def __init__(self, model, batch_size=1, input_size=1):
self.model = model
# self.bins = np.linspace(-1, 1, self.model.num_classes)
_, self.bins = mu_law_bins(self.num_classes)
inputs = tf.placeholder(tf.float32, [batch_size, input_size],
name='inputs')
print('Make Generator.')
count = 0
h = inputs
init_ops = []
push_ops = []
for b in range(self.model.num_blocks):
for i in range(self.model.num_layers):
rate = 2**i
name = 'b{}-l{}'.format(b, i)
if count == 0:
state_size = 1
else:
state_size = self.model.num_hidden
q = tf.FIFOQueue(rate,
dtypes=tf.float32,
shapes=(batch_size, state_size))
init = q.enqueue_many(tf.zeros((rate, batch_size, state_size)))
state_ = q.dequeue()
push = q.enqueue([h])
init_ops.append(init)
push_ops.append(push)
h = _causal_linear(h, state_, name=name, activation=tf.nn.relu)
count += 1
outputs = _output_linear(h)
out_ops = [tf.nn.softmax(outputs)]
out_ops.extend(push_ops)
self.inputs = inputs
self.init_ops = init_ops
self.out_ops = out_ops
# Initialize queues.
self.model.sess.run(self.init_ops)
def __init__(self, model, batch_size=1, input_size=1):
self.model = model
self.bins = np.linspace(-1, 1, 256)
inputs = tf.placeholder(tf.float32, [batch_size, input_size],
name='inputs')
print 'Make Generator.'
count = 0
h = inputs
push_ops = []
for b in range(2):
for i in range(14):
rate = 2**i
name = 'b{}-l{}'.format(b, i)
if count == 0:
state_size = 1
else:
state_size = 128
q = Queue(batch_size=batch_size,
state_size=state_size,
buffer_size=rate,
name=name)
state_ = q.pop()
push = q.push(h)
push_ops.append(push)
h = _causal_linear(h, state_, name=name, activation=tf.nn.relu)
count += 1
outputs = _output_linear(h)
out_ops = [tf.argmax(tf.nn.softmax(outputs), 1)]
out_ops.extend(push_ops)
# Initialize new variables
new_vars = [
var for var in tf.trainable_variables()
if 'pointer' in var.name or 'state_buffer' in var.name
]
self.model.sess.run(tf.initialize_variables(new_vars))
self.inputs = inputs
self.out_ops = out_ops