def op(self, concat, c, forget_bias):
batches = concat.shape[0]
vec_len = concat.shape[1]/4
assert c.shape[0] == concat.shape[0]
assert c.shape[1] == vec_len
assert c.dtype == concat.dtype
pos = position_in([batches, vec_len])
cur_batch = pos[0]
cur_elem = pos[1]
i = concat[cur_batch, cur_elem]
j = concat[cur_batch, cur_elem + vec_len]
f = concat[cur_batch, cur_elem + 2*vec_len]
o = concat[cur_batch, cur_elem + 3*vec_len]
c_cur = c[cur_batch, cur_elem]
new_c = output_like(c)
new_h = output_like(c)
new_c_cur = c_cur*sig(f + forget_bias) + sig(i) * tanh(j)
new_c[pos] = new_c_cur
new_h[pos] = tanh(new_c_cur) * sig(o)
return new_c, new_h
def sum_sq(input0, input1, input2):
pos = position_in(input0.shape)
out0 = output_like(input0)
a = input0[pos]
b = input1[pos]
c = input2[pos]
d = a * a + b * b + c * c
out0[pos] = d
return out0
def sum_sq(input0, input1, input2):
pos = position_in(input0.shape)
out0 = output_like(input0)
a = input0[pos]
b = input1[pos]
c = input2[pos]
d = a*a + b*b + c*c
out0[pos] = d
return out0
def multi_op(input0, input1, input2):
pos = position_in(input0.shape)
output0 = output_like(input0)
output1 = output_like(input0)
output2 = output_like(input0)
a = input0[pos]
b = input1[pos]
c = input2[pos]
d = a + b
output0[pos] = d
output1[pos] = d * c
output2[pos] = d + c
return output0, output1, output2
def op(self, input0, input1, input2):
pos = position_in(input0.shape)
output0 = output_like(input0)
output1 = output_like(input0)
output2 = output_like(input0)
a = input0[pos]
b = input1[pos]
c = input2[pos]
d = a + b
output0[pos] = d
output1[pos] = d*c
output2[pos] = d+c
return output0, output1, output2
def op(self, concat, c, forget_bias, d_concat, d_c):
batches = concat.shape[0]
vec_len = concat.shape[1]/4
assert c.shape[0] == concat.shape[0]
assert c.shape[1] == vec_len
assert c.dtype == concat.dtype
assert d_concat.tensor_type == concat.tensor_type
assert d_c.tensor_type == c.tensor_type
pos = position_in([batches, vec_len])
cur_batch = pos[0]
cur_elem = pos[1]
i = concat[cur_batch, cur_elem]
j = concat[cur_batch, cur_elem + vec_len]
f = concat[cur_batch, cur_elem + 2*vec_len]
o = concat[cur_batch, cur_elem + 3*vec_len]
c_cur = c[cur_batch, cur_elem]
new_c_cur = c_cur*sig(f + forget_bias) + sig(i) * tanh(j)
d_new_c = output_like(c)
d_new_h = output_like(c)
d_i = d_concat[cur_batch, cur_elem]
d_j = d_concat[cur_batch, cur_elem + vec_len]
d_f = d_concat[cur_batch, cur_elem + 2*vec_len]
d_o = d_concat[cur_batch, cur_elem + 3*vec_len]
d_c_cur = d_c[cur_batch, cur_elem]
d_new_c_cur = c_cur*sig_grad(f+forget_bias)*d_f + \
sig(f+forget_bias)*d_c_cur + \
tanh(j)*sig_grad(i)*d_i + \
sig(i)*tanh_grad(j)*d_j
d_new_h_cur = sig(o)*tanh_grad(new_c_cur)*d_new_c_cur + \
tanh(new_c_cur)*sig_grad(o)*d_o
d_new_c[pos] = d_new_c_cur
d_new_h[pos] = d_new_h_cur
return d_new_c, d_new_h
def add(x, y):
pos = position_in(x.shape)
out = output_like(x)
out[pos] = x[pos] + y[pos]
return out
def op(self, x, y):
pos = position_in(x.shape)
out = output_like(x)
out[pos] = x[pos] + y[pos]
return out
def add_inputs(op_input0, op_input1):
output = output_like(op_input0)
pos = position_in(op_input0.shape)
output[pos] = op_input0[pos] + op_input1[pos]
return output