def run_training(train_dir):
"""Run training"""
resume = os.path.exists(train_dir)
with tf.Graph().as_default():
model = FeedModel()
saver = tf.train.Saver()
session = tf.Session()
summary_writer = tf.summary.FileWriter(train_dir,
graph_def=session.graph_def,
flush_secs=10)
if resume:
print("Resuming: ", train_dir)
saver.restore(session, tf.train.latest_checkpoint(train_dir))
else:
print("Starting new training: ", train_dir)
session.run(model.init)
run_inference = make_run_inference(session, model)
get_q_values = make_get_q_values(session, model)
experience_collector = ExperienceCollector()
batcher = ExperienceBatcher(experience_collector, run_inference,
get_q_values, STATE_NORMALIZE_FACTOR)
test_experiences = experience_collector.collect(
play.random_strategy, 100)
for state_batch, targets, actions in batcher.get_batches_stepwise():
global_step, _ = session.run(
[model.global_step, model.train_op],
feed_dict={
model.state_batch_placeholder: state_batch,
model.targets_placeholder: targets,
model.actions_placeholder: actions,
})
if global_step % 1e3 == 0 and global_step != 0:
saver.save(session,
train_dir + "/checkpoint",
global_step=global_step)
loss = write_summaries(session, batcher, model,
test_experiences, summary_writer)
print("Step:", global_step, "Loss:", loss)
def test_experiences_to_batches(target_computer_class_mock):
compute = target_computer_class_mock.return_value.compute
compute.return_value = np.array([42, 43])
state1 = np.arange(16).reshape((4, 4)) + 1
state2 = np.arange(16).reshape((4, 4)) + 2
state3 = np.arange(16).reshape((4, 4)) + 3
experiences = [
Experience(state1, 1, 2, state2, False, False, [3]),
Experience(state2, 3, 4, state3, True, False, [])
]
run_inference = Mock(
side_effect=[np.array([[0, 0, 0, -0.5], [0, 0, 0, 0]])])
batcher = ExperienceBatcher(None, run_inference, None, 1.0 / 15.0)
state_batch, targets, actions = batcher.experiences_to_batches(experiences)
reward_batch = np.array([2, 4])
bad_action_batch = np.array([False, True])
next_state_batch = np.array([state2.flatten(), state3.flatten()]) / 15.0
available_actions_batch = np.array([[False, False, False, True],
[False, False, False, False]])
assert (compute.call_args_list[0][0][0] == reward_batch).all()
assert (compute.call_args_list[0][0][1] == bad_action_batch).all()
assert (compute.call_args_list[0][0][2] == next_state_batch).all()
assert (compute.call_args_list[0][0][3] == available_actions_batch).all()
expected_state_batch = np.array([state1.flatten(),
state2.flatten()]) / 15.0
assert (state_batch == expected_state_batch).all()
assert (targets == np.array([42, 43])).all()
assert (actions == np.array([1, 3])).all()
