def __init__(self, inpt, nin, nunits, conv_sz=1, learn_init_state=True):
# inpt is transposed a priori
tablet_wd, _ = inpt.shape
if conv_sz > 1:
inpt_clipped = inpt[:conv_sz * (tablet_wd // conv_sz), :]
inpt_conv = inpt_clipped.reshape(
(tablet_wd // conv_sz, nin * conv_sz))
else:
inpt_conv = inpt
wio = share(init_wts(nin * conv_sz, nunits)) # input to output
woo = share(init_wts(nunits, nunits)) # output to output
bo = share(init_wts(nunits))
h0 = share(init_wts(nunits))
def step(in_t, out_tm1):
return tt.tanh(tt.dot(out_tm1, woo) + tt.dot(in_t, wio) + bo)
self.output, _ = theano.scan(step,
sequences=[inpt_conv],
outputs_info=[h0])
self.params = [wio, woo, bo]
if learn_init_state:
self.params += [h0]
self.nout = nunits
def __init__(self, inpt, nin, nunits, conv_sz=1,
learn_init_state=True):
# inpt is transposed a priori
tablet_wd, _ = inpt.shape
if conv_sz > 1:
inpt_clipped = inpt[:conv_sz * (tablet_wd // conv_sz), :]
inpt_conv = inpt_clipped.reshape(
(tablet_wd // conv_sz, nin * conv_sz))
else:
inpt_conv = inpt
wio = share(init_wts(nin * conv_sz, nunits)) # input to output
woo = share(init_wts(nunits, nunits)) # output to output
bo = share(init_wts(nunits))
h0 = share(init_wts(nunits))
def step(in_t, out_tm1):
return TT.tanh(TT.dot(out_tm1, woo) + TT.dot(in_t, wio) + bo)
self.output, _ = theano.scan(
step,
sequences=[inpt_conv],
outputs_info=[h0]
)
self.params = [wio, woo, bo]
if learn_init_state:
self.params += [h0]
self.nout = nunits
def __init__(self, inpt, in_sz, n_classes, tied=False):
if tied:
b = share(init_wts(n_classes-1))
w = share(init_wts(in_sz, n_classes-1))
w1 = TT.horizontal_stack(w, TT.zeros((in_sz, 1)))
b1 = TT.concatenate((b, TT.zeros(1)))
self.output = TT.dot(inpt, w1) + b1
else:
b = share(init_wts(n_classes))
w = share(init_wts(in_sz, n_classes))
self.output = TT.dot(inpt, w) + b
self.params = [w, b]
def __init__(self, inpt, in_sz, n_classes, tied=False):
if tied:
b = share(init_wts(n_classes-1))
w = share(init_wts(in_sz, n_classes-1))
w1 = tt.horizontal_stack(w, tt.zeros((in_sz, 1)))
b1 = tt.concatenate((b, tt.zeros(1)))
self.output = tt.dot(inpt, w1) + b1
else:
b = share(init_wts(n_classes))
w = share(init_wts(in_sz, n_classes))
self.output = tt.dot(inpt, w) + b
self.params = [w, b]
