def network(klen, sd, nbase=smt.DEFAULT_NBASE, nfeature=1, winlen=11, stride=2, size=[32, 96, 128]):
""" Create fat Nanonet with GRUs and convolution input layer
:param klen: Length of kmer
:param sd: Standard Deviation of initialisation noise
:param nfeature: Number of features per time-step
:param winlen: Length of convolution window over data
:param stride: Stride over data
:param size: Sizes of hidden recurrent layers
:returns: a `class`:layer.Layer:
"""
_prn = smt.partial(smt.truncated_normal, sd=sd)
nstate = smt.nstate(klen, nbase=nbase)
gru_act = smt.tanh
ff_act = smt.tanh
inlayer = smt.Convolution(nfeature, size[0], winlen, stride, init=_prn, has_bias=True, fun=ff_act)
fwd1 = smt.Gru(size[0], size[1], init=_prn, has_bias=True, fun=gru_act)
bwd1 = smt.Gru(size[0], size[1], init=_prn, has_bias=True, fun=gru_act)
layer1 = smt.birnn(fwd1, bwd1)
layer2 = smt.FeedForward(2 * size[1], size[2], has_bias=True, fun=ff_act)
fwd3 = smt.Gru(size[2], size[1], init=_prn, has_bias=True, fun=gru_act)
bwd3 = smt.Gru(size[2], size[1], init=_prn, has_bias=True, fun=gru_act)
layer3 = smt.birnn(fwd3, bwd3)
layer4 = smt.FeedForward(2 * size[1], size[2], init=_prn, has_bias=True, fun=ff_act)
outlayer = smt.Softmax(size[2], nstate, init=_prn, has_bias=True)
return smt.Serial([inlayer, layer1, layer2, layer3, layer4, outlayer])
def network(klen, sd, nbase=smt.DEFAULT_NBASE, nfeature=4, winlen=3, stride=1, size=4):
""" Create standard Nanonet with GRU units
:param klen: Length of kmer
:param sd: Standard Deviation of initialisation noise
:param nfeature: Number of features per time-step
:param winlen: Length of window over data
:param stride: Stride over data
:param size: Size of hidden recurrent layers
:returns: a `class`:layer.Layer:
"""
assert stride == 1, "Model only supports stride of 1"
_prn = smt.partial(smt.truncated_normal, sd=sd)
nstate = smt.nstate(klen, nbase=nbase)
gru_act = smt.tanh
ff_act = smt.tanh
insize = nfeature * winlen
inlayer = smt.Window(nfeature, winlen)
fwd1 = smt.Gru(insize, size, init=_prn, has_bias=True, fun=gru_act)
bwd1 = smt.Gru(insize, size, init=_prn, has_bias=True, fun=gru_act)
layer1 = smt.birnn(fwd1, bwd1)
layer2 = smt.FeedForward(2 * size, size, has_bias=True, fun=ff_act)
outlayer = smt.Softmax(size, nstate, init=_prn, has_bias=True)
return smt.Serial([inlayer, layer1, layer2, outlayer])
def network(klen, sd, nbase=smt.DEFAULT_NBASE, nfeature=1, winlen=11, stride=5):
""" Create a network with convolution input layer and five alternating-in-direction GRU layers
:param klen: Length of kmer
:param sd: Standard Deviation of initialisation noise
:param nbase: Number of distinct bases
:param nfeature: Number of features per time-step
:param winlen: Length of window over data
:param stride: Stride over data
:returns: a `class`:layer.Layer:
"""
n = 96
fun = smt.tanh
init = smt.partial(smt.truncated_normal, sd=sd)
return smt.Serial([smt.Convolution(nfeature, n, winlen, stride, init=init, has_bias=True, fun=smt.elu),
smt.Reverse(smt.Gru(n, n, init=init, has_bias=True, fun=fun)),
smt.Gru(n, n, init=init, has_bias=True, fun=fun),
smt.Reverse(smt.Gru(n, n, init=init, has_bias=True, fun=fun)),
smt.Gru(n, n, init=init, has_bias=True, fun=fun),
smt.Reverse(smt.Gru(n, n, init=init, has_bias=True, fun=fun)),
smt.Softmax(n, smt.nstate(klen, nbase=nbase), init=init, has_bias=True)
])
