def inception_inception(self, kvals):
(p1, p2, p3, p4) = kvals
branch1 = [Conv(**self.conv_params(1, p1[0], padding=0, strides=1))]
i2_branch1 = [
Conv(**self.conv_params((1, 3), p2[1], padding='SAME', strides=1))
]
i2_branch2 = [
Conv(**self.conv_params((3, 1), p2[1], padding='SAME', strides=1))
]
branch2 = [
Conv(**self.conv_params(1, p2[0], padding=0, strides=1)),
MergeBroadcast(layers=[i2_branch1, i2_branch2], merge="depth")
]
i3_branch1 = [
Conv(**self.conv_params((1, 3), p3[2], padding='SAME', strides=1))
]
i3_branch2 = [
Conv(**self.conv_params((3, 1), p3[2], padding='SAME', strides=1))
]
branch3 = [
Conv(**self.conv_params(1, p3[0], padding=0, strides=1)),
Conv(**self.conv_params(3, p3[1], padding='SAME', strides=1)),
MergeBroadcast(layers=[i3_branch1, i3_branch2], merge="depth")
]
branch4 = [
Pooling(**self.pool3s1p1),
Conv(**self.conv_params(1, p4[0], padding=0, strides=1))
]
return MergeBroadcast(layers=[branch1, branch2, branch3, branch4],
merge="depth")
def inception(kvals):
(p1, p2, p3, p4) = kvals
branch1 = [Conv((1, 1, p1[0]), **common)]
branch2 = [Conv((1, 1, p2[0]), **common), Conv((3, 3, p2[1]), **commonp1)]
branch3 = [Conv((1, 1, p3[0]), **common), Conv((5, 5, p3[1]), **commonp2)]
branch4 = [Pooling(op="max", **pool3s1p1), Conv((1, 1, p4[0]), **common)]
return MergeBroadcast(layers=[branch1, branch2, branch3, branch4], merge="depth")
def inception(kvals, name="i"):
(p1, p2, p3) = kvals
branch1 = [Sequential([Conv(fshape(1, p1[0]), **common)])] if p1[0] else []
branch2 = [Sequential([Conv(fshape(1, p2[0]), **common),
Conv(fshape(3, p2[1]), **commonp1)])]
branch3 = [Sequential([Pooling(op=p3[0], **pool3s1p1)] + (
[Conv(fshape(1, p3[1]), **common)] if p3[1] else []))]
partitions = branch1 + branch2 + branch3
return [MergeBroadcast(layers=partitions, merge="depth")]
def inception(kvals, name):
(p1, p2, p3, p4) = kvals
branch1 = [Conv((1, 1, p1[0]), name=name+'1x1', **common)]
branch2 = [Conv((1, 1, p2[0]), name=name+'3x3_reduce', **common),
Conv((3, 3, p2[1]), name=name+'3x3', **commonp1)]
branch3 = [Conv((1, 1, p3[0]), name=name+'5x5_reduce', **common),
Conv((5, 5, p3[1]), name=name+'5x5', **commonp2)]
branch4 = [Pooling(op="max", name=name+'pool', **pool3s1p1),
Conv((1, 1, p4[0]), name=name+'pool_proj', **common)]
return MergeBroadcast(layers=[branch1, branch2, branch3, branch4], merge="depth")
def conv_net(self, activation, init=Kaiming(), version=-1):
width = max([self.width, self.lookup_dim])
if version == -1:
if self.lookup_size:
pre_layers = [
LookupTable(vocab_size=self.lookup_size,
embedding_dim=width,
init=GlorotUniform()),
Reshape((1, self.num_words, width)),
]
first_width = width
else:
pre_layers = [
Conv((1, width, width),
padding=0,
init=init,
activation=activation)
]
first_width = 1
return pre_layers + \
[
MergeBroadcast(
[
[
Conv((3, first_width, 15), padding={'pad_h': 1, 'pad_w': 0}, init=init,
activation=activation)
],
[
Conv((5, first_width, 15), padding={'pad_h': 2, 'pad_w': 0}, init=init,
activation=activation)
],
[
Conv((7, first_width, 15), padding={'pad_h': 3, 'pad_w': 0}, init=init,
activation=activation)
],
],
merge='depth'
),
NoisyDropout(keep=0.5, noise_pct=1.0, noise_std=0.001),
Conv((5, 1, 15), strides={'str_h': 2 if self.num_words > 59 else 1,
'str_w': 1}, padding=0, init=init,
activation=activation),
NoisyDropout(keep=0.9, noise_pct=1.0, noise_std=0.00001),
Conv((3, 1, 9), strides={'str_h': 2, 'str_w': 1}, padding=0, init=init,
activation=activation),
NoisyDropout(keep=0.9, noise_pct=1.0, noise_std=0.00001),
Conv((9, 1, 9), strides={'str_h': 2, 'str_w': 1}, padding=0, init=init,
activation=activation)
]
def inception(self, kvals, b2fsz=5):
(p1, p2, p3, p4) = kvals
branch1 = [Conv(**self.conv_params(1, p1[0], padding=0, strides=1))]
branch2 = [
Conv(**self.conv_params(1, p2[0], padding=0, strides=1)),
Conv(**self.conv_params(b2fsz, p2[1], padding='SAME', strides=1))
],
branch3 = [
Conv(**self.conv_params(1, p3[0], padding=0, strides=1)),
Conv(**self.conv_params(3, p3[1], padding='SAME', strides=1)),
Conv(**self.conv_params(3, p3[1], padding='SAME', strides=1))
]
branch4 = [
Pooling(**self.pool3s1p1),
Conv(**self.conv_params(1, p4[0], padding=0, strides=1))
]
return MergeBroadcast(layers=[branch1, branch2, branch3, branch4],
merge="depth")
