def one_run(env, n_turns, steepness, noise):
env.max_turns = n_turns
env.steepness = steepness
env.noise_factor = noise
trials = int(20 * 400 / n_turns)
t = time.time()
metrics_mcts_v3 = []
for i in range(trials):
env.reset()
m = Metric('step', 'score')
root = Node(0, 10)
mcts = Mcts(root)
done = False
while not done:
action = mcts.decide()
_, r, done, _ = env.step(action)
mcts.register(r)
for j, r in enumerate(root.results):
m.add_record(j, r)
metrics_mcts_v3.append(m)
metrics_mcts_v3 = sum(metrics_mcts_v3)
print('Time for MCTSv3:', time.time() - t)
t = time.time()
import random
metrics_rnd = []
for i in range(trials):
env.reset()
m = Metric('step', 'score')
rand_results = []
done = False
while not done:
action = random.random() * 10
_, r, done, _ = env.step(action)
rand_results.append(r)
for j, r in enumerate(rand_results):
m.add_record(j, r)
metrics_rnd.append(m)
print('Time for RND:', time.time() - t)
plot_group({
'mcts_v3': metrics_mcts_v3,
'random': sum(metrics_rnd)
},
'temp', name=f'{n_turns}_st{steepness}_n{noise}')
def one_run(ga):
ga.reset()
m = Metric('steps', 'score')
while not ga.done:
r = ga.step()
m.add_record(ga.ctr, r)
return m
def one_try(env, agent_constructor, episodes):
agent = agent_constructor(env)
training_reward = Metric('episode', 'training reward')
test_reward = Metric('episode', 'test reward')
for episode in range(episodes):
r_train = train_episode(agent, env)
print(episode, f"Train: reward: {r_train:.2g}")
training_reward.add_record(episode, r_train)
if episode % 3 == 0:
r_test = trial_episode(agent, env)
print(episode, f"Test: reward: {r_test:.2f}")
test_reward.add_record(episode, r_test)
return training_reward, test_reward
def run_trials():
metrics_mcts = []
for i in range(trials):
env.reset()
m = Metric('step', 'score')
root = Node(0, 10)
mcts = Mcts(run_action, root)
done = False
while not done:
done = mcts.step()
for j, r in enumerate(root.results):
m.add_record(j, r)
metrics_mcts.append(m)
print('Score by MCTS:', sum(root.results))
def train_model(epochs: int, tag: str, train_epoch_callable, model_callable):
d = model_callable()
m_batch_loss = Metric(
name=f"batch_loss_{tag}", x_label="images trained", y_label="loss_batches"
)
m_batch_acc = Metric(
name=f"batch_acc_{tag}", x_label="images trained", y_label="acc_batches"
)
m_train_loss = Metric(
name=f"train_loss_{tag}", x_label="images trained", y_label="loss_training_set"
)
m_test_loss = Metric(
name=f"test_loss_{tag}", x_label="images trained", y_label="loss_test_set"
)
m_train_acc = Metric(
name=f"train_acc_{tag}", x_label="images trained", y_label="acc_training_set"
)
m_test_acc = Metric(
name=f"test_acc_{tag}", x_label="images trained", y_label="acc_test_set"
)
print("Started training")
for epoch in range(epochs):
_globals.epoch = epoch
print("Epoch", epoch)
t = time.time()
chronicles = train_epoch_callable(d)
if chronicles:
loss_chronicle, acc_chronicle = chronicles
m_batch_loss.add_dict(loss_chronicle)
m_batch_acc.add_dict(acc_chronicle)
_globals.epoch += 1
loss, tr_acc = d.evaluate(train_images, train_labels, verbose=0)
m_train_acc.add_record(calc_n_images(0), tr_acc)
loss, tst_acc = d.evaluate(test_images, test_labels, verbose=0)
m_test_acc.add_record(calc_n_images( 0), tst_acc)
print(f"Epoch {epoch} done in {time.time() - t:.3f}")
y_pred = tf.convert_to_tensor(d.predict(train_images))
losses = cross_entropy.call(train_labels, y_pred)
plot_statistic(
losses,
f"losses_{epoch}",
folder=os.path.join("plots", tag, "losses"),
)
m_train_loss.add_record(calc_n_images( 0), np.mean(losses))
if tag == Tags.Prio:
cifar_sampler.update_chances(losses)
plot_statistic(
cifar_sampler._chances,
f"chances_{epoch}",
folder=os.path.join("plots", tag, "chances"),
)
test_pred = tf.convert_to_tensor(d.predict(test_images))
test_losses = cross_entropy.call(test_labels, test_pred)
plot_statistic(
test_losses,
f"test_losses_{epoch}",
folder=os.path.join("plots", tag, "test_losses"),
)
m_test_loss.add_record(calc_n_images( 0), np.mean(test_losses))
print(np.mean(losses), np.mean(test_losses), tr_acc, tst_acc)
print(f"Chances changed in {time.time() - t:.3f}")
if tag == Tags.Prio:
m_batch_loss.save()
m_batch_acc.save()
m_train_loss.save()
m_test_loss.save()
m_train_acc.save()
m_test_acc.save()