def r_lstm(seq_len, num_hidden, C):
import my_layer as ml
T = seq_len
cs = [S.Variable('c')]
hs = [S.Variable('h')]
preds = []
datas = [S.Variable('data%d' % i) for i in range(T)]
param = LSTMParam(x2g_weight=S.Variable("x2g_weight"),
x2g_bias=S.Variable("x2g_bias"),
h2g_weight=S.Variable("h2g_weight"),
h2g_bias=S.Variable("h2g_bias"),
Y_weight=S.Variable("Y_weight"),
Y_bias=S.Variable("Y_bias"))
for t in range(T):
pred, c, h = r_lstm_step(datas[t],
num_hidden,
C,
c=cs[-1],
h=hs[-1],
param=param)
pred = S.LogisticRegressionOutput(data=pred, name='logis%d' % t)
if t != 0:
useless = S.Custom(prev_data=preds[-1],
this_data=pred,
name='regloss',
op_type='regloss')
preds.append(pred)
cs.append(c)
hs.append(h)
return S.Group(preds)
def unroll_lstm(seq_len, num_hidden, C, H, W):
T = seq_len
cs = [S.Variable('c')]
hs = [S.Variable('h')]
preds = []
datas = [S.Variable('data%d' % i) for i in range(T)]
param = LSTMParam(i2h_weight=S.Variable("i2h_weight"),
i2h_bias=S.Variable("i2h_bias"),
h2h_weight=S.Variable("h2h_weight"),
h2h_bias=S.Variable("h2h_bias"),
Y_weight=S.Variable("Y_weight"),
Y_bias=S.Variable("Y_bias"))
for t in range(T):
pred, c, h = LSTM(datas[t],
num_hidden,
C,
H,
W,
c=cs[-1],
h=hs[-1],
param=param)
pred = S.LogisticRegressionOutput(data=pred, name='logis%d' % t)
preds.append(pred)
cs.append(c)
hs.append(h)
return S.Group(preds)
def train(self):
return S.Group(self.last_states + [self.train_loss])
def eval(self):
return S.Group(self.last_states + [self.eval_loss])
