def _lstm_block_cell(x,
cs_prev,
h_prev,
w,
b,
wci=None,
wcf=None,
wco=None,
forget_bias=None,
cell_clip=None,
use_peephole=None,
name=None):
r"""Computes the LSTM cell forward propagation for 1 time step.
This implementation uses 1 weight matrix and 1 bias vector, and there's an
optional peephole connection.
This kernel op implements the following mathematical equations:
```python
xh = [x, h_prev]
[i, ci, f, o] = xh * w + b
f = f + forget_bias
if not use_peephole:
wci = wcf = wco = 0
i = sigmoid(cs_prev * wci + i)
f = sigmoid(cs_prev * wcf + f)
ci = tanh(ci)
cs = ci .* i + cs_prev .* f
cs = clip(cs, cell_clip)
o = sigmoid(cs * wco + o)
co = tanh(cs)
h = co .* o
```
Args:
x: A `Tensor`. Must be one of the following types: `float32`.
The input to the LSTM cell, shape (batch_size, num_inputs).
cs_prev: A `Tensor`. Must have the same type as `x`.
Value of the cell state at previous time step.
h_prev: A `Tensor`. Must have the same type as `x`.
Output of the previous cell at previous time step.
w: A `Tensor`. Must have the same type as `x`. The weight matrix.
b: A `Tensor`. Must have the same type as `x`. The bias vector.
wci: A `Tensor`. Must have the same type as `x`.
The weight matrix for input gate peephole connection.
wcf: A `Tensor`. Must have the same type as `x`.
The weight matrix for forget gate peephole connection.
wco: A `Tensor`. Must have the same type as `x`.
The weight matrix for output gate peephole connection.
forget_bias: An optional `float`. Defaults to `1`. The forget gate bias.
cell_clip: An optional `float`. Defaults to `3`.
Value to clip the 'cs' value to. Disable by setting to negative value.
use_peephole: An optional `bool`. Defaults to `False`.
Whether to use peephole weights.
name: A name for the operation (optional).
Returns:
A tuple of `Tensor` objects (i, cs, f, o, ci, co, h).
i: A `Tensor`. Has the same type as `x`. The input gate.
cs: A `Tensor`. Has the same type as `x`. The cell state before the tanh.
f: A `Tensor`. Has the same type as `x`. The forget gate.
o: A `Tensor`. Has the same type as `x`. The output gate.
ci: A `Tensor`. Has the same type as `x`. The cell input.
co: A `Tensor`. Has the same type as `x`. The cell after the tanh.
h: A `Tensor`. Has the same type as `x`. The output h vector.
Raises:
ValueError: If cell_size is None.
"""
if wci is None:
cell_size = cs_prev.get_shape().with_rank(2)[1].value
if cell_size is None:
raise ValueError("cell_size from `cs_prev` should not be None.")
wci = array_ops.constant(0, dtype=dtypes.float32, shape=[cell_size])
wco = wci
wcf = wci
# pylint: disable=protected-access
return gen_lstm_ops.lstm_block_cell(
x=x,
cs_prev=cs_prev,
h_prev=h_prev,
w=w,
wci=wci,
wco=wco,
wcf=wcf,
b=b,
forget_bias=forget_bias,
cell_clip=cell_clip,
use_peephole=use_peephole,
name=name)
def _lstm_block_cell(x,
cs_prev,
h_prev,
w,
b,
wci=None,
wcf=None,
wco=None,
forget_bias=None,
cell_clip=None,
use_peephole=None,
name=None):
r"""Computes the LSTM cell forward propagation for 1 time step.
This implementation uses 1 weight matrix and 1 bias vector, and there's an
optional peephole connection.
This kernel op implements the following mathematical equations:
```python
xh = [x, h_prev]
[i, ci, f, o] = xh * w + b
f = f + forget_bias
if not use_peephole:
wci = wcf = wco = 0
i = sigmoid(cs_prev * wci + i)
f = sigmoid(cs_prev * wcf + f)
ci = tanh(ci)
cs = ci .* i + cs_prev .* f
cs = clip(cs, cell_clip)
o = sigmoid(cs * wco + o)
co = tanh(cs)
h = co .* o
```
Args:
x: A `Tensor`. Must be one of the following types: `float32`.
The input to the LSTM cell, shape (batch_size, num_inputs).
cs_prev: A `Tensor`. Must have the same type as `x`.
Value of the cell state at previous time step.
h_prev: A `Tensor`. Must have the same type as `x`.
Output of the previous cell at previous time step.
w: A `Tensor`. Must have the same type as `x`. The weight matrix.
b: A `Tensor`. Must have the same type as `x`. The bias vector.
wci: A `Tensor`. Must have the same type as `x`.
The weight matrix for input gate peephole connection.
wcf: A `Tensor`. Must have the same type as `x`.
The weight matrix for forget gate peephole connection.
wco: A `Tensor`. Must have the same type as `x`.
The weight matrix for output gate peephole connection.
forget_bias: An optional `float`. Defaults to `1`. The forget gate bias.
cell_clip: An optional `float`. Defaults to `-1` (no clipping).
Value to clip the 'cs' value to. Disable by setting to negative value.
use_peephole: An optional `bool`. Defaults to `False`.
Whether to use peephole weights.
name: A name for the operation (optional).
Returns:
A tuple of `Tensor` objects (i, cs, f, o, ci, co, h).
i: A `Tensor`. Has the same type as `x`. The input gate.
cs: A `Tensor`. Has the same type as `x`. The cell state before the tanh.
f: A `Tensor`. Has the same type as `x`. The forget gate.
o: A `Tensor`. Has the same type as `x`. The output gate.
ci: A `Tensor`. Has the same type as `x`. The cell input.
co: A `Tensor`. Has the same type as `x`. The cell after the tanh.
h: A `Tensor`. Has the same type as `x`. The output h vector.
Raises:
ValueError: If cell_size is None.
"""
if wci is None:
cell_size = cs_prev.get_shape().with_rank(2)[1].value
if cell_size is None:
raise ValueError("cell_size from `cs_prev` should not be None.")
wci = array_ops.constant(0, dtype=dtypes.float32, shape=[cell_size])
wcf = wci
wco = wci
# pylint: disable=protected-access
return gen_lstm_ops.lstm_block_cell(
x=x,
cs_prev=cs_prev,
h_prev=h_prev,
w=w,
wci=wci,
wcf=wcf,
wco=wco,
b=b,
forget_bias=forget_bias,
cell_clip=cell_clip if cell_clip is not None else -1,
use_peephole=use_peephole,
name=name)
