def test_log(self):
log_dir = "./newtest/test_log.txt"
if os.path.exists(log_dir):
shutil.rmtree(log_dir)
logger = Logger(log_dir)
logger.log("test text")
logger.log_performance(1, 1)
logger.log_performance(2, 2)
logger.log_performance(3, 3)
logger.close_files()
logger.plot('aaa')
state = env.reset()
for timestep in range(timesteps):
action = agent.step(state)
next_state, reward, done = env.step(action)
ts = (state, action, reward, next_state, done)
agent.feed(ts)
if timestep % evaluate_every == 0:
rewards = []
state = eval_env.reset()
for _ in range(evaluate_num):
action, _ = agent.eval_step(state)
_, reward, done = env.step(action)
if done:
rewards.append(reward)
logger.log_performance(env.timestep, np.mean(rewards))
# Close files in the logger
logger.close_files()
# Plot the learning curve
logger.plot('DQN')
# Save model
save_dir = 'models/uno_single_dqn'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
saver = tf.train.Saver()
saver.save(sess, os.path.join(save_dir, 'model'))
# The paths for saving the logs and learning curves
log_dir = './experiments/leduc_holdem_cfr_result/'
# Set a global seed
set_global_seed(0)
# Initilize CFR Agent
agent = CFRAgent(env)
agent.load() # If we have saved model, we first load the model
# Evaluate CFR against pre-trained NFSP
eval_env.set_agents([agent, models.load('leduc-holdem-nfsp').agents[0]])
# Init a Logger to plot the learning curve
logger = Logger(log_dir)
for episode in range(episode_num):
agent.train()
print('\rIteration {}'.format(episode), end='')
# Evaluate the performance. Play with NFSP agents.
if episode % evaluate_every == 0:
agent.save() # Save model
logger.log_performance(env.timestep,
tournament(eval_env, evaluate_num)[0])
# Close files in the logger
logger.close_files()
# Plot the learning curve
logger.plot('CFR')
class ExperimentRunner:
def __init__(self, env, eval_env, log_every, save_every, base_dir, config,
training_agent, vs_agent, feed_function, save_function):
self.save_dir = "{}/{}".format(base_dir,
datetime.now().strftime("%Y%m%d"))
self.log_dir = os.path.join(self.save_dir, "logs/")
self.model_dir = os.path.join(self.save_dir, "model/")
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
self.log_every = log_every
self.save_every = save_every
self.config = config
self.env = env
self.eval_env = eval_env
self.agent = training_agent
self.training_agents = [self.agent, vs_agent]
self.env.set_agents(self.training_agents)
self.logger = Logger(self.log_dir)
self.logger.log("CONFIG: ")
self.logger.log(str(config))
self.stat_logger = YanivStatLogger(self.logger)
self.feed_function = feed_function
self.save_function = save_function
self.action_space = utils.JOINED_ACTION_SPACE if config[
'single_step_actions'] else utils.ACTION_SPACE
def feed_game(self, agent, trajectories, player_id):
self.feed_function(agent, trajectories[player_id])
if self.config.get("feed_both_games"):
other_traj = trajectories[player_id +
1 % len(self.training_agents)]
if self.training_agents[player_id +
1 % len(self.training_agents)].use_raw:
self.feed_function(
agent,
list(
map(
lambda t: [t[0], self.action_space[t[1]], *t[2:]],
other_traj,
)))
else:
self.feed_function(agent, other_traj)
def run_training(self, episode_num, eval_every, eval_vs, eval_num):
for episode in trange(episode_num, desc="Episodes", file=sys.stdout):
# Generate data from the environment
trajectories, _ = self.env.run(is_training=True)
self.stat_logger.add_game(trajectories, self.env, 0)
self.feed_game(self.agent, trajectories, 0)
if self.config['feed_both_agents']:
self.feed_game(self.training_agents[1], trajectories, 1)
if episode != 0 and episode % self.log_every == 0:
self.stat_logger.log_stats()
if episode != 0 and episode % self.save_every == 0:
self.save_function(self.agent, self.model_dir)
if episode != 0 and episode % eval_every == 0:
self.logger.log(
"\n\n########## Evaluation {} ##########".format(episode))
self.evaluate_perf(eval_vs, eval_num)
self.evaluate_perf(eval_vs, eval_num)
self.save_function(self.agent, self.model_dir)
def evaluate_perf(self, eval_vs, eval_num):
if isinstance(eval_vs, list):
for vs in eval_vs:
self.run_evaluation(vs, eval_num)
else:
self.run_evaluation(eval_vs, eval_num)
def run_evaluation(self, vs, num):
self.eval_env.set_agents([self.agent, vs])
self.logger.log("eval vs {}".format(vs.__class__.__name__))
r = tournament(self.eval_env, num)
eval_vs = "eval_{}_".format(vs.__class__.__name__)
wandb.log(
{
eval_vs + "payoff": r["payoffs"][0],
eval_vs + "draws": r["draws"],
eval_vs + "roundlen": r["roundlen"],
eval_vs + "assafs": r["assafs"][0],
eval_vs + "win_rate": r["wins"][0] / num,
}, )
self.logger.log("Timestep: {}, avg roundlen: {}".format(
self.env.timestep, r["roundlen"]))
for i in range(self.env.player_num):
self.logger.log(
"Agent {}:\nWins: {}, Draws: {}, Assafs: {}, Payoff: {}".
format(
i,
r["wins"][i],
r["draws"],
r["assafs"][i],
r["payoffs"][i],
))
self.logger.log_performance(self.env.timestep, r["payoffs"][0])
def train_mahjong():
# Make environment
env = rlcard.make('mahjong', config={'seed': 0})
eval_env = rlcard.make('mahjong', config={'seed': 0})
# Set the iterations numbers and how frequently we evaluate the performance
evaluate_every = 1000
evaluate_num = 1000
episode_num = 10000
# The intial memory size
memory_init_size = 1000
# Train the agent every X steps
train_every = 64
# The paths for saving the logs and learning curves
log_dir = './experiments/mahjong_nfsp_result/'
# Set a global seed
set_global_seed(0)
with tf.Session() as sess:
# Initialize a global step
global_step = tf.Variable(0, name='global_step', trainable=False)
# Set up the agents
agents = []
for i in range(env.player_num):
agent = NFSPAgent(sess,
scope='nfsp' + str(i),
action_num=env.action_num,
state_shape=env.state_shape,
hidden_layers_sizes=[512, 512],
anticipatory_param=0.5,
batch_size=256,
rl_learning_rate=0.00005,
sl_learning_rate=0.00001,
min_buffer_size_to_learn=memory_init_size,
q_replay_memory_size=int(1e5),
q_replay_memory_init_size=memory_init_size,
train_every=train_every,
q_train_every=train_every,
q_batch_size=256,
q_mlp_layers=[512, 512])
agents.append(agent)
random_agent = RandomAgent(action_num=eval_env.action_num)
env.set_agents(agents)
eval_env.set_agents(
[agents[0], random_agent, random_agent, random_agent])
# Initialize global variables
sess.run(tf.global_variables_initializer())
# Init a Logger to plot the learning curvefrom rlcard.agents.random_agent import RandomAgent
logger = Logger(log_dir)
for episode in tqdm(range(episode_num)):
# First sample a policy for the episode
for agent in agents:
agent.sample_episode_policy()
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for i in range(env.player_num):
for ts in trajectories[i]:
agents[i].feed(ts)
# Evaluate the performance. Play with random agents.
if episode % evaluate_every == 0:
logger.log_performance(env.timestep,
tournament(eval_env, evaluate_num)[0])
# Close files in the logger
logger.close_files()
# Plot the learning curve
logger.plot('NFSP')
# Save model
save_dir = 'models/mahjong_nfsp'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
saver = tf.train.Saver()
saver.save(sess, os.path.join(save_dir, 'model'))
sess.run(tf.global_variables_initializer())
# Init a Logger to plot the learning curve
logger = Logger(log_dir)
for episode in range(episode_num):
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for ts in trajectories[0]:
agent.feed(ts)
# Evaluate the performance. Play with random agents.
if episode % evaluate_every == 0:
t = tournament(eval_env, evaluate_num)[0]
logger.log_performance(env.timestep, t)
# Close files in the logger
logger.close_files()
# Plot the learning curve
logger.plot('DQN')
# Save model
save_dir = 'models/yaniv_dqn'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
saver = tf.train.Saver()
saver.save(sess, os.path.join(save_dir, 'model'))
i, step_counters[i], sl_loss),
end='')
### Evaluation Step: Evaluate the agent every 100 steps and generate the plot for analysis. ###
# Evaluate the performance. Play with random agents.
if episode % evaluate_every == 0:
reward = 0
for eval_episode in range(evaluate_num):
_, payoffs = eval_env.run(is_training=False)
reward += payoffs[0]
logger.log('\n########## Evaluation ##########')
logger.log('Timestep: {} Average reward is {}'.format(
env.timestep,
float(reward) / evaluate_num))
# Add point to logger
logger.log_performance(env.timestep, float(reward) / evaluate_num)
# Make plot
if episode % 100000 == 0 and episode > 0:
if not os.path.exists(save_dir):
os.makedirs(save_dir)
saver = tf.train.Saver()
saver.save(
sess, os.path.join(save_dir, 'model' + str(episode // 100000)))
# logger.plot('NSFP')
# Make the final plot
# plot(csv_path, figure_path + 'fig.png', 'NSFP')
state_shape=env.state_shape,
batch_size=64,
mlp_layers=[64])
env.set_agents([
agent,
RandomAgent(action_num=env.action_num),
RandomAgent(action_num=env.action_num),
RandomAgent(action_num=env.action_num)
])
eval_env.set_agents([
agent,
RandomAgent(action_num=env.action_num),
RandomAgent(action_num=env.action_num),
RandomAgent(action_num=env.action_num)
])
logger = Logger('.')
for episode in range(100000):
# Generate data from the environment
trajectories, _ = env.run(is_training=True)
# Feed transitions into agent memory, and train the agent
for ts in trajectories[0]:
agent.feed(ts)
# Evaluate the performance. Play with random agents.
if episode % 5000 == 0:
logger.log_performance(env.timestep, tournament(eval_env, 10000)[0])
logger.close_files()
logger.plot('DQN')
