def __init__(self, n_features, eps=None):
df.Module.__init__(self)
self.weight, self.grad_weight = create_param_and_grad(
n_features, df.init.const(1), name='W_BN_{}'.format(n_features))
self.bias, self.grad_bias = create_param_and_grad(
n_features, df.init.const(0), name='b_BN_{}'.format(n_features))
self.inference_weight = create_param(
n_features,
df.init.const(1),
name='W_BN_{}_inf'.format(n_features))
self.inference_bias = create_param(
n_features,
df.init.const(0),
name='b_BN_{}_inf'.format(n_features))
# These are buffers for collecting the minibatch statistics.
self.buffer_variance = create_param(
n_features, df.init.const(1), name='BN_var_{}'.format(n_features))
self.buffer_mean = create_param(n_features,
df.init.const(0),
name='BN_mean_{}'.format(n_features))
self.buffer_counts = df.th.shared(
_np.asarray(0, dtype=df.floatX),
name='BN_count_{}'.format(n_features))
self.eps = eps or 1e-5
self.batch_mean = None
self.batch_var = None
def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, pad_w=0, pad_h=0, mode='cross', with_bias=True, initW=df.init.xavier(), initB=df.init.const(0), border=None):
# mode='cross' is the default in Lasagne[1], Torch[2], matConvNet[3], Caffee[4].
#
# 1: https://github.com/Lasagne/Lasagne/blob/63d44a0d/lasagne/layers/dnn.py#L299
# 2: https://github.com/soumith/cudnn.torch/blob/840f0228/SpatialConvolution.lua#L83
# 3: https://github.com/vlfeat/matconvnet/blob/b7dd9c96/matlab/src/bits/impl/nnconv_cudnn.cu#L133
# 4: https://github.com/BVLC/caffe/blob/50ab52cb/include/caffe/util/cudnn.hpp#L104
#
# `border` is an alternative way to specify `pad_w` and `pad_h` so that Theano strings can be used. Better documentation to follow soon.
df.Module.__init__(self)
self.n_input_plane = n_input_plane
self.n_output_plane = n_output_plane
self.k_w = k_w
self.k_h = k_h
self.d_w = d_w
self.d_h = d_h
self.mode = mode
self.with_bias = with_bias
# 'same' is a (common) shortcut for "zero-padding so that outshape == inshape".
self.border = border or (pad_h, pad_w)
if self.border == 'same':
assert self.k_w % 2 == 1 and self.k_h % 2 == 1, "'same' convolution only supports odd filter sizes."
self.border = ((self.k_h-1)//2, (self.k_w-1)//2)
w_shape = (n_output_plane, n_input_plane, k_h, k_w)
w_fan = (n_input_plane*k_w*k_h, n_output_plane*k_w*k_h)
self.weight, self.grad_weight = create_param_and_grad(w_shape, initW, fan=w_fan, name='Wconv_{},{}@{}x{}'.format(n_input_plane, n_output_plane, k_w, k_h))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(n_output_plane, initB, name='bconv_{}'.format(n_output_plane))
def __init__(self,
n_input_vol,
n_output_vol,
k_w,
k_h,
k_d,
with_bias=True,
initW=df.init.xavier(),
initB=df.init.const(0),
border_mode='valid',
volshape=None):
df.Module.__init__(self)
self.n_input_vol = n_input_vol
self.n_output_vol = n_output_vol
self.k_w = k_w
self.k_h = k_h
self.k_d = k_d
self.with_bias = with_bias
self.border_mode = border_mode
self.volshape = volshape
self.w_shape = (n_output_vol, k_d, n_input_vol, k_h, k_w)
w_fan = (n_input_vol * k_w * k_h * k_d, n_output_vol * k_w * k_h * k_d)
self.weight, self.grad_weight = create_param_and_grad(
self.w_shape,
initW,
fan=w_fan,
name='Wconv_{},{},{},{},{}'.format(n_output_vol, k_d, n_input_vol,
k_h, k_w))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
(n_output_vol, ), initB, name='bconv_{}'.format(n_output_vol))
def __init__(
self,
n_input_vol,
n_output_vol,
k_w,
k_h,
k_d,
with_bias=True,
initW=df.init.xavier(),
initB=df.init.const(0),
border_mode="valid",
volshape=None,
):
df.Module.__init__(self)
self.n_input_vol = n_input_vol
self.n_output_vol = n_output_vol
self.k_w = k_w
self.k_h = k_h
self.k_d = k_d
self.with_bias = with_bias
self.border_mode = border_mode
self.volshape = volshape
self.w_shape = (n_output_vol, k_d, n_input_vol, k_h, k_w)
w_fan = (n_input_vol * k_w * k_h * k_d, n_output_vol * k_w * k_h * k_d)
self.weight, self.grad_weight = create_param_and_grad(
self.w_shape, initW, fan=w_fan, name="Wconv_{},{},{},{},{}".format(n_output_vol, k_d, n_input_vol, k_h, k_w)
)
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
(n_output_vol,), initB, name="bconv_{}".format(n_output_vol)
)
def __init__(self,
n_input_plane,
n_output_plane,
k_w,
k_h,
d_w=1,
d_h=1,
pad_w=0,
pad_h=0,
with_bias=True,
initW=xavier(),
initB=const(0)):
Module.__init__(self)
self.n_input_plane = n_input_plane
self.n_output_plane = n_output_plane
self.k_w = k_w
self.k_h = k_h
self.d_w = d_w
self.d_h = d_h
self.pad_w = pad_w
self.pad_h = pad_h
self.with_bias = with_bias
w_shape = (n_output_plane, n_input_plane, k_h, k_w)
w_fan = (n_input_plane * k_w * k_h, n_output_plane * k_w * k_h)
self.weight, self.grad_weight = create_param_and_grad(
w_shape,
initW,
fan=w_fan,
name='Wconv_{},{}@{}x{}'.format(n_input_plane, n_output_plane, k_w,
k_h))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
n_output_plane, initB, name='bconv_{}'.format(n_output_plane))
def __init__(self,
n_input_plane,
n_output_plane,
k_w,
k_h,
d_w=1,
d_h=1,
with_bias=True,
initW=df.init.xavier(),
initB=df.init.const(0),
border_mode='valid',
imshape=None):
df.Module.__init__(self)
self.n_input_plane = n_input_plane
self.n_output_plane = n_output_plane
self.k_w = k_w
self.k_h = k_h
self.d_w = d_w
self.d_h = d_h
self.with_bias = with_bias
self.border_mode = border_mode
self.imshape = imshape
self.w_shape = (n_output_plane, n_input_plane, k_h, k_w)
w_fan = (n_input_plane * k_w * k_h, n_output_plane * k_w * k_h)
self.weight, self.grad_weight = create_param_and_grad(
self.w_shape,
initW,
fan=w_fan,
name='Wconv_{},{}@{}x{}'.format(n_input_plane, n_output_plane, k_w,
k_h))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
n_output_plane, initB, name='bconv_{}'.format(n_output_plane))
def __init__(self, nin, nout, with_bias=True, initW=df.init.xavier(), initB=df.init.const(0)):
df.Module.__init__(self)
self.nin = nin
self.nout = nout
self.with_bias = with_bias
self.weight, self.grad_weight = create_param_and_grad((nin, nout), initW, fan=(nin, nout), name='Wlin_{}x{}'.format(nin, nout))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(nout, initB, name='blin_{}'.format(nout))
def __init__(self,
n_input_plane,
n_output_plane,
k_w,
k_h,
d_w=1,
d_h=1,
pad_w=0,
pad_h=0,
mode='cross',
with_bias=True,
initW=df.init.xavier(),
initB=df.init.const(0),
border=None):
# mode='cross' is the default in Lasagne[1], Torch[2], matConvNet[3], Caffee[4].
#
# 1: https://github.com/Lasagne/Lasagne/blob/63d44a0d/lasagne/layers/dnn.py#L299
# 2: https://github.com/soumith/cudnn.torch/blob/840f0228/SpatialConvolution.lua#L83
# 3: https://github.com/vlfeat/matconvnet/blob/b7dd9c96/matlab/src/bits/impl/nnconv_cudnn.cu#L133
# 4: https://github.com/BVLC/caffe/blob/50ab52cb/include/caffe/util/cudnn.hpp#L104
#
# `border` is an alternative way to specify `pad_w` and `pad_h` so that Theano strings can be used. Better documentation to follow soon.
df.Module.__init__(self)
self.n_input_plane = n_input_plane
self.n_output_plane = n_output_plane
self.k_w = k_w
self.k_h = k_h
self.d_w = d_w
self.d_h = d_h
self.mode = mode
self.with_bias = with_bias
# 'same' is a (common) shortcut for "zero-padding so that outshape == inshape".
self.border = border or (pad_h, pad_w)
if self.border == 'same':
assert self.k_w % 2 == 1 and self.k_h % 2 == 1, "'same' convolution only supports odd filter sizes."
self.border = ((self.k_h - 1) // 2, (self.k_w - 1) // 2)
w_shape = (n_output_plane, n_input_plane, k_h, k_w)
w_fan = (n_input_plane * k_w * k_h, n_output_plane * k_w * k_h)
self.weight, self.grad_weight = create_param_and_grad(
w_shape,
initW,
fan=w_fan,
name='Wconv_{},{}@{}x{}'.format(n_input_plane, n_output_plane, k_w,
k_h))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
n_output_plane, initB, name='bconv_{}'.format(n_output_plane))
def __init__(self, n_input_plane, n_output_plane, k_w, k_h, d_w=1, d_h=1, pad_w=0, pad_h=0, with_bias=True, initW=xavier(), initB=const(0)):
Module.__init__(self)
self.n_input_plane = n_input_plane
self.n_output_plane = n_output_plane
self.k_w = k_w
self.k_h = k_h
self.d_w = d_w
self.d_h = d_h
self.pad_w = pad_w
self.pad_h = pad_h
self.with_bias = with_bias
w_shape = (n_output_plane, n_input_plane, k_h, k_w)
w_fan = (n_input_plane*k_w*k_h, n_output_plane*k_w*k_h)
self.weight, self.grad_weight = create_param_and_grad(w_shape, initW, fan=w_fan, name='Wconv_{},{}@{}x{}'.format(n_input_plane, n_output_plane, k_w, k_h))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(n_output_plane, initB, name='bconv_{}'.format(n_output_plane))
def __init__(self, n_features, eps=None):
df.Module.__init__(self)
self.weight, self.grad_weight = create_param_and_grad(n_features, df.init.const(1), name='W_BN_{}'.format(n_features))
self.bias, self.grad_bias = create_param_and_grad(n_features, df.init.const(0), name='b_BN_{}'.format(n_features))
self.inference_weight = create_param(n_features, df.init.const(1), name='W_BN_{}_inf'.format(n_features))
self.inference_bias = create_param(n_features, df.init.const(0), name='b_BN_{}_inf'.format(n_features))
# These are buffers for collecting the minibatch statistics.
self.buffer_variance = create_param(n_features, df.init.const(1), name='BN_var_{}'.format(n_features))
self.buffer_mean = create_param(n_features, df.init.const(0), name='BN_mean_{}'.format(n_features))
self.buffer_counts = df.th.shared(_np.asarray(0, dtype=df.floatX), name='BN_count_{}'.format(n_features))
self.eps = eps or 1e-5
self.batch_mean = None
self.batch_var = None
def __init__(self,
nin,
nout,
with_bias=True,
initW=df.init.xavier(),
initB=df.init.const(0)):
df.Module.__init__(self)
self.nin = nin
self.nout = nout
self.with_bias = with_bias
self.weight, self.grad_weight = create_param_and_grad(
(nin, nout),
initW,
fan=(nin, nout),
name='Wlin_{}x{}'.format(nin, nout))
if self.with_bias:
self.bias, self.grad_bias = create_param_and_grad(
nout, initB, name='blin_{}'.format(nout))
