delta = sample_dict['reward'][i] + gamma * sample_dict[ 'value_estimate'][i + 1].numpy() * sample_dict['not_done'][ i] - sample_dict['value_estimate'][i].numpy() gae = delta + gamma * my_lambda * sample_dict['not_done'][i] * gae # Insert advantage in front to get correct order sample_dict['advantage'].insert(0, gae) # Center advantage around zero sample_dict['advantage'] -= np.mean(sample_dict['advantage']) # Remove keys that are no longer used sample_dict.pop('value_estimate') sample_dict.pop('state_new') sample_dict.pop('reward') sample_dict.pop('not_done') samples = dict_to_dict_of_datasets(sample_dict, batch_size=optimization_batch_size) print('optimizing...') actor_losses = [] critic_losses = [] losses = [] for state_batch, action_batch, advantage_batch, returns_batch, log_prob_batch in zip( samples['state'], samples['action'], samples['advantage'], samples['monte_carlo'], samples['log_prob']): with tf.GradientTape() as tape: #print('ACTION:\n',action_batch) # Old policy old_log_prob = log_prob_batch #print('OLD_LOGPROB:\n',old_log_prob)
agent = manager.get_agent() for e in range(epochs): # training core # experience replay print("collecting experience..") data = manager.get_data(total_steps=100) manager.store_in_buffer(data) # sample data to optimize on from buffer sample_dict = manager.sample(sample_size) print(f"collected data for: {sample_dict.keys()}") # create and batch tf datasets data_dict = dict_to_dict_of_datasets(sample_dict, batch_size=64) print("optimizing...") # for each batch for state, action, reward, state_new, not_done in zip( data_dict['state'], data_dict['action'], data_dict['reward'], data_dict['state_new'], data_dict['not_done']): q_target = tf.cast(reward, tf.float64) + ( tf.cast(not_done, tf.float64) * tf.cast(gamma * agent.max_q(state_new), tf.float64)) with tf.GradientTape() as tape: prediction = agent.q_val(state, action) loss = loss_function(prediction, q_target) gradients = tape.gradient(loss, agent.model.trainable_variables)
def train_td3(args, model, action_dimension=None): print(args) tf.keras.backend.set_floatx('float32') ray.init(log_to_driver=False) # hyper parameters buffer_size = args.buffer_size # 10e6 in their repo, not possible with our ram epochs = args.epochs saving_path = os.getcwd() + "/" + args.saving_dir saving_after = 5 sample_size = args.sample_size optim_batch_size = args.batch_size gamma = args.gamma test_steps = 100 # 1000 in their repo policy_delay = 2 rho = .046 policy_noise = args.policy_noise policy_noise_clip = .5 msg_dim = args.msg_dim # 32 in their repo learning_rate = args.learning_rate save_args(args, saving_path) env_test_instance = gym.make('BipedalWalker-v3') if action_dimension is None: action_dimension = copy(env_test_instance.action_space.shape[0]) model_kwargs = { # action dimension for modular actions 'action_dimension': action_dimension, 'min_action': copy(env_test_instance.action_space.low)[0], 'max_action': copy(env_test_instance.action_space.high)[0], 'msg_dimension': msg_dim, 'fix_sigma': True, 'hidden_units': args.hidden_units } del env_test_instance manager = SampleManager(model, 'BipedalWalker-v3', num_parallel=(os.cpu_count() - 1), total_steps=150, action_sampling_type="continuous_normal_diagonal", is_tf=True, model_kwargs=model_kwargs) optim_keys = [ 'state', 'action', 'reward', 'state_new', 'not_done', ] manager.initialize_buffer(buffer_size, optim_keys) manager.initialize_aggregator(path=saving_path, saving_after=saving_after, aggregator_keys=["loss", "reward"]) agent = manager.get_agent() optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate) # fill buffer print("Filling buffer before training..") while len(manager.buffer.buffer[ manager.buffer.keys[0]]) < manager.buffer.size: # Gives you state action reward trajectories data = manager.get_data() manager.store_in_buffer(data) # track time while training timer = time.time() last_t = timer target_agent = manager.get_agent() for e in range(epochs): # off policy sample_dict = manager.sample(sample_size, from_buffer=True) print(f"collected data for: {sample_dict.keys()}") # cast values to float32 and create data dict sample_dict['state'] = tf.cast(sample_dict['state'], tf.float32) sample_dict['action'] = tf.cast(sample_dict['action'], tf.float32) sample_dict['reward'] = tf.cast(sample_dict['reward'], tf.float32) sample_dict['state_new'] = tf.cast(sample_dict['state_new'], tf.float32) sample_dict['not_done'] = tf.cast(sample_dict['not_done'], tf.float32) data_dict = dict_to_dict_of_datasets(sample_dict, batch_size=optim_batch_size) total_loss = 0 for state, action, reward, state_new, not_done in \ zip(data_dict['state'], data_dict['action'], data_dict['reward'], data_dict['state_new'], data_dict['not_done']): action_new = target_agent.act(state_new) # add noise to action_new action_new = action_new + tf.clip_by_value( tf.random.normal(action_new.shape, 0., policy_noise), -policy_noise_clip, policy_noise_clip) # clip action_new to action space action_new = tf.clip_by_value( action_new, manager.env_instance.action_space.low, manager.env_instance.action_space.high) # calculate target with double-Q-learning state_action_new = tf.concat([state_new, action_new], axis=-1) q_values0 = target_agent.model.critic0(state_action_new) q_values1 = target_agent.model.critic1(state_action_new) q_values = tf.concat([q_values0, q_values1], axis=-1) q_targets = tf.squeeze(tf.reduce_min(q_values, axis=-1)) critic_target = reward + gamma * not_done * q_targets state_action = tf.concat([state, action], axis=-1) # update critic 0 with tf.GradientTape() as tape: q_output = agent.model.critic0(state_action) loss = tf.keras.losses.MSE(tf.squeeze(critic_target), tf.squeeze(q_output)) total_loss += loss gradients = tape.gradient(loss, agent.model.critic0.trainable_variables) optimizer.apply_gradients( zip(gradients, agent.model.critic0.trainable_variables)) # update critic 1 with tf.GradientTape() as tape: q_output = agent.model.critic1(state_action) loss = tf.keras.losses.MSE(tf.squeeze(critic_target), tf.squeeze(q_output)) total_loss += loss gradients = tape.gradient(loss, agent.model.critic1.trainable_variables) optimizer.apply_gradients( zip(gradients, agent.model.critic1.trainable_variables)) # update actor with delayed policy update if e % policy_delay == 0: with tf.GradientTape() as tape: actor_output = agent.model.actor(state) action = reparam_action(actor_output, agent.model.action_dimension, agent.model.min_action, agent.model.max_action) state_action = tf.concat([state, action], axis=-1) q_val = agent.model.critic0(state_action) actor_loss = -tf.reduce_mean(q_val) total_loss += actor_loss actor_gradients = tape.gradient( actor_loss, agent.model.actor.trainable_variables) optimizer.apply_gradients( zip(actor_gradients, agent.model.actor.trainable_variables)) # Update agent manager.set_agent(agent.get_weights()) agent = manager.get_agent() if e % policy_delay == 0: # Polyak averaging new_weights = list(rho * np.array(target_agent.get_weights()) + (1. - rho) * np.array(agent.get_weights())) target_agent.set_weights(new_weights) reward = manager.test(test_steps, evaluation_measure="reward") manager.update_aggregator(loss=total_loss, reward=reward) print( f"epoch ::: {e} loss ::: {total_loss} avg reward ::: {np.mean(reward)}" ) if e % saving_after == 0: manager.save_model(saving_path, e) # needed time and remaining time estimation current_t = time.time() time_needed = (current_t - last_t) / 60. time_remaining = (current_t - timer) / 60. / (e + 1) * (epochs - (e + 1)) print( 'Finished epoch %d of %d. Needed %1.f min for this epoch. Estimated time remaining: %.1f min' % (e + 1, epochs, time_needed, time_remaining)) last_t = current_t manager.load_model(saving_path) print("done") print("testing optimized agent") manager.test(test_steps, test_episodes=10, render=True) ray.shutdown()
for e in range(epochs): sample_dict = manager.sample(sample_size, from_buffer=False) print(f"collected data for: {sample_dict.keys()}") # Shift value estimate by one to the left to get the value estimate of next state state_value = tf.squeeze(sample_dict['value_estimate']) state_value_new = tf.roll(state_value, -1, axis=0) not_done = tf.cast(sample_dict['not_done'], tf.bool) state_value_new = tf.where(not_done, state_value_new, 0) # Calculate advantate estimate q(s,a)-b(s)=r+v(s')-v(s) advantage_estimate = -state_value + sample_dict[ 'reward'] + gamma * state_value_new sample_dict['advantage_estimate'] = advantage_estimate data_dict = dict_to_dict_of_datasets(sample_dict, batch_size=optim_batch_size) total_loss = 0 for state, action, reward, state_new, not_done, mc, advantage_estimate, value_estimate, old_action_prob in \ zip(data_dict['state'], data_dict['action'], data_dict['reward'], data_dict['state_new'], data_dict['not_done'], data_dict['monte_carlo'], data_dict['advantage_estimate'], data_dict['value_estimate'], data_dict['log_prob']): old_action_prob = tf.cast(old_action_prob, tf.float32) mc = tf.cast(mc, tf.float32)
manager.test(MAX_TEST_STEPS, 5, evaluation_measure='time_and_reward', do_print=True, render=True) # get the initial agent agent = manager.get_agent() print('# =============== START TRAINING ================ #') for e in range(1, EPOCHS+1): print(f'# ============== EPOCH {e}/{EPOCHS} ============== #') print('# ============= collecting samples ============== #') # collect experience and save it to ERP buffer data = manager.get_data(do_print=False) manager.store_in_buffer(data) # get some samples from the ERP buffer and create a dataset sample_dict = manager.sample(sample_size=SAMPLE_SIZE) data_dict = dict_to_dict_of_datasets(sample_dict, batch_size=BATCH_SIZE) dataset = tf.data.Dataset.zip((data_dict['state'], data_dict['action'], data_dict['reward'], data_dict['state_new'], data_dict['not_done'])) print('# ================= optimizing ================== #') losses = [] for s, a, r, ns, nd in dataset: # ensure that the datasets have at least 10 elements # otherwise we run into problems with the MSE loss if len(s) >= 10: loss = train_q_network(agent, s, a, r, ns, nd, optimizer) losses.append(loss) print(f'average loss: {np.mean(losses)}') # update the weights of the manager