def setup_slow_walk(self):
self.walk_input = theano.shared(np.ones((2, 2), dtype='int8'))
self.walk_time = theano.shared(np.array(0, dtype='int64'))
self.walk_hiddens = [
theano.shared(np.ones((2, 2), dtype=theano.config.floatX))
for layer in self.time_model.layers if has_hidden(layer)
]
# correct for dropout
if self.dropout > 0:
masks = [1 - self.dropout for layer in self.time_model.layers]
masks[0] = None
else:
masks = []
new_states = self.time_model.forward(self.walk_input,
prev_hiddens=self.walk_hiddens,
dropout=masks)
# Now new_states is a list of matrix [layer](notes, hidden_states) for each layer
time_final = get_last_layer(new_states)
start_note_values = theano.tensor.alloc(np.array(0, dtype=np.int8), 2)
note_outputs_info = ([
initial_state_with_taps(layer) for layer in self.pitch_model.layers
] + [dict(initial=start_note_values, taps=[-1])])
notes_result, updates = theano.scan(fn=self._predict_step_note,
sequences=[time_final],
outputs_info=note_outputs_info)
# Now notes_result is a list of matrix [layer/output](notes, onOrArtic)
output = get_last_layer(notes_result)
next_input = OutputFormToInputFormOp()(output, self.walk_time +
1) # TODO: Fix time
#next_input = T.cast(T.alloc(0, 3, 4),'int64')
slow_walk_results = (new_states[:-1] + notes_result[:-1] +
[next_input, output])
updates.update({
self.walk_time: self.walk_time + 1,
self.walk_input: next_input
})
updates.update({
hidden: newstate
for hidden, newstate, layer in zip(self.walk_hiddens, new_states,
self.time_model.layers)
if has_hidden(layer)
})
self.slow_walk_fun = theano.function(inputs=[self.conservativity],
outputs=slow_walk_results,
updates=updates,
allow_input_downcast=True)
def step_time(*states):
# States is [ *hiddens, prev_result, time]
hiddens = list(states[:-2])
in_data = states[-2]
time = states[-1]
# correct for dropout
if self.dropout > 0:
masks = [1 - self.dropout for layer in self.time_model.layers]
masks[0] = None
else:
masks = []
new_states = self.time_model.forward(in_data,
prev_hiddens=hiddens,
dropout=masks)
# Now new_states is a list of matrix [layer](notes, hidden_states) for each layer
time_final = get_last_layer(new_states)
start_note_values = theano.tensor.alloc(np.array(0, dtype=np.int8),
2)
# This gets a little bit complicated. In the training case, we can pass in a combination of the
# time net's activations with the known choices. But in the prediction case, those choices don't
# exist yet. So instead of iterating over the combination, we iterate over only the activations,
# and then combine in the previous outputs in the step. And then since we are passing outputs to
# previous inputs, we need an additional outputs_info for the initial "previous" output of zero.
note_outputs_info = ([
initial_state_with_taps(layer)
for layer in self.pitch_model.layers
] + [dict(initial=start_note_values, taps=[-1])])
notes_result, updates = theano.scan(fn=self._predict_step_note,
sequences=[time_final],
outputs_info=note_outputs_info)
# Now notes_result is a list of matrix [layer/output](notes, onOrArtic)
output = get_last_layer(notes_result)
# TODO: input should be changed to next segment instead of previous output
next_input = OutputFormToInputFormOp()(output,
time + 1) # TODO: Fix time
#next_input = T.cast(T.alloc(0, 3, 4),'int64')
return (ensure_list(new_states) +
[next_input, time + 1, output]), updates