def trace2mem(args):
from deeprl_hw2.preprocessors import AtariPreprocessor
from deeprl_hw2.core import ReplayMemory
import glob
import pickle
memory = ReplayMemory(args)
atari_processor = AtariPreprocessor()
count = 0
for trace_path in glob.glob("%s/*.dmp" % args.trace_dir):
with open(trace_path, 'rb') as tdump:
trace = pickle.load(tdump)
for state, action, reward, done in zip(trace["state"], trace["action"],
trace["reward"], trace["done"]):
processed_state = atari_processor.process_state_for_memory(state)
processed_reward = atari_processor.process_reward(reward)
memory.append(processed_state, action, processed_reward, done)
count += 1
if len(trace["state"]) > len(trace["reward"]):
processed_state = atari_processor.process_state_for_memory(
trace["state"][-1])
memory.append(processed_state, trace["action"][-1], 0,
trace["done"][-1])
count += 1
with open(args.mem_dump, 'wb') as mdump:
print(count)
pickle.dump(memory, mdump)
class DQNAgent:
"""Class implementing DQN.
This is a basic outline of the functions/parameters you will need
in order to implement the DQNAgnet. This is just to get you
started. You may need to tweak the parameters, add new ones, etc.
Feel free to change the functions and funciton parameters that the
class provides.
We have provided docstrings to go along with our suggested API.
Parameters
----------
q_network: keras.models.Model
Your Q-network model.
preprocessor: deeprl_hw2.core.Preprocessor
The preprocessor class. See the associated classes for more
details.
memory: deeprl_hw2.core.Memory
Your replay memory.
gamma: float
Discount factor.
target_update_freq: float
Frequency to update the target network. You can either provide a
number representing a soft target update (see utils.py) or a
hard target update (see utils.py and Atari paper.)
num_burn_in: int
Before you begin updating the Q-network your replay memory has
to be filled up with some number of samples. This number says
how many.
train_freq: int
How often you actually update your Q-Network. Sometimes
stability is improved if you collect a couple samples for your
replay memory, for every Q-network update that you run.
batch_size: int
How many samples in each minibatch.
"""
def __init__(self, q_network, target_netwrok, policy, gamma, num_burn_in,
train_freq, batch_size, config):
self.q = q_network
self.q_target = target_netwrok
self.memory = ReplayMemory(config)
self.policy = policy
self.gamma = gamma
self.num_burn_in = num_burn_in
self.train_freq = train_freq
self.batch_size = batch_size
self.currentIter = 0
self.currentEps = 0
self.currentReward = 0
self.config = config
#####
self.historyPre = HistoryPreprocessor(config)
self.AtariPre = AtariPreprocessor(config)
pass
def compile(self, optimizer, loss_func):
"""Setup all of the TF graph variables/ops.
This is inspired by the compile method on the
keras.models.Model class.
This is a good place to create the target network, setup your
loss function and any placeholders you might need.
You should use the mean_huber_loss function as your
loss_function. You can also experiment with MSE and other
losses.
The optimizer can be whatever class you want. We used the
keras.optimizers.Optimizer class. Specifically the Adam
optimizer.
"""
pass
def calc_q_values(self, state, network):
"""Given a state (or batch of states) calculate the Q-values.
Basically run your network on these states.
Return
------
Q-values for the state(s)
"""
state_pre = np.zeros((1, 4, 84, 84), dtype=np.float32)
state_pre[0] = state
q_values = network.predict(state_pre, batch_size=1)[0]
return q_values
def select_action(self, state, network, **kwargs):
"""Select the action based on the current state.
You will probably want to vary your behavior here based on
which stage of training your in. For example, if you're still
collecting random samples you might want to use a
UniformRandomPolicy.
If you're testing, you might want to use a GreedyEpsilonPolicy
with a low epsilon.
If you're training, you might want to use the
LinearDecayGreedyEpsilonPolicy.
This would also be a good place to call
process_state_for_network in your preprocessor.
Returns
--------
selected action
"""
state_pre = np.zeros((1, 4, 84, 84), dtype=np.float32)
state_pre[0] = state
q_values = network.predict(state_pre, batch_size=1)[0]
return self.policy.select_action(q_values)
def fit(self, env, num_iterations, max_episode_length=None):
"""Fit your model to the provided environment.
Its a good idea to print out things like loss, average reward,
Q-values, etc to see if your agent is actually improving.
You should probably also periodically save your network
weights and any other useful info.
This is where you should sample actions from your network,
collect experience samples and add them to your replay memory,
and update your network parameters.
Parameters
----------
env: gym.Env
This is your Atari environment. You should wrap the
environment using the wrap_atari_env function in the
utils.py
num_iterations: int
How many samples/updates to perform.
max_episode_length: int
How long a single episode should last before the agent
resets. Can help exploration.
"""
cnt = np.long(0)
episode_rwd = 0
_screen_raw = self.process_env_reset(env) # Save to history
mse_loss, mae_metric = 0, 0
self.policy = UniformRandomPolicy(env.action_space.n)
evaluation_interval_cnt = 0
while cnt < num_iterations:
cnt += 1
evaluation_interval_cnt += 1
current_state = self.historyPre.get_current_state()
action = self.select_action(current_state, self.q) # Get action
_screen_next_raw, reward, isterminal, _ = env.step(
action) # take action, observe new
episode_rwd += reward
_screen_raw = self.process_one_screen(
_screen_raw, action, reward, _screen_next_raw, isterminal,
True) # Save to history, Memory
# print "\t state: %d, Step: %d, reward: %d, terminal: %d, Observe: %d" \
# % (np.matrix(_screen).sum(), action, reward, isterminal, np.matrix(_screen_next).sum())
# env.render()
if isterminal: # reset
if evaluation_interval_cnt >= self.config.evaluation_interval:
Aver_reward = self.evaluate(env,
self.config.eval_batch_num)
# print ("----------Evaluate, Average reward", Aver_reward)
evaluation_interval_cnt = 0
with open(self.config.rewardlog, "a") as log:
log.write(",".join([
str(int(cnt / self.config.evaluation_interval)),
str(Aver_reward)
]) + "\n")
_screen_raw = self.process_env_reset(env)
# print ("Episode End, iter: ", cnt, "last batch loss: ", mse_loss, 'last mae Metric: ', mae_metric, "Episode reward: ", episode_rwd)
episode_rwd = 0
if cnt >= self.num_burn_in and cnt % self.train_freq == 0: # update
samples = self.AtariPre.process_batch(
self.memory.sample(self.batch_size))
x = np.zeros(
(self.batch_size, self.config.history_length,
self.config.screen_height, self.config.screen_width),
dtype=np.float32)
y = np.zeros((self.batch_size, int(action_size(env))),
dtype=np.float32)
for _index in range(len(samples)):
sample = samples[_index]
x[_index] = np.copy(sample.state)
if sample.is_terminal:
y[_index] = self.calc_q_values(sample.state, self.q)
y[_index][sample.action] = sample.reward
else:
y[_index] = self.calc_q_values(sample.state, self.q)
q_next = max(
self.calc_q_values(
sample.next_state,
self.q_target)) # Use max to update
y[_index][sample.
action] = sample.reward + self.gamma * q_next
mse_loss, mae_metric = self.q.train_on_batch(x, y)
with open(self.config.losslog, "a") as log:
log.write(",".join(
[str(cnt /
4), str(mse_loss),
str(mae_metric)]) + "\n")
# print(cnt, mse_loss, mae_metric)
if cnt % self.config.target_q_update_step == 0: # Set q == q^
self.q_target.set_weights(self.q.get_weights())
if cnt == self.config.memory_size: # change Policy
self.policy = LinearDecayGreedyEpsilonPolicy(
1, 0.05, self.config.decayNum)
if cnt % (num_iterations / 3) == 0: # Save model
TimeStamp = datetime.datetime.strftime(datetime.datetime.now(),
"%y-%m-%d_%H-%M")
self.q.save_weights(
str(self.config.modelname) + '_' + TimeStamp +
'_weights.h5')
return mse_loss, mae_metric, self.q, self.q_target
def process_one_screen(self, screen_raw, action, reward, screen_next_raw,
isterminal, Is_train):
screen_32_next = self.AtariPre.process_state_for_network(
screen_next_raw)
screen_8 = self.AtariPre.process_state_for_memory(screen_raw)
self.historyPre.insert_screen(screen_32_next)
if Is_train:
self.memory.append(screen_8, action, reward, isterminal)
return screen_next_raw
def process_env_reset(self, env):
self.historyPre.reset()
screen_raw = env.reset()
screen_32 = self.AtariPre.process_state_for_network(screen_raw)
self.historyPre.insert_screen(screen_32)
return screen_raw
def evaluate(self, env, num_episodes):
"""Test your agent with a provided environment.
You shouldn't update your network parameters here. Also if you
have any layers that vary in behavior between train/test time
(such as dropout or batch norm), you should set them to test.
Basically run your policy on the environment and collect stats
like cumulative reward, average episode length, etc.
You can also call the render function here if you want to
visually inspect your policy.
"""
eval_policy = GreedyEpsilonPolicy(self.config.epsilon)
cumu_reward = 0
epscnt = 0
while epscnt < num_episodes:
isterminal = False
_screen_raw = self.process_env_reset(env) # Save to history
while not isterminal:
current_state = self.historyPre.get_current_state()
action = self.select_action_test(current_state,
eval_policy) # Get action
_screen_next_raw, reward, isterminal, _ = env.step(
action) # take action, observe new
cumu_reward += reward
_screen_raw = self.process_one_screen(
_screen_raw, action, reward, _screen_next_raw, isterminal,
True) # Save to history, Memory
epscnt += 1
return cumu_reward / num_episodes
def select_action_test(self, state, policy, **kwargs):
"""Select the action based on the current state.
You will probably want to vary your behavior here based on
which stage of training your in. For example, if you're still
collecting random samples you might want to use a
UniformRandomPolicy.
If you're testing, you might want to use a GreedyEpsilonPolicy
with a low epsilon.
If you're training, you might want to use the
LinearDecayGreedyEpsilonPolicy.
This would also be a good place to call
process_state_for_network in your preprocessor.
Returns
--------
selected action
"""
state_pre = np.zeros((1, 4, 84, 84), dtype=np.float32)
state_pre[0] = state
q_values = self.q.predict(state_pre, batch_size=1)[0]
return policy.select_action(q_values)
