Deep Q-Network implementation

This notebook shamelessly demands you to implement a DQN - an approximate q-learning algorithm with experience replay and target networks - and see if it works any better this way.



In [ ]:

    
import sys, os
if 'google.colab' in sys.modules:
    %tensorflow_version 1.x
    
    if not os.path.exists('.setup_complete'):
        !wget -q https://raw.githubusercontent.com/yandexdataschool/Practical_RL/spring20/setup_colab.sh -O- | bash

        !wget -q https://raw.githubusercontent.com/yandexdataschool/Practical_RL/spring20/week04_approx_rl/framebuffer.py
        !wget -q https://raw.githubusercontent.com/yandexdataschool/Practical_RL/spring20/week04_approx_rl/replay_buffer.py

        !touch .setup_complete

# This code creates a virtual display to draw game images on.
# It will have no effect if your machine has a monitor.
if type(os.environ.get("DISPLAY")) is not str or len(os.environ.get("DISPLAY")) == 0:
    !bash ../xvfb start
    os.environ['DISPLAY'] = ':1'

Frameworks - we'll accept this homework in any deep learning framework. This particular notebook was designed for tensorflow, but you will find it easy to adapt it to almost any python-based deep learning framework.



In [ ]:

    
import gym
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline

Let's play some old videogames

This time we're gonna apply approximate q-learning to an atari game called Breakout. It's not the hardest thing out there, but it's definitely way more complex than anything we tried before.

Processing game image

Raw atari images are large, 210x160x3 by default. However, we don't need that level of detail in order to learn them.

We can thus save a lot of time by preprocessing game image, including

Resizing to a smaller shape, 64 x 64
Converting to grayscale
Cropping irrelevant image parts (top & bottom)



In [ ]:

    
from gym.core import ObservationWrapper
from gym.spaces import Box

import cv2


class PreprocessAtari(ObservationWrapper):
    def __init__(self, env):
        """A gym wrapper that crops, scales image into the desired shapes and optionally grayscales it."""
        ObservationWrapper.__init__(self, env)

        self.img_size = (64, 64, 1)
        self.observation_space = Box(0.0, 1.0, self.img_size)

    def _to_gray_scale(self, rgb_image, channel_weights=[0.8, 0.1, 0.1]):\n",
        <YOUR CODE>

    def observation(self, img):
        """what happens to each observation"""

        # Here's what you need to do:
        #  * Crop image, remove irrelevant parts.
        #  * Resize image to self.img_size. Use cv2.resize or any other library you want,
        #    e.g. PIL or Keras. Do not use skimage.transform.resize because it is roughly
        #    6x slower than cv2.resize.
        #  * Cast image to grayscale.
        #  * Convert image pixels to (0, 1) range, float32 type.

        <YOUR CODE>
        return <YOUR CODE>



In [ ]:

    
import gym
# spawn game instance for tests
env = gym.make("BreakoutNoFrameskip-v4")  # create raw env
env = PreprocessAtari(env)

observation_shape = env.observation_space.shape
n_actions = env.action_space.n

obs = env.reset()


# test observation
assert obs.ndim == 3, "observation must be [height, width, channels] even if there's just one channel"
assert obs.shape == observation_shape
assert obs.dtype == 'float32'
assert len(np.unique(obs)) > 2, "your image must not be binary"
assert 0 <= np.min(obs) and np.max(obs) <= 1, "convert image pixels to (0,1) range"
assert np.max(obs) >= 0.5, "It would be easier to see a brighter observation"
assert np.mean(obs) >= 0.1, "It would be easier to see a brighter observation"

print("Formal tests seem fine. Here's an example of what you'll get.")

plt.title("what your network gonna see")
plt.imshow(obs, interpolation='none', cmap='gray')

Frame buffer

Our agent can only process one observation at a time, so we gotta make sure it contains enough information to fing optimal actions. For instance, agent has to react to moving objects so he must be able to measure object's velocity.

To do so, we introduce a buffer that stores 4 last images. This time everything is pre-implemented for you.



In [ ]:

    
from framebuffer import FrameBuffer


def make_env():
    env = gym.make("BreakoutNoFrameskip-v4")
    env = PreprocessAtari(env)
    env = FrameBuffer(env, n_frames=4, dim_order='tensorflow')
    return env


env = make_env()
env.reset()
n_actions = env.action_space.n
state_dim = env.observation_space.shape



In [ ]:

    
for _ in range(50):
    obs, _, _, _ = env.step(env.action_space.sample())


plt.title("Game image")
plt.imshow(env.render("rgb_array"))
plt.show()
plt.title("Agent observation (4 frames left to right)")
plt.imshow(obs.transpose([0, 2, 1]).reshape([state_dim[0], -1]))

Building a network

We now need to build a neural network that can map images to state q-values. This network will be called on every agent's step so it better not be resnet-152 unless you have an array of GPUs. Instead, you can use strided convolutions with a small number of features to save time and memory.

You can build any architecture you want, but for reference, here's something that will more or less work:



In [ ]:

    
import tensorflow as tf
tf.reset_default_graph()
sess = tf.InteractiveSession()



In [ ]:

    
from keras.layers import Conv2D, Dense, Flatten


class DQNAgent:
    def __init__(self, name, state_shape, n_actions, epsilon=0, reuse=False):
        """A simple DQN agent"""
        with tf.variable_scope(name, reuse=reuse):

            <YOUR CODE: define your network body here. Please make sure you don't use any layers created elsewhere>

            # prepare a graph for agent step
            self.state_t = tf.placeholder('float32', [None, ] + list(state_shape))
            self.qvalues_t = self.get_symbolic_qvalues(self.state_t)

        self.weights = tf.get_collection(
            tf.GraphKeys.TRAINABLE_VARIABLES, scope=name)
        self.epsilon = epsilon

    def get_symbolic_qvalues(self, state_t):
        """takes agent's observation, returns qvalues. Both are tf Tensors"""
        <YOUR CODE: apply your network layers here>
        qvalues = <YOUR CODE: symbolic tensor for q-values>

        assert tf.is_numeric_tensor(qvalues) and qvalues.shape.ndims == 2, \
            "please return 2d tf tensor of qvalues [you got %s]" % repr(qvalues)
        assert int(qvalues.shape[1]) == n_actions

        return qvalues

    def get_qvalues(self, state_t):
        """Same as symbolic step except it operates on numpy arrays"""
        sess = tf.get_default_session()
        return sess.run(self.qvalues_t, {self.state_t: state_t})

    def sample_actions(self, qvalues):
        """pick actions given qvalues. Uses epsilon-greedy exploration strategy. """
        epsilon = self.epsilon
        batch_size, n_actions = qvalues.shape
        random_actions = np.random.choice(n_actions, size=batch_size)
        best_actions = qvalues.argmax(axis=-1)
        should_explore = np.random.choice([0, 1], batch_size, p=[1-epsilon, epsilon])
        return np.where(should_explore, random_actions, best_actions)



In [ ]:

    
agent = DQNAgent("dqn_agent", state_dim, n_actions, epsilon=0.5)
sess.run(tf.global_variables_initializer())

Now let's try out our agent to see if it raises any errors.



In [ ]:

    
def evaluate(env, agent, n_games=1, greedy=False, t_max=10000):
    """ Plays n_games full games. If greedy, picks actions as argmax(qvalues). Returns mean reward. """
    rewards = []
    for _ in range(n_games):
        s = env.reset()
        reward = 0
        for _ in range(t_max):
            qvalues = agent.get_qvalues([s])
            action = qvalues.argmax(axis=-1)[0] if greedy else agent.sample_actions(qvalues)[0]
            s, r, done, _ = env.step(action)
            reward += r
            if done:
                break

        rewards.append(reward)
    return np.mean(rewards)



In [ ]:

    
evaluate(env, agent, n_games=1)

Experience replay

For this assignment, we provide you with experience replay buffer. If you implemented experience replay buffer in last week's assignment, you can copy-paste it here to get 2 bonus points.

The interface is fairly simple:

exp_replay.add(obs, act, rw, next_obs, done) - saves (s,a,r,s',done) tuple into the buffer
exp_replay.sample(batch_size) - returns observations, actions, rewards, next_observations and is_done for batch_size random samples.
len(exp_replay) - returns number of elements stored in replay buffer.



In [ ]:

    
from replay_buffer import ReplayBuffer
exp_replay = ReplayBuffer(10)

for _ in range(30):
    exp_replay.add(env.reset(), env.action_space.sample(), 1.0, env.reset(), done=False)

obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(5)

assert len(exp_replay) == 10, "experience replay size should be 10 because that's what maximum capacity is"



In [ ]:

    
def play_and_record(agent, env, exp_replay, n_steps=1):
    """
    Play the game for exactly n steps, record every (s,a,r,s', done) to replay buffer. 
    Whenever game ends, add record with done=True and reset the game.
    It is guaranteed that env has done=False when passed to this function.

    PLEASE DO NOT RESET ENV UNLESS IT IS "DONE"

    :returns: return sum of rewards over time
    """
    # initial state
    s = env.framebuffer

    # Play the game for n_steps as per instructions above
    <YOUR CODE>



In [ ]:

    
# testing your code. This may take a minute...
exp_replay = ReplayBuffer(20000)

play_and_record(agent, env, exp_replay, n_steps=10000)

# if you're using your own experience replay buffer, some of those tests may need correction.
# just make sure you know what your code does
assert len(exp_replay) == 10000, (
    "play_and_record should have added exactly 10000 steps, " +
    "but instead added %i") % len(exp_replay)
is_dones = list(zip(*exp_replay._storage))[-1]

assert 0 < np.mean(is_dones) < 0.1, (
    "Please make sure you restart the game whenever it is 'done' " +
    "and record the is_done correctly into the buffer." +
    "Got %f is_done rate over %i steps. [If you think it's your tough luck, just re-run the test]"
) % (np.mean(is_dones), len(exp_replay))

for _ in range(100):
    obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(
        10)
    assert obs_batch.shape == next_obs_batch.shape == (10,) + state_dim
    assert act_batch.shape == (10,), "actions batch should have shape (10,) but is instead %s" % str(act_batch.shape)
    assert reward_batch.shape == (10,), "rewards batch should have shape (10,) but is instead %s" % str(reward_batch.shape)
    assert is_done_batch.shape == (10,), "is_done batch should have shape (10,) but is instead %s" % str(is_done_batch.shape)
    assert [int(i) in (0, 1) for i in is_dones], "is_done should be strictly True or False"
    assert [0 <= a <= n_actions for a in act_batch], "actions should be within [0, n_actions]"

print("Well done!")

Target networks

We also employ the so called "target network" - a copy of neural network weights to be used for reference Q-values:

The network itself is an exact copy of agent network, but it's parameters are not trained. Instead, they are moved here from agent's actual network every so often.

$$ Q_{reference}(s,a) = r + \gamma \cdot \max _{a'} Q_{target}(s',a') $$



In [ ]:

    
target_network = DQNAgent("target_network", state_dim, n_actions)



In [ ]:

    
def load_weigths_into_target_network(agent, target_network):
    """ assign target_network.weights variables to their respective agent.weights values. """
    assigns = []
    for w_agent, w_target in zip(agent.weights, target_network.weights):
        assigns.append(tf.assign(w_target, w_agent, validate_shape=True))
    # tf.get_default_session().run(assigns)
    return assigns



In [ ]:

    
# create the tf copy graph only once.
copy_step = load_weigths_into_target_network(agent, target_network)
sess.run(copy_step)
# check that it works
sess.run([tf.assert_equal(w, w_target) for w, w_target in zip(agent.weights, target_network.weights)])
print("It works!")

Learning with... Q-learning

Here we write a function similar to agent.update from tabular q-learning.



In [ ]:

    
# placeholders that will be fed with exp_replay.sample(batch_size)
obs_ph = tf.placeholder(tf.float32, shape=(None,) + state_dim)
actions_ph = tf.placeholder(tf.int32, shape=[None])
rewards_ph = tf.placeholder(tf.float32, shape=[None])
next_obs_ph = tf.placeholder(tf.float32, shape=(None,) + state_dim)
is_done_ph = tf.placeholder(tf.float32, shape=[None])

is_not_done = 1 - is_done_ph
gamma = 0.99

Take q-values for actions agent just took



In [ ]:

    
current_qvalues = agent.get_symbolic_qvalues(obs_ph)
current_action_qvalues = tf.reduce_sum(tf.one_hot(actions_ph, n_actions) * current_qvalues, axis=1)

Compute Q-learning TD error:

$$ L = { 1 \over N} \sum_i [ Q_{\theta}(s,a) - Q_{reference}(s,a) ] ^2 $$

With Q-reference defined as

$$ Q_{reference}(s,a) = r(s,a) + \gamma \cdot max_{a'} Q_{target}(s', a') $$

Where

$Q_{target}(s',a')$ denotes q-value of next state and next action predicted by target_network
$s, a, r, s'$ are current state, action, reward and next state respectively
$\gamma$ is a discount factor defined two cells above.



In [ ]:

    
next_qvalues_target = <YOUR CODE: compute q-values for NEXT states with target network>
next_state_values_target = <YOUR CODE: compute state values by taking max over next_qvalues_target for all actions>
reference_qvalues = <YOUR CODE: compute Q_reference(s,a) as per formula above>

# Define loss function for sgd.
td_loss = (current_action_qvalues - reference_qvalues) ** 2
td_loss = tf.reduce_mean(td_loss)

train_step = tf.train.AdamOptimizer(1e-3).minimize(td_loss, var_list=agent.weights)



In [ ]:

    
sess.run(tf.global_variables_initializer())



In [ ]:

    
for chk_grad in tf.gradients(reference_qvalues, agent.weights):
    error_msg = "Reference q-values should have no gradient w.r.t. agent weights. Make sure you used target_network qvalues! "
    error_msg += "If you know what you're doing, ignore this assert."
    assert chk_grad is None or np.allclose(sess.run(chk_grad), sess.run(chk_grad * 0)), error_msg

assert tf.gradients(reference_qvalues, is_not_done)[0] is not None, "make sure you used is_not_done"
assert tf.gradients(reference_qvalues, rewards_ph)[0] is not None, "make sure you used rewards"
assert tf.gradients(reference_qvalues, next_obs_ph)[0] is not None, "make sure you used next states"
assert tf.gradients(reference_qvalues, obs_ph)[0] is None, "reference qvalues shouldn't depend on current observation!"  # ignore if you're certain it's ok
print("Splendid!")

Main loop

It's time to put everything together and see if it learns anything.



In [ ]:

    
from tqdm import trange
import pandas as pd
from IPython.display import clear_output
import matplotlib.pyplot as plt
%matplotlib inline

def moving_average(x, span=100, **kw):
    return pd.DataFrame({'x': np.asarray(x)}).x.ewm(span=span, **kw).mean().values

mean_rw_history = []
td_loss_history = []



In [ ]:

    
exp_replay = ReplayBuffer(10**5)
play_and_record(agent, env, exp_replay, n_steps=10000)


def sample_batch(exp_replay, batch_size):
    obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(batch_size)
    return {
        obs_ph: obs_batch,
        actions_ph: act_batch,
        rewards_ph: reward_batch,
        next_obs_ph: next_obs_batch,
        is_done_ph: is_done_batch,
    }



In [ ]:

    
for i in trange(10**5):
    # play
    play_and_record(agent, env, exp_replay, 10)

    # train
    _, loss_t = sess.run([train_step, td_loss], sample_batch(exp_replay, batch_size=64))
    td_loss_history.append(loss_t)

    # adjust agent parameters
    if i % 500 == 0:
        # You could think that loading weights onto a target network is simply
        #     load_weigths_into_target_network(agent, target_network)
        # but actually calling this function repeatedly creates a TF copy operator
        # again and again, which bloats memory consumption with each training step.
        # Instead, you should create 'copy_step' once.
        sess.run(copy_step)
        agent.epsilon = max(agent.epsilon * 0.99, 0.01)
        mean_rw_history.append(evaluate(make_env(), agent, n_games=3))

    if i % 100 == 0:
        clear_output(True)
        print("buffer size = %i, epsilon = %.5f" % (len(exp_replay), agent.epsilon))

        plt.subplot(1, 2, 1)
        plt.title("mean reward per game")
        plt.plot(mean_rw_history)
        plt.grid()

        assert not np.isnan(loss_t)
        plt.figure(figsize=[12, 4])
        plt.subplot(1, 2, 2)
        plt.title("TD loss history (moving average)")
        plt.plot(moving_average(np.array(td_loss_history), span=100, min_periods=100))
        plt.grid()
        plt.show()



In [ ]:

    
assert np.mean(mean_rw_history[-10:]) > 10.
print("That's good enough for tutorial.")

How to interpret plots:

This aint no supervised learning so don't expect anything to improve monotonously.

TD loss is the MSE between agent's current Q-values and target Q-values. It may slowly increase or decrease, it's ok. The "not ok" behavior includes going NaN or stayng at exactly zero before agent has perfect performance.
mean reward is the expected sum of r(s,a) agent gets over the full game session. It will oscillate, but on average it should get higher over time (after a few thousand iterations...).
- In basic q-learning implementation it takes 5-10k steps to "warm up" agent before it starts to get better.
buffer size - this one is simple. It should go up and cap at max size.
epsilon - agent's willingness to explore. If you see that agent's already at 0.01 epsilon before it's average reward is above 0 - it means you need to increase epsilon. Set it back to some 0.2 - 0.5 and decrease the pace at which it goes down.
Also please ignore first 100-200 steps of each plot - they're just oscillations because of the way moving average works.

At first your agent will lose quickly. Then it will learn to suck less and at least hit the ball a few times before it loses. Finally it will learn to actually score points.

Training will take time. A lot of it actually. An optimistic estimate is to say it's gonna start winning (average reward > 10) after 10k steps.

But hey, look on the bright side of things:

Video



In [ ]:

    
# Don't forget to reset epsilon back to previous value if you want to go on training
agent.epsilon = 0



In [ ]:

    
# Record sessions

import gym.wrappers

with gym.wrappers.Monitor(make_env(), directory="videos", force=True) as env_monitor:
    sessions = [evaluate(env_monitor, agent, n_games=1) for _ in range(100)]



In [ ]:

    
# Show video. This may not work in some setups. If it doesn't
# work for you, you can download the videos and view them locally.

from pathlib import Path
from IPython.display import HTML

video_names = sorted([s for s in Path('videos').iterdir() if s.suffix == '.mp4'])

HTML("""
<video width="640" height="480" controls>
  <source src="{}" type="video/mp4">
</video>
""".format(video_names[-1]))  # You can also try other indices

Assignment part I (5 pts)

We'll start by implementing target network to stabilize training.

To do that you should use TensorFlow functionality.

We recommend thoroughly debugging your code on simple tests before applying it in atari dqn.

Bonus I (2+ pts)

Implement and train double q-learning.

This task contains of

Implementing double q-learning or dueling q-learning or both (see tips below)
Training a network till convergence
- Full points will be awwarded if your network gets average score of >=10 (see "evaluating results")
- Higher score = more points as usual
- If you're running out of time, it's okay to submit a solution that hasn't converged yet and updating it when it converges. Lateness penalty will not increase for second submission, so submitting first one in time gets you no penalty.

Tips:

Implementing double q-learning shouldn't be a problem if you've already have target networks in place.
- You will probably need tf.argmax to select best actions
- Here's an original article
Dueling architecture is also quite straightforward if you have standard DQN.
- You will need to change network architecture, namely the q-values layer
- It must now contain two heads: V(s) and A(s,a), both dense layers
- You should then add them up via elemwise sum layer.
- Here's an article

Bonus II (5+ pts): Prioritized experience replay

In this section, you're invited to implement prioritized experience replay

You will probably need to provide a custom data structure
Once pool.update is called, collect the pool.experience_replay.observations, actions, rewards and is_alive and store them in your data structure
You can now sample such transitions in proportion to the error (see article) for training.

It's probably more convenient to explicitly declare inputs for "sample observations", "sample actions" and so on to plug them into q-learning.

Prioritized (and even normal) experience replay should greatly reduce amount of game sessions you need to play in order to achieve good performance.

While it's effect on runtime is limited for atari, more complicated envs (further in the course) will certainly benefit for it.

Prioritized experience replay only supports off-policy algorithms, so pls enforce n_steps=1 in your q-learning reference computation (default is 10).



In [ ]: