Multiple improvements and bug fixes (#66)

* Multiple improvements and bug fixes:

    * Using lazy stacking to save on memory when using a replay buffer
    * Remove step counting for evaluation episodes
    * Reset game between heatup and training
    * Major bug fixes in NEC (is reproducing the paper results for pong now)
    * Image input rescaling to 0-1 is now optional
    * Change the terminal title to be the experiment name
    * Observation cropping for atari is now optional
    * Added random number of noop actions for gym to match the dqn paper
    * Fixed a bug where the evaluation episodes won't start with the max possible ale lives
    * Added a script for plotting the results of an experiment over all the atari games

.gitignore

@@ -15,3 +15,8 @@ roboschool
 *.orig
 docs/site
 coach_env
+build
+rl_coach.egg*
+contrib
+test_log_*
+dist

agents/agent.py

@@ -24,6 +24,8 @@ except:
 import copy
 from renderer import Renderer
 from configurations import Preset
+from collections import deque
+from utils import LazyStack
 from collections import OrderedDict
 from utils import RunPhase, Signal, is_empty, RunningStat
 from architectures import *
@@ -214,6 +216,8 @@ class Agent(object):
                 network.online_network.curr_rnn_c_in = network.online_network.middleware_embedder.c_init
                 network.online_network.curr_rnn_h_in = network.online_network.middleware_embedder.h_init
 
+        self.prepare_initial_state()
+
     def preprocess_observation(self, observation):
         """
         Preprocesses the given observation.
@@ -291,9 +295,8 @@ class Agent(object):
         """
         current_states = {}
         next_states = {}
-
-        current_states['observation'] = np.array([transition.state['observation'] for transition in batch])
-        next_states['observation'] = np.array([transition.next_state['observation'] for transition in batch])
+        current_states['observation'] = np.array([np.array(transition.state['observation']) for transition in batch])
+        next_states['observation'] = np.array([np.array(transition.next_state['observation']) for transition in batch])
         actions = np.array([transition.action for transition in batch])
         rewards = np.array([transition.reward for transition in batch])
         game_overs = np.array([transition.game_over for transition in batch])
@@ -349,6 +352,23 @@ class Agent(object):
             input_state[input_name] = np.expand_dims(np.array(curr_state[input_name]), 0)
         return input_state
         
+    def prepare_initial_state(self):
+        """
+        Create an initial state when starting a new episode
+        :return: None
+        """
+        observation = self.preprocess_observation(self.env.state['observation'])
+        self.curr_stack = deque([observation]*self.tp.env.observation_stack_size, maxlen=self.tp.env.observation_stack_size)
+        observation = LazyStack(self.curr_stack, -1)
+
+        self.curr_state = {
+            'observation': observation
+        }
+        if self.tp.agent.use_measurements:
+            self.curr_state['measurements'] = self.env.measurements
+            if self.tp.agent.use_accumulated_reward_as_measurement:
+                self.curr_state['measurements'] = np.append(self.curr_state['measurements'], 0)
+
     def act(self, phase=RunPhase.TRAIN):
         """
         Take one step in the environment according to the network prediction and store the transition in memory
@@ -356,34 +376,12 @@ class Agent(object):
         :return: A boolean value that signals an episode termination
         """
 
+        if phase != RunPhase.TEST:
             self.total_steps_counter += 1
         self.current_episode_steps_counter += 1
 
         # get new action
-        action_info = {"action_probability": 1.0 / self.env.action_space_size, "action_value": 0}
-        is_first_transition_in_episode = (self.curr_state == {})
-        if is_first_transition_in_episode:
-            if not isinstance(self.env.state, dict):
-                raise ValueError((
-                    'expected state to be a dictionary, found {}'
-                ).format(type(self.env.state)))
-
-            state = self.env.state
-            # TODO: modify preprocess_observation to modify the entire state
-            # for now, only preprocess the observation
-            state['observation'] = self.preprocess_observation(state['observation'])
-
-            # TODO: provide option to stack more than just the observation
-            # TODO: this should probably be happening in an environment wrapper anyway
-            state['observation'] = stack_observation([], state['observation'], self.tp.env.observation_stack_size)
-
-            self.curr_state = state
-            if self.tp.agent.use_measurements:
-                # TODO: this should be handled in the environment
-                self.curr_state['measurements'] = self.env.measurements
-
-                if self.tp.agent.use_accumulated_reward_as_measurement:
-                    self.curr_state['measurements'] = np.append(self.curr_state['measurements'], 0)
+        action_info = {"action_probability": 1.0 / self.env.action_space_size, "action_value": 0, "max_action_value": 0}
 
         if phase == RunPhase.HEATUP and not self.tp.heatup_using_network_decisions:
             action = self.env.get_random_action()
@@ -409,8 +407,10 @@ class Agent(object):
 
         # initialize the next state
         # TODO: provide option to stack more than just the observation
-        next_state['observation'] = stack_observation(self.curr_state['observation'], next_state['observation'], self.tp.env.observation_stack_size)
+        self.curr_stack.append(next_state['observation'])
+        observation = LazyStack(self.curr_stack, -1)
 
+        next_state['observation'] = observation
         if self.tp.agent.use_measurements and 'measurements' in result.keys():
             next_state['measurements'] = result['state']['measurements']
             if self.tp.agent.use_accumulated_reward_as_measurement:
@@ -516,6 +516,7 @@ class Agent(object):
         self.exploration_policy.change_phase(RunPhase.TRAIN)
         training_start_time = time.time()
         model_snapshots_periods_passed = -1
+        self.reset_game()
 
         while self.training_iteration < self.tp.num_training_iterations:
             # evaluate
@@ -526,7 +527,7 @@ class Agent(object):
                               self.training_iteration % self.tp.evaluate_every_x_training_iterations == 0)
 
             if evaluate_agent:
-                self.env.reset()
+                self.env.reset(force_environment_reset=True)
                 self.last_episode_evaluation_ran = self.current_episode
                 self.evaluate(self.tp.evaluation_episodes)
 

agents/nec_agent.py

@@ -27,10 +27,7 @@ class NECAgent(ValueOptimizationAgent):
         ValueOptimizationAgent.__init__(self, env, tuning_parameters, replicated_device, thread_id,
                                         create_target_network=False)
         self.current_episode_state_embeddings = []
-        self.current_episode_actions = []
         self.training_started = False
-        # if self.tp.checkpoint_restore_dir:
-        #     self.load_dnd(self.tp.checkpoint_restore_dir)
 
     def learn_from_batch(self, batch):
         if not self.main_network.online_network.output_heads[0].DND.has_enough_entries(self.tp.agent.number_of_knn):
@@ -41,83 +38,57 @@ class NECAgent(ValueOptimizationAgent):
                 screen.log_title("Finished collecting initial entries in DND. Starting to train network...")
 
         current_states, next_states, actions, rewards, game_overs, total_return = self.extract_batch(batch)
-        result = self.main_network.train_and_sync_networks(current_states, total_return)
+
+        TD_targets = self.main_network.online_network.predict(current_states)
+
+        #  only update the action that we have actually done in this transition
+        for i in range(self.tp.batch_size):
+            TD_targets[i, actions[i]] = total_return[i]
+
+        # train the neural network
+        result = self.main_network.train_and_sync_networks(current_states, TD_targets)
+
         total_loss = result[0]
 
         return total_loss
 
-    def choose_action(self, curr_state, phase=RunPhase.TRAIN):
-        """
-        this method modifies the superclass's behavior in only 3 ways:
-
-        1) the embedding is saved and stored in self.current_episode_state_embeddings
-        2) the dnd output head is only called if it has a minimum number of entries in it
-            ideally, the dnd had would do this on its own, but in my attempt in encoding this
-            behavior in tensorflow, I ran into problems. Would definitely be worth
-            revisiting in the future
-        3) during training, actions are saved and stored in self.current_episode_actions
-            if behaviors 1 and 2 were handled elsewhere, this could easily be implemented
-            as a wrapper around super instead of overriding this method entirelysearch
-        """
-
-        # get embedding
+    def act(self, phase=RunPhase.TRAIN):
+        if self.in_heatup:
+            # get embedding in heatup (otherwise we get it through choose_action)
             embedding = self.main_network.online_network.predict(
-            self.tf_input_state(curr_state),
+                self.tf_input_state(self.curr_state),
                 outputs=self.main_network.online_network.state_embedding)
             self.current_episode_state_embeddings.append(embedding)
 
-        # TODO: support additional heads.  Right now all other heads are ignored
-        if self.main_network.online_network.output_heads[0].DND.has_enough_entries(self.tp.agent.number_of_knn):
-            # if there are enough entries in the DND then we can query it to get the action values
-            # actions_q_values = []
-            feed_dict = {
-                self.main_network.online_network.state_embedding: [embedding],
-            }
-            actions_q_values = self.main_network.sess.run(
-                self.main_network.online_network.output_heads[0].output, feed_dict=feed_dict)
-        else:
-            # get only the embedding so we can insert it to the DND
-            actions_q_values = [0] * self.action_space_size
+        return super().act(phase)
 
-        # choose action according to the exploration policy and the current phase (evaluating or training the agent)
-        if phase == RunPhase.TRAIN:
-            action = self.exploration_policy.get_action(actions_q_values)
-            # NOTE: this next line is not in the parent implementation
-            # NOTE: it could be implemented as a wrapper around the parent since action is returned
-            self.current_episode_actions.append(action)
-        else:
-            action = np.argmax(actions_q_values)
+    def get_prediction(self, curr_state):
+        # get the actions q values and the state embedding
+        embedding, actions_q_values = self.main_network.online_network.predict(
+            self.tf_input_state(curr_state),
+            outputs=[self.main_network.online_network.state_embedding,
+                     self.main_network.online_network.output_heads[0].output]
+        )
 
-        # store the q values statistics for logging
-        self.q_values.add_sample(actions_q_values)
-
-        # store information for plotting interactively (actual plotting is done in agent)
-        if self.tp.visualization.plot_action_values_online:
-            for idx, action_name in enumerate(self.env.actions_description):
-                self.episode_running_info[action_name].append(actions_q_values[idx])
-
-        action_value = {"action_value": actions_q_values[action]}
-        return action, action_value
+        # store the state embedding for inserting it to the DND later
+        self.current_episode_state_embeddings.append(embedding.squeeze())
+        actions_q_values = actions_q_values[0][0]
+        return actions_q_values
 
     def reset_game(self, do_not_reset_env=False):
-        ValueOptimizationAgent.reset_game(self, do_not_reset_env)
+        super().reset_game(do_not_reset_env)
 
-        # make sure we already have at least one episode
-        if self.memory.num_complete_episodes() >= 1 and not self.in_heatup:
         # get the last full episode that we have collected
-            episode = self.memory.get(-2)
-            returns = []
-            for i in range(episode.length()):
-                returns.append(episode.get_transition(i).total_return)
-            # Just to deal with the end of heatup where there might be a case where it ends in a middle
-            # of an episode, and thus when getting the episode out of the ER, it will be a complete one whereas
-            # the other statistics collected here, are collected only during training.
-            returns = returns[-len(self.current_episode_actions):]
+        episode = self.memory.get_last_complete_episode()
+        if episode is not None:
+            # the indexing is only necessary because the heatup can end in the middle of an episode
+            # this won't be required after fixing this so that when the heatup is ended, the episode is closed
+            returns = episode.get_transitions_attribute('total_return')[:len(self.current_episode_state_embeddings)]
+            actions = episode.get_transitions_attribute('action')[:len(self.current_episode_state_embeddings)]
             self.main_network.online_network.output_heads[0].DND.add(self.current_episode_state_embeddings,
-                                                                     self.current_episode_actions, returns)
+                                                                     actions, returns)
 
         self.current_episode_state_embeddings = []
-        self.current_episode_actions = []
 
     def save_model(self, model_id):
         self.main_network.save_model(model_id)

agents/value_optimization_agent.py

@@ -73,5 +73,5 @@ class ValueOptimizationAgent(Agent):
             for idx, action_name in enumerate(self.env.actions_description):
                 self.episode_running_info[action_name].append(actions_q_values[idx])
 
-        action_value = {"action_value": actions_q_values[action]}
+        action_value = {"action_value": actions_q_values[action], "max_action_value": np.max(actions_q_values)}
         return action, action_value

architectures/network_wrapper.py

@@ -125,14 +125,15 @@ class NetworkWrapper(object):
         """
         self.online_network.apply_gradients(self.online_network.accumulated_gradients)
 
-    def train_and_sync_networks(self, inputs, targets):
+    def train_and_sync_networks(self, inputs, targets, additional_fetches=[]):
         """
         A generic training function that enables multi-threading training using a global network if necessary.
         :param inputs: The inputs for the network.
         :param targets: The targets corresponding to the given inputs
+        :param additional_fetches: Any additional tensor the user wants to fetch
         :return: The loss of the training iteration
         """
-        result = self.online_network.accumulate_gradients(inputs, targets)
+        result = self.online_network.accumulate_gradients(inputs, targets, additional_fetches=additional_fetches)
         self.apply_gradients_and_sync_networks()
         return result
 

architectures/tensorflow_components/general_network.py

@@ -56,7 +56,8 @@ class GeneralTensorFlowNetwork(TensorFlowArchitecture):
         # the observation can be either an image or a vector
         def get_observation_embedding(with_timestep=False):
             if self.input_height > 1:
-                return ImageEmbedder((self.input_height, self.input_width, self.input_depth), name="observation")
+                return ImageEmbedder((self.input_height, self.input_width, self.input_depth), name="observation",
+                                     input_rescaler=self.tp.agent.input_rescaler)
             else:
                 return VectorEmbedder((self.input_width + int(with_timestep), self.input_depth), name="observation")
 
@@ -191,7 +192,7 @@ class GeneralTensorFlowNetwork(TensorFlowArchitecture):
             if tuning_parameters.agent.optimizer_type == 'Adam':
                 self.optimizer = tf.train.AdamOptimizer(learning_rate=tuning_parameters.learning_rate)
             elif tuning_parameters.agent.optimizer_type == 'RMSProp':
-                self.optimizer = tf.train.RMSPropOptimizer(self.tp.learning_rate, decay=0.9, epsilon=0.01)
+                self.optimizer = tf.train.RMSPropOptimizer(tuning_parameters.learning_rate, decay=0.9, epsilon=0.01)
             elif tuning_parameters.agent.optimizer_type == 'LBFGS':
                 self.optimizer = tf.contrib.opt.ScipyOptimizerInterface(self.total_loss, method='L-BFGS-B',
                                                                         options={'maxiter': 25})

architectures/tensorflow_components/heads.py

@@ -58,6 +58,7 @@ class Head(object):
             self.regularizations = force_list(self.regularizations)
             if self.is_local:
                 self.set_loss()
+            self._post_build()
 
         if self.is_local:
             return self.output, self.target, self.input
@@ -76,6 +77,14 @@ class Head(object):
         """
         pass
 
+    def _post_build(self):
+        """
+        Optional function that allows adding any extra definitions after the head has been fully defined
+        For example, this allows doing additional calculations that are based on the loss
+        :return: None
+        """
+        pass
+
     def get_name(self):
         """
         Get a formatted name for the module
@@ -271,6 +280,9 @@ class DNDQHead(Head):
         else:
             self.loss_type = tf.losses.mean_squared_error
         self.tp = tuning_parameters
+        self.dnd_embeddings = [None]*self.num_actions
+        self.dnd_values = [None]*self.num_actions
+        self.dnd_indices = [None]*self.num_actions
 
     def _build_module(self, input_layer):
         # DND based Q head
@@ -281,29 +293,29 @@ class DNDQHead(Head):
         else:
             self.DND = differentiable_neural_dictionary.QDND(
                 self.DND_size, input_layer.get_shape()[-1], self.num_actions, self.new_value_shift_coefficient,
-                key_error_threshold=self.DND_key_error_threshold)
+                key_error_threshold=self.DND_key_error_threshold, learning_rate=self.tp.learning_rate)
 
         # Retrieve info from DND dictionary
-        # self.action = tf.placeholder(tf.int8, [None], name="action")
-        # self.input = self.action
-        self.output = [
+        # We assume that all actions have enough entries in the DND
+        self.output = tf.transpose([
             self._q_value(input_layer, action)
             for action in range(self.num_actions)
-        ]
+        ])
 
     def _q_value(self, input_layer, action):
         result = tf.py_func(self.DND.query,
-                            [input_layer, [action], self.number_of_nn],
-                            [tf.float64, tf.float64])
-        dnd_embeddings = tf.to_float(result[0])
-        dnd_values = tf.to_float(result[1])
+                            [input_layer, action, self.number_of_nn],
+                            [tf.float64, tf.float64, tf.int64])
+        self.dnd_embeddings[action] = tf.to_float(result[0])
+        self.dnd_values[action] = tf.to_float(result[1])
+        self.dnd_indices[action] = result[2]
 
         # DND calculation
-        square_diff = tf.square(dnd_embeddings - tf.expand_dims(input_layer, 1))
+        square_diff = tf.square(self.dnd_embeddings[action] - tf.expand_dims(input_layer, 1))
         distances = tf.reduce_sum(square_diff, axis=2) + [self.l2_norm_added_delta]
         weights = 1.0 / distances
         normalised_weights = weights / tf.reduce_sum(weights, axis=1, keep_dims=True)
-        return tf.reduce_sum(dnd_values * normalised_weights, axis=1)
+        return tf.reduce_sum(self.dnd_values[action] * normalised_weights, axis=1)
 
 
 class NAFHead(Head):

benchmarks/README.md

@@ -116,6 +116,14 @@ python3 coach.py -p Doom_Health_MMC -r
 
 ## NEC
 
+## Pong_NEC
+
+```bash
+python3 coach.py -p Pong_NEC -r
+```
+
+<img src="img/Pong_NEC.png" alt="Pong_NEC" width="400"/>
+
 ## Doom_Basic_NEC
 
 ```bash

coach.py

@@ -34,11 +34,6 @@ import sys
 import subprocess
 from threading import Thread
 
-try:
-    from Queue import Queue, Empty
-except ImportError:
-    from queue import Queue, Empty # for Python 3.x
-
 if len(set(failed_imports)) > 0:
     screen.warning("Warning: failed to import the following packages - {}".format(', '.join(set(failed_imports))))
 
@@ -258,7 +253,8 @@ if __name__ == "__main__":
     # dump documentation
     logger.set_dump_dir(run_dict['experiment_path'], add_timestamp=True)
     if not args.no_summary:
-        atexit.register(logger.print_summary)
+        atexit.register(logger.summarize_experiment)
+        screen.change_terminal_title(logger.experiment_name)
 
     # Single-threaded runs
     if run_dict['num_threads'] == 1:
@@ -300,7 +296,7 @@ if __name__ == "__main__":
            "--worker_hosts={}".format(worker_hosts),
            "--job_name=ps",
         ]
-        parameter_server = Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=1)
+        parameter_server = Popen(cmd)
 
         screen.log_title("*** Distributed Training ***")
         time.sleep(1)
@@ -325,7 +321,7 @@ if __name__ == "__main__":
                             "--job_name=worker",
                             "--load_json={}".format(json_run_dict_path)]
 
-            p = Popen(workers_args, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=1)
+            p = Popen(workers_args)
 
             if i != run_dict['num_threads']:
                 workers.append(p)

configurations.py

@@ -115,6 +115,7 @@ class AgentParameters(Parameters):
     replace_mse_with_huber_loss = False
     load_memory_from_file_path = None
     collect_new_data = True
+    input_rescaler = 255.0
 
     # PPO related params
     target_kl_divergence = 0.01
@@ -154,6 +155,8 @@ class EnvironmentParameters(Parameters):
     desired_observation_width = 76
     desired_observation_height = 60
     normalize_observation = False
+    crop_observation = False
+    random_initialization_steps = 0
     reward_scaling = 1.0
     reward_clipping_min = None
     reward_clipping_max = None
@@ -290,6 +293,8 @@ class Atari(EnvironmentParameters):
     desired_observation_width = 84
     reward_clipping_max = 1.0
     reward_clipping_min = -1.0
+    random_initialization_steps = 30
+    crop_observation = False  # in the original paper the observation is cropped but not in the Nature paper
 
 
 class Doom(EnvironmentParameters):
@@ -355,6 +360,7 @@ class DQN(AgentParameters):
 
 class DDQN(DQN):
     type = 'DDQNAgent'
+    num_steps_between_copying_online_weights_to_target = 30000
 
 
 class DuelingDQN(DQN):
@@ -384,17 +390,19 @@ class QuantileRegressionDQN(DQN):
 
 class NEC(AgentParameters):
     type = 'NECAgent'
-    optimizer_type = 'RMSProp'
+    optimizer_type = 'Adam'
     input_types = {'observation': InputTypes.Observation}
     output_types = [OutputTypes.DNDQ]
     loss_weights = [1.0]
     dnd_size = 500000
     l2_norm_added_delta = 0.001
-    new_value_shift_coefficient = 0.1
+    new_value_shift_coefficient = 0.1  # alpha
     number_of_knn = 50
     n_step = 100
     bootstrap_total_return_from_old_policy = True
-    DND_key_error_threshold = 0.1
+    DND_key_error_threshold = 0
+    input_rescaler = 1.0
+    num_consecutive_playing_steps = 4
 
 
 class ActorCritic(AgentParameters):

environments/gym_environment_wrapper.py

@@ -19,6 +19,7 @@ from logger import *
 import gym
 import numpy as np
 import time
+import random
 try:
     import roboschool
     from OpenGL import GL
@@ -59,7 +60,7 @@ class GymEnvironmentWrapper(EnvironmentWrapper):
         # self.env_spec = gym.spec(self.env_id)
         self.env.frameskip = self.frame_skip
         self.discrete_controls = type(self.env.action_space) != gym.spaces.box.Box
-
+        self.random_initialization_steps = 0
         self.state = self.reset(True)['state']
 
         # render
@@ -113,6 +114,7 @@ class GymEnvironmentWrapper(EnvironmentWrapper):
         else:
             self.timestep_limit = None
         self.measurements_size = len(self.step(0)['info'].keys())
+        self.random_initialization_steps = self.tp.env.random_initialization_steps
 
     def _wrap_state(self, state):
         if isinstance(self.env.observation_space, gym.spaces.Dict):
@@ -155,8 +157,9 @@ class GymEnvironmentWrapper(EnvironmentWrapper):
 
     def _preprocess_state(self, state):
         # TODO: move this into wrapper
-        if any(env in self.env_id for env in ["Breakout", "Pong"]):
-            # crop image
+        # crop image for atari games
+        # the image from the environment is 210x160
+        if self.tp.env.crop_observation and hasattr(self.env, 'env') and hasattr(self.env.env, 'ale'):
             state['observation'] = state['observation'][34:195, :, :]
         return state
 
@@ -170,7 +173,16 @@ class GymEnvironmentWrapper(EnvironmentWrapper):
             self.env.seed(self.seed)
 
         self.state = self._wrap_state(self.env.reset())
-        while self.state is None:
+
+        # initialize the number of lives
+        if hasattr(self.env, 'env') and hasattr(self.env.env, 'ale'):
+            self.current_ale_lives = self.env.env.ale.lives()
+
+        # simulate a random initial environment state by stepping for a random number of times between 0 and 30
+        step_count = 0
+        random_initialization_steps = random.randint(0, self.random_initialization_steps)
+        while self.state is None or step_count < random_initialization_steps:
+            step_count += 1
             self.step(0)
 
         return self.state

logger.py

@@ -115,6 +115,14 @@ class ScreenLogger(object):
                 if default is not None:
                     return default
 
+    def change_terminal_title(self, title: str):
+        """
+        Changes the title of the terminal window
+        :param title: The new title
+        :return: None
+        """
+        print("\x1b]2;{}\x07".format(title))
+
 
 class BaseLogger(object):
     def __init__(self):
@@ -157,6 +165,7 @@ class Logger(BaseLogger):
         self.time = None
         self.experiments_path = ""
         self.last_line_idx_written_to_csv = 0
+        self.experiment_name = ""
 
     def set_current_time(self, time):
         self.time = time
@@ -205,7 +214,9 @@ class Logger(BaseLogger):
 
     def signal_value_exists(self, time, signal_name):
         try:
-            self.get_signal_value(time, signal_name)
+            value = self.get_signal_value(time, signal_name)
+            if value != value:  # value is nan
+                return False
         except:
             return False
         return True
@@ -229,7 +240,8 @@ class Logger(BaseLogger):
         if self.start_time:
             self.create_signal_value('Wall-Clock Time', time.time() - self.start_time, time=episode)
         else:
-            self.create_signal_value('Wall-Clock Time', time.time(), time=episode)
+            self.create_signal_value('Wall-Clock Time', 0, time=episode)
+            self.start_time = time.time()
 
     def create_gif(self, images, fps=10, name="Gif"):
         output_file = '{}_{}.gif'.format(datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S'), name)
@@ -243,7 +255,7 @@ class Logger(BaseLogger):
     def remove_experiment_dir(self):
         shutil.rmtree(self.experiments_path)
 
-    def print_summary(self):
+    def summarize_experiment(self):
         screen.separator()
         screen.log_title("Results stored at: {}".format(self.experiments_path))
         screen.log_title("Total runtime: {}".format(datetime.datetime.now() - self.time_started))
@@ -273,7 +285,8 @@ class Logger(BaseLogger):
                 screen.error('Experiment name must be composed only of alphanumeric letters, '
                              'underscores and dashes and should not be longer than 100 characters.')
 
-        return match.group(0)
+        self.experiment_name = match.group(0)
+        return self.experiment_name
 
     def get_experiment_path(self, experiment_name, create_path=True):
         general_experiments_path = os.path.join('./experiments/', experiment_name)

memories/differentiable_neural_dictionary.py

@@ -83,6 +83,11 @@ class AnnoyDictionary(object):
 
     # Returns the stored embeddings and values of the closest embeddings
     def query(self, keys, k):
+        if not self.has_enough_entries(k):
+            # this will only happen when the DND is not yet populated with enough entries, which is only during heatup
+            # these values won't be used and therefore they are meaningless
+            return [0.0], [0.0], [0]
+
         _, indices = self._get_k_nearest_neighbors_indices(keys, k)
 
         embeddings = []
@@ -94,7 +99,7 @@ class AnnoyDictionary(object):
 
         self.current_timestamp += 1
 
-        return embeddings, values
+        return embeddings, values, indices
 
     def has_enough_entries(self, k):
         return self.curr_size > k and (self.built_capacity > k)
@@ -133,9 +138,11 @@ class AnnoyDictionary(object):
 
 
 class QDND:
-    def __init__(self, dict_size, key_width, num_actions, new_value_shift_coefficient=0.1, key_error_threshold=0.01):
+    def __init__(self, dict_size, key_width, num_actions, new_value_shift_coefficient=0.1, key_error_threshold=0.01,
+                 learning_rate=0.01):
         self.num_actions = num_actions
         self.dicts = []
+        self.learning_rate = learning_rate
 
         # create a dict for each action
         for a in range(num_actions):
@@ -155,16 +162,18 @@ class QDND:
             self.dicts[a].add(curr_action_embeddings, curr_action_values)
         return True
 
-    def query(self, embeddings, actions, k):
+    def query(self, embeddings, action, k):
         # query for nearest neighbors to the given embeddings
         dnd_embeddings = []
         dnd_values = []
-        for i, action in enumerate(actions):
-            embedding, value = self.dicts[action].query([embeddings[i]], k)
+        dnd_indices = []
+        for i in range(len(embeddings)):
+            embedding, value, indices = self.dicts[action].query([embeddings[i]], k)
             dnd_embeddings.append(embedding[0])
             dnd_values.append(value[0])
+            dnd_indices.append(indices[0])
 
-        return dnd_embeddings, dnd_values
+        return dnd_embeddings, dnd_values, dnd_indices
 
     def has_enough_entries(self, k):
         # check if each of the action dictionaries has at least k entries
@@ -193,4 +202,5 @@ def load_dnd(model_dir):
                 DND.dicts[a].index.add_item(idx, key)
 
             DND.dicts[a].index.build(50)
+
     return DND

memories/episodic_experience_replay.py

@@ -16,6 +16,7 @@
 
 from memories.memory import *
 import threading
+from typing import Union
 
 
 class EpisodicExperienceReplay(Memory):
@@ -103,7 +104,8 @@ class EpisodicExperienceReplay(Memory):
         if transition.game_over:
             self._num_transitions_in_complete_episodes += last_episode.length()
             self._length += 1
-            self.buffer[-1].update_returns(self.discount, is_bootstrapped=self.return_is_bootstrapped,
+            self.buffer[-1].update_returns(self.discount,
+                                           is_bootstrapped=self.tp.agent.bootstrap_total_return_from_old_policy,
                                            n_step_return=self.tp.agent.n_step)
             self.buffer[-1].update_measurements_targets(self.tp.agent.num_predicted_steps_ahead)
             # self.buffer[-1].update_actions_probabilities() # used for off-policy policy optimization
@@ -146,6 +148,17 @@ class EpisodicExperienceReplay(Memory):
     def get(self, index):
         return self.get_episode(index)
 
+    def get_last_complete_episode(self) -> Union[None, Episode]:
+        """
+        Returns the last complete episode in the memory or None if there are no complete episodes
+        :return: None or the last complete episode
+        """
+        last_complete_episode_index = self.num_complete_episodes()-1
+        if last_complete_episode_index >= 0:
+            return self.get(last_complete_episode_index)
+        else:
+            return None
+
     def update_last_transition_info(self, info):
         episode = self.buffer[-1]
         if episode.length() == 0:

memories/memory.py

@@ -80,9 +80,12 @@ class Episode(object):
             total_return += current_discount * np.pad(rewards[i:], (0, i), 'constant', constant_values=0)
             current_discount *= discount
 
+        # calculate the bootstrapped returns
+        bootstraps = np.array([np.squeeze(t.info['max_action_value']) for t in self.transitions[n_step_return:]])
+        bootstrapped_return = total_return + current_discount * np.pad(bootstraps, (0, n_step_return), 'constant',
+                                                                       constant_values=0)
         if is_bootstrapped:
-            bootstraps = np.array([np.squeeze(t.info['action_value']) for t in self.transitions[n_step_return:]])
-            total_return += current_discount * np.pad(bootstraps, (0, n_step_return), 'constant', constant_values=0)
+            total_return = bootstrapped_return
 
         for transition_idx in range(self.length()):
             self.transitions[transition_idx].total_return = total_return[transition_idx]
@@ -114,7 +117,13 @@ class Episode(object):
         return self.returns_table
 
     def get_returns(self):
-        return [t.total_return for t in self.transitions]
+        return self.get_transitions_attribute('total_return')
+
+    def get_transitions_attribute(self, attribute_name):
+        if hasattr(self.transitions[0], attribute_name):
+            return [t.__dict__[attribute_name] for t in self.transitions]
+        else:
+            raise ValueError("The transitions have no such attribute name")
 
     def to_batch(self):
         batch = []
@@ -141,14 +150,12 @@ class Transition(object):
         :param game_over: A boolean which should be True if the episode terminated after
                           the execution of the action.
         """
-        self.state = copy.deepcopy(state)
-        self.state['observation'] = np.array(self.state['observation'], copy=False)
+        self.state = state
         self.action = action
         self.reward = reward
         self.total_return = None
         if not next_state:
             next_state = state
-        self.next_state = copy.deepcopy(next_state)
-        self.next_state['observation'] = np.array(self.next_state['observation'], copy=False)
+        self.next_state = next_state
         self.game_over = game_over
         self.info = {}

plot_atari.py

@@ -0,0 +1,105 @@
+import argparse
+import matplotlib
+import matplotlib.pyplot as plt
+from dashboard import SignalsFile
+import os
+
+
+class FigureMaker(object):
+    def __init__(self, path, cols, smoothness, signal_to_plot, x_axis):
+        self.experiments_path = path
+        self.environments = self.list_environments()
+        self.cols = cols
+        self.rows = int((len(self.environments) + cols - 1) / cols)
+        self.smoothness = smoothness
+        self.signal_to_plot = signal_to_plot
+        self.x_axis = x_axis
+
+        params = {
+            'axes.labelsize': 8,
+            'font.size': 10,
+            'legend.fontsize': 14,
+            'xtick.labelsize': 8,
+            'ytick.labelsize': 8,
+            'text.usetex': False,
+            'figure.figsize': [16, 30]
+        }
+        matplotlib.rcParams.update(params)
+
+    def list_environments(self):
+        environments = sorted([e.name for e in os.scandir(args.path) if e.is_dir()])
+        filtered_environments = self.filter_environments(environments)
+        return filtered_environments
+
+    def filter_environments(self, environments):
+        filtered_environments = []
+        for idx, environment in enumerate(environments):
+            path = os.path.join(args.path, environment)
+            experiments = [e.name for e in os.scandir(path) if e.is_dir()]
+
+            # take only the last updated experiment directory
+            last_experiment_dir = max([os.path.join(path, root) for root in experiments], key=os.path.getctime)
+
+            # make sure there is a csv file inside it
+            for file_path in os.listdir(last_experiment_dir):
+                full_file_path = os.path.join(last_experiment_dir, file_path)
+                if os.path.isfile(full_file_path) and file_path.endswith('.csv'):
+                    filtered_environments.append((environment, full_file_path))
+
+        return filtered_environments
+
+    def plot_figures(self):
+        for idx, (environment, full_file_path) in enumerate(self.environments):
+            print(environment)
+            axis = plt.subplot(self.rows, self.cols, idx + 1)
+            signals = SignalsFile(full_file_path)
+            signals.change_averaging_window(self.smoothness, force=True, signals=[self.signal_to_plot])
+            steps = signals.bokeh_source.data[self.x_axis]
+            rewards = signals.bokeh_source.data[self.signal_to_plot]
+
+            yloc = plt.MaxNLocator(4)
+            axis.yaxis.set_major_locator(yloc)
+            axis.ticklabel_format(style='sci', axis='x', scilimits=(0, 0))
+            plt.title(environment, fontsize=10, y=1.08)
+            plt.plot(steps, rewards, linewidth=0.8)
+            plt.subplots_adjust(hspace=2.0, wspace=0.4)
+
+    def save_pdf(self, name):
+        plt.savefig(name + ".pdf", bbox_inches='tight')
+
+    def show_figures(self):
+        plt.show()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-p', '--path',
+                        help="(string) Root directory of the experiments",
+                        default=None,
+                        type=str)
+    parser.add_argument('-c', '--cols',
+                        help="(int) Number of plot columns",
+                        default=6,
+                        type=int)
+    parser.add_argument('-s', '--smoothness',
+                        help="(int) Number of consequent episodes to average over",
+                        default=200,
+                        type=int)
+    parser.add_argument('-sig', '--signal',
+                        help="(str) The name of the signal to plot",
+                        default='Evaluation Reward',
+                        type=str)
+    parser.add_argument('-x', '--x_axis',
+                        help="(str) The meaning of the x axis",
+                        default='Total steps',
+                        type=str)
+    parser.add_argument('-pdf', '--pdf',
+                        help="(str) A name of a pdf to save to",
+                        default='atari',
+                        type=str)
+    args = parser.parse_args()
+
+    maker = FigureMaker(args.path, cols=args.cols, smoothness=args.smoothness, signal_to_plot=args.signal, x_axis=args.x_axis)
+    maker.plot_figures()
+    maker.save_pdf(args.pdf)
+    maker.show_figures()

presets.py

@@ -89,7 +89,6 @@ class Doom_Basic_QRDQN(Preset):
         self.num_heatup_steps = 1000
 
 
-
 class Doom_Basic_OneStepQ(Preset):
     def __init__(self):
         Preset.__init__(self, NStepQ, Doom, ExplorationParameters)
@@ -408,8 +407,67 @@ class Breakout_DQN(Preset):
         self.exploration.evaluation_policy = 'EGreedy'
         self.exploration.evaluation_epsilon = 0.05
         self.num_heatup_steps = 50000
+        self.agent.num_consecutive_playing_steps = 4
         self.evaluation_episodes = 1
-        self.evaluate_every_x_episodes = 100
+        self.evaluate_every_x_episodes = 25
+        self.agent.replace_mse_with_huber_loss = True
+        # self.env.crop_observation = True  # TODO: remove
+        # self.rescaling_interpolation_type = 'nearest' # TODO: remove
+
+
+class Breakout_DDQN(Preset):
+    def __init__(self):
+        Preset.__init__(self, DDQN, Atari, ExplorationParameters)
+        self.env.level = 'BreakoutDeterministic-v4'
+        self.agent.num_steps_between_copying_online_weights_to_target = 30000
+        self.learning_rate = 0.00025
+        self.agent.num_transitions_in_experience_replay = 1000000
+        self.exploration.initial_epsilon = 1.0
+        self.exploration.final_epsilon = 0.01
+        self.exploration.epsilon_decay_steps = 1000000
+        self.exploration.evaluation_policy = 'EGreedy'
+        self.exploration.evaluation_epsilon = 0.001
+        self.num_heatup_steps = 50000
+        self.agent.num_consecutive_playing_steps = 4
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 25
+        self.agent.replace_mse_with_huber_loss = True
+
+
+class Breakout_Dueling_DDQN(Preset):
+    def __init__(self):
+        Preset.__init__(self, DDQN, Atari, ExplorationParameters)
+        self.env.level = 'BreakoutDeterministic-v4'
+        self.agent.output_types = [OutputTypes.DuelingQ]
+        self.agent.num_steps_between_copying_online_weights_to_target = 30000
+        self.learning_rate = 0.00025
+        self.agent.num_transitions_in_experience_replay = 1000000
+        self.exploration.initial_epsilon = 1.0
+        self.exploration.final_epsilon = 0.01
+        self.exploration.epsilon_decay_steps = 1000000
+        self.exploration.evaluation_policy = 'EGreedy'
+        self.exploration.evaluation_epsilon = 0.001
+        self.num_heatup_steps = 50000
+        self.agent.num_consecutive_playing_steps = 4
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 25
+        self.agent.replace_mse_with_huber_loss = True
+
+class Alien_DQN(Preset):
+    def __init__(self):
+        Preset.__init__(self, DQN, Atari, ExplorationParameters)
+        self.env.level = 'AlienDeterministic-v4'
+        self.agent.num_steps_between_copying_online_weights_to_target = 10000
+        self.learning_rate = 0.00025
+        self.agent.num_transitions_in_experience_replay = 1000000
+        self.exploration.initial_epsilon = 1.0
+        self.exploration.final_epsilon = 0.1
+        self.exploration.epsilon_decay_steps = 1000000
+        self.exploration.evaluation_policy = 'EGreedy'
+        self.exploration.evaluation_epsilon = 0.05
+        self.num_heatup_steps = 50000
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 5
 
 
 class Breakout_C51(Preset):
@@ -846,7 +904,8 @@ class CartPole_NEC(Preset):
         self.num_heatup_steps = 1000
         self.exploration.epsilon_decay_steps = 1000
         self.exploration.final_epsilon = 0.1
-        self.agent.discount = 1.0
+        self.agent.discount = 0.99
+        self.seed = 0
 
         self.test = True
         self.test_max_step_threshold = 200
@@ -857,10 +916,16 @@ class Doom_Basic_NEC(Preset):
     def __init__(self):
         Preset.__init__(self, NEC, Doom, ExplorationParameters)
         self.env.level = 'basic'
-        self.agent.num_episodes_in_experience_replay = 200
-        self.learning_rate = 0.00025
-        self.num_heatup_steps = 1000
-        self.agent.num_playing_steps_between_two_training_steps = 1
+        self.learning_rate = 0.00001
+        self.agent.num_transitions_in_experience_replay = 100000
+        # self.exploration.initial_epsilon = 0.1  # TODO: try exploration
+        # self.exploration.final_epsilon = 0.1
+        # self.exploration.epsilon_decay_steps = 1000000
+        self.num_heatup_steps = 200
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 5
+        self.seed = 123
+
 
 
 class Montezuma_NEC(Preset):
@@ -877,12 +942,20 @@ class Breakout_NEC(Preset):
     def __init__(self):
         Preset.__init__(self, NEC, Atari, ExplorationParameters)
         self.env.level = 'BreakoutDeterministic-v4'
-        self.learning_rate = 0.00025
+        self.agent.num_steps_between_copying_online_weights_to_target = 10000
+        self.learning_rate = 0.00001
         self.agent.num_transitions_in_experience_replay = 1000000
-        self.exploration.initial_epsilon = 1.0
+        self.exploration.initial_epsilon = 0.1
         self.exploration.final_epsilon = 0.1
         self.exploration.epsilon_decay_steps = 1000000
-        self.num_heatup_steps = 50000
+        self.exploration.evaluation_policy = 'EGreedy'
+        self.exploration.evaluation_epsilon = 0.05
+        self.num_heatup_steps = 1000
+        self.env.reward_clipping_max = None
+        self.env.reward_clipping_min = None
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 25
+        self.seed = 123
 
 
 class Doom_Health_NEC(Preset):
@@ -924,12 +997,54 @@ class Pong_NEC(Preset):
     def __init__(self):
         Preset.__init__(self, NEC, Atari, ExplorationParameters)
         self.env.level = 'PongDeterministic-v4'
-        self.learning_rate = 0.001
+        self.learning_rate = 0.00001
         self.agent.num_transitions_in_experience_replay = 100000
-        self.exploration.initial_epsilon = 0.5
+        self.exploration.initial_epsilon = 0.1  # TODO: try exploration
         self.exploration.final_epsilon = 0.1
         self.exploration.epsilon_decay_steps = 1000000
+        self.num_heatup_steps = 2000
+        self.env.reward_clipping_max = None
+        self.env.reward_clipping_min = None
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 5
+        self.env.crop_observation = True  # TODO: remove
+        self.env.random_initialization_steps = 1  # TODO: remove
+        # self.seed = 123
+
+
+class Alien_NEC(Preset):
+    def __init__(self):
+        Preset.__init__(self, NEC, Atari, ExplorationParameters)
+        self.env.level = 'AlienDeterministic-v4'
+        self.learning_rate = 0.0001
+        self.agent.num_transitions_in_experience_replay = 100000
+        self.exploration.initial_epsilon = 0.1  # TODO: try exploration
+        self.exploration.final_epsilon = 0.1
+        self.exploration.epsilon_decay_steps = 1000000
+        self.num_heatup_steps = 3000
+        self.env.reward_clipping_max = None
+        self.env.reward_clipping_min = None
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 5
+        self.seed = 123
+
+
+class Pong_DQN(Preset):
+    def __init__(self):
+        Preset.__init__(self, DQN, Atari, ExplorationParameters)
+        self.env.level = 'PongDeterministic-v4'
+        self.agent.num_steps_between_copying_online_weights_to_target = 10000
+        self.learning_rate = 0.00025
+        self.agent.num_transitions_in_experience_replay = 1000000
+        self.exploration.initial_epsilon = 1.0
+        self.exploration.final_epsilon = 0.1
+        self.exploration.epsilon_decay_steps = 1000000
+        self.exploration.evaluation_policy = 'EGreedy'
+        self.exploration.evaluation_epsilon = 0.05
         self.num_heatup_steps = 50000
+        self.evaluation_episodes = 1
+        self.evaluate_every_x_episodes = 5
+        self.seed = 123
 
 
 class CartPole_A3C(Preset):

run_test.py

@@ -50,6 +50,9 @@ if __name__ == '__main__':
     parser.add_argument('-v', '--verbose',
                         help="(flag) display verbose logs in the event of an error",
                         action='store_true')
+    parser.add_argument('-l', '--list_presets',
+                        help="(flag) list all the presets that are tested",
+                        action='store_true')
     parser.add_argument('--stop_after_first_failure',
                         help="(flag) stop executing tests after the first error",
                         action='store_true')
@@ -73,6 +76,14 @@ if __name__ == '__main__':
         presets_to_ignore = args.ignore_presets.split(',')
     else:
         presets_to_ignore = []
+
+    if args.list_presets:
+        for idx, preset_name in enumerate(presets_lists):
+            preset = eval('presets.{}()'.format(preset_name))
+            if preset.test and preset_name not in presets_to_ignore:
+                print(preset_name)
+        exit(0)
+
     for idx, preset_name in enumerate(presets_lists):
         preset = eval('presets.{}()'.format(preset_name))
         if preset.test and preset_name not in presets_to_ignore:

utils.py

@@ -21,6 +21,7 @@ import numpy as np
 import threading
 from subprocess import call, Popen
 import signal
+import copy
 
 killed_processes = []
 
@@ -333,6 +334,23 @@ def switch_axes_order(observation, from_type='channels_first', to_type='channels
         return np.transpose(observation, (1, 0))
 
 
+class LazyStack(object):
+    """
+    A lazy version of np.stack which avoids copying the memory until it is
+    needed.
+    """
+
+    def __init__(self, history, axis=None):
+        self.history = copy.copy(history)
+        self.axis = axis
+
+    def __array__(self, dtype=None):
+        array = np.stack(self.history, axis=self.axis)
+        if dtype is not None:
+            array = array.astype(dtype)
+        return array
+
+
 def stack_observation(curr_stack, observation, stack_size):
     """
     Adds a new observation to an existing stack of observations from previous time-steps.