(Xper, self.kernel_size)
convolved = (self.cols * self.coeffs.f[:,::-1]).sum(2) convolved = (self.cols * self.coeffs.f[:,::-1]).sum(2)
if self.bias:
convolved += self.biases.f
return convolved return convolved
def backward(self, dY): def backward(self, dY):
self.coeffs.g += (dY[:,:,None] * self.cols).sum(0)[:,::-1].sum(0, keepdims=True) self.coeffs.g += (dY[:,:,None] * self.cols).sum(0)[:,::-1].sum(0, keepdims=True)
if self.bias:
self.biases.g += dY.sum(0, keepdims=True)
return (dY[:,:,None] * self.coeffs.f[:,::-1]).sum(2) return (dY[:,:,None] * self.coeffs.f[:,::-1]).sum(2)
class LayerNorm(Layer): class LayerNorm(Layer):