public void test_ppo_model_dc_vector()
{
tf.reset_default_graph();
tf_with(tf.variable_scope("FakeGraphScope"), delegate
{
model = new PPOModel(make_brain_parameters(discrete_action: true, visual_inputs: 0));
init = tf.global_variables_initializer();
});
using (var sess = tf.Session())
{
sess.run(init);
var results = sess.run
(
// run list
(model.output,
model.all_log_probs,
model.value,
model.entropy,
model.learning_rate),
// feed dict
(model.batch_size, 2),
(model.sequence_length, 1),
(model.vector_in, np.array(new int[, ] {
{ 1, 2, 3, 1, 2, 3 }, { 3, 4, 5, 3, 4, 5 }
})),
(model.action_masks, np.ones((2, 2)))
);
print(results);
}
}
// Use this for initialization
void Start()
{
PPONetworkContinuousSimple network;
if (environment.is3D)
{
network = new PPONetworkContinuousSimple(8, 2, 2, 32, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 10000, 200);
trainer.ClipEpsilon = 0.1f;
}
else
{
network = new PPONetworkContinuousSimple(5, 2, 2, 32, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 10000, 200);
}
//test
//trainer.RewardDiscountFactor = 0.5f;
loss = new AutoAverage(iterationForEachTrain);
episodePointAve = new AutoAverage(episodeToRunForEachTrain);
}
// Use this for initialization
void Start()
{
var network = new PPONetworkDiscreteSimple(6, 3, 4, 64, DeviceDescriptor.CPUDevice, 0.01f);
model = new PPOModel(network);
trainer = new TrainerPPOSimple(model, LearnerDefs.AdamLearner(learningRate), 1, 50000, 2000);
//test
//trainer.RewardDiscountFactor = 0.5f;
loss = new AutoAverage(iterationForEachTrain);
episodePointAve = new AutoAverage(episodeToRunForEachTrain);
}
