def _numpy_to_tf(self, batch_size, max_seq_len, input_size, hidden_size):
with tf.Graph().as_default():
with tf.Session() as sess:
tensors = dynamic_rnn_minimal.random_inputs_tf(
batch_size, max_seq_len, input_size, hidden_size)
inputs, seq_len, w, b, init_state = tensors
tf_from_np = conversion.convert(
dynamic_rnn_minimal.numpy, numpy_to_tf,
[variables, functions, control_flow])
ops = tf_from_np(inputs, seq_len, w, b, init_state)
def target():
sess.run(ops)
self.time_execution(
('NumPy_TF', batch_size, max_seq_len, input_size,
hidden_size),
target,
extras={
'max_seq_len': max_seq_len,
'batch_size': batch_size,
'input_size': input_size,
'hidden_size': hidden_size,
})
def _autograph_baseline(self, batch_size, max_seq_len, input_size,
hidden_size):
with tf.Graph().as_default():
with tf.Session() as sess:
inputs, seq_len, w, b, init_state = dynamic_rnn_minimal.random_inputs_tf(
batch_size, max_seq_len, input_size, hidden_size)
converted_fn = tf.autograph.to_graph(
dynamic_rnn_minimal.eager,
experimental_optional_features=None)
ops = converted_fn(inputs, seq_len, w, b, init_state)
def target():
sess.run(ops)
self.time_execution(
('AutoGraph', batch_size, max_seq_len, input_size,
hidden_size),
target,
extras={
'max_seq_len': max_seq_len,
'batch_size': batch_size,
'input_size': input_size,
'hidden_size': hidden_size,
})
def _eager_baseline(self, batch_size, max_seq_len, input_size,
hidden_size):
inputs, seq_len, w, b, init_state = dynamic_rnn_minimal.random_inputs_tf(
batch_size, max_seq_len, input_size, hidden_size)
def target():
dynamic_rnn_minimal.eager(inputs, seq_len, w, b,
init_state).numpy()
self.time_execution(
('Eager', batch_size, max_seq_len, input_size, hidden_size),
target,
extras={
'max_seq_len': max_seq_len,
'batch_size': batch_size,
'input_size': input_size,
'hidden_size': hidden_size,
})
def _numpy_to_eager(self, batch_size, max_seq_len, input_size,
hidden_size):
inputs, seq_len, w, b, init_state = dynamic_rnn_minimal.random_inputs_tf(
batch_size, max_seq_len, input_size, hidden_size)
eager_from_np = conversion.convert(
dynamic_rnn_minimal.numpy, numpy_to_tf,
[variables, functions, control_flow])
def target():
eager_from_np(inputs, seq_len, w, b, init_state).numpy()
self.time_execution(
('NumPy_Eager', batch_size, max_seq_len, input_size, hidden_size),
target,
extras={
'max_seq_len': max_seq_len,
'batch_size': batch_size,
'input_size': input_size,
'hidden_size': hidden_size,
})
def _tf_baseline(self, batch_size, max_seq_len, input_size, hidden_size):
with tf.Graph().as_default():
with tf.Session() as sess:
tensors = dynamic_rnn_minimal.random_inputs_tf(
batch_size, max_seq_len, input_size, hidden_size)
inputs, seq_len, w, b, init_state = tensors
ops = dynamic_rnn_minimal.tf_(inputs, seq_len, w, b,
init_state)
def target():
sess.run(ops)
self.time_execution(
('TF', batch_size, max_seq_len, input_size, hidden_size),
target,
extras={
'max_seq_len': max_seq_len,
'batch_size': batch_size,
'input_size': input_size,
'hidden_size': hidden_size,
})
