def lstm_layer(num_timesteps: int, num_inputs: int, num_units: int,
init=glorot_uniform, bias_init=sym.zeros):
"""
Create a single cell and replicate it `num_timesteps` times for training.
Return X,[(batch_size,num_classes) x num_timesteps]
"""
X = Variable("X", shape=(_batch_size, num_timesteps, num_inputs), dtype='float32')
U_shape = (num_inputs + num_units, num_units)
b_shape = (1, num_units)
Ug = Variable("Ug", init=init(U_shape))
bg = Variable("bg", init=bias_init(shape=b_shape))
Ui = Variable("Ui", init=init(U_shape))
bi = Variable("bi", init=bias_init(shape=b_shape))
Uf = Variable("Uf", init=init(U_shape))
bf = Variable("bf", init=bias_init(shape=b_shape) + sym.ones(shape=b_shape))
Uo = Variable("Uo", init=init(U_shape))
bo = Variable("bo", init=bias_init(shape=b_shape))
def cell(x_t, s_t, h_t):
xh_t = sym.concatenate(x_t, h_t, axis=1)
g = lstm_gate(sym.tanh, Ug, bg, xh_t, num_units)
i = lstm_gate(sym.sigmoid, Ui, bi, xh_t, num_units)
f = lstm_gate(sym.sigmoid, Uf, bf, xh_t, num_units)
o = lstm_gate(sym.sigmoid, Uo, bo, xh_t, num_units)
s_t1 = s_t * f + g * i
h_t1 = sym.tanh(s_t1) * o
return (s_t1, h_t1)
xs = sym.split(X, indices_or_sections=num_timesteps, axis=1)
xs = [sym.squeeze(x, axis=1) for x in xs]
# in TF:
# batch_size = sym.shape(X)[0]
# s_shape = sym.stack([batch_size, num_units], name="s_shape")
#
# s = sym.zeros(s_shape, dtype=np.float32)
if num_units > num_inputs:
s_like = sym.pad(xs[0], pad_width=((0, 0), (0, num_units-num_inputs)))
else:
s_like = xs[0][:, 0:num_units] # TODO untested
s = sym.zeros_like(s_like)
h = s
outputs = []
for x in xs:
# x = sym.squeeze(x, axis=1)
s, h = cell(x, s, h)
outputs.append(h)
return X, outputs
def _get_model(dshape):
data = sym.Variable('data', shape=dshape)
fc1 = sym.dense(data, units=dshape[-1]*2, use_bias=True)
left, right = sym.split(fc1, indices_or_sections=2, axis=1)
return sym.Group(((left + 1), (right - 1)))
def _get_model(dshape):
data = sym.Variable('data', shape=dshape)
fc1 = sym.dense(data, units=dshape[-1] * 2, use_bias=True)
left, right = sym.split(fc1, indices_or_sections=2, axis=1)
return sym.Group(((left + 1), (right - 1)))