def forward(self, inputs, weights):
del weights
x1, x2 = inputs
x1_split = np.split(x1, self._n_sections, self._axis)
x2_split = np.split(x2, self._n_sections, self._axis)
res = [np.concatenate(ys, -1) for ys in zip(x1_split, x2_split)]
return tuple(res)
def forward(self, inputs, weights):
x, lstm_state = inputs
# LSTM state consists of c and h.
c, h = np.split(lstm_state, 2, axis=-1)
# Dense layer on the concatenation of x and h.
w, b = weights
y = np.dot(np.concatenate([x, h], axis=-1), w) + b
# i = input_gate, j = new_input, f = forget_gate, o = output_gate
i, j, f, o = np.split(y, 4, axis=-1)
new_c = c * backend.sigmoid(f) + backend.sigmoid(i) * np.tanh(j)
new_h = np.tanh(new_c) * backend.sigmoid(o)
return new_h, np.concatenate([new_c, new_h], axis=-1)
def dataset_to_stream(dataset, input_name, n_chunks=0):
"""Takes a tf.Dataset and creates a numpy stream of ready batches."""
for example in backend.dataset_as_numpy(dataset):
features = example[0]
inp, out = features[input_name], example[1]
mask = features['mask'] if 'mask' in features else None
# All input-pipeline processing should be on CPU.
with tf.device('cpu:0'):
# Some accelerators don't handle uint8 well, cast to int.
if isinstance(inp, np.uint8):
inp = inp.astype(np.int32)
if isinstance(out, np.uint8):
out = out.astype(np.int32)
if len(out.shape) > 1 and out.shape[-1] == 1:
out = np.squeeze(out, axis=-1)
if n_chunks > 0:
inp = tuple(np.split(inp, n_chunks, axis=1))
out = tuple(np.split(out, n_chunks, axis=1))
yield (inp, out) if mask is None else (inp, out, mask)
def reverse(self, output, weights=(), state=(), new_state=(), **kwargs):
del weights, kwargs
x1_split = []
x2_split = []
for y in output:
y1, y2 = np.split(y, 2, -1)
x1_split.append(y1)
x2_split.append(y2)
x1 = np.concatenate(x1_split, self._axis)
x2 = np.concatenate(x2_split, self._axis)
return (x1, x2)
def forward(self, inputs, weights):
x, gru_state = inputs
# Dense layer on the concatenation of x and h.
w1, b1, w2, b2 = weights
y = np.dot(np.concatenate([x, gru_state], axis=-1), w1) + b1
# Update and reset gates.
u, r = np.split(backend.sigmoid(y), 2, axis=-1)
# Candidate.
c = np.dot(np.concatenate([x, r * gru_state], axis=-1), w2) + b2
new_gru_state = u * gru_state + (1 - u) * np.tanh(c)
return new_gru_state, new_gru_state
