def train(args):
# Make environments, CFR only supports Leduc Holdem
env = rlcard.make('leduc-holdem', config={'seed': 0, 'allow_step_back':True})
eval_env = rlcard.make('leduc-holdem', config={'seed': 0})
# Seed numpy, torch, random
set_seed(args.seed)
# Initilize CFR Agent
agent = CFRAgent(env, os.path.join(args.log_dir, 'cfr_model'))
agent.load() # If we have saved model, we first load the model
# Evaluate CFR against random
eval_env.set_agents([agent, RandomAgent(num_actions=env.num_actions)])
# Start training
with Logger(args.log_dir) as logger:
for episode in range(args.num_episodes):
agent.train()
print('\rIteration {}'.format(episode), end='')
# Evaluate the performance. Play with Random agents.
if episode % args.evaluate_every == 0:
agent.save() # Save model
logger.log_performance(env.timestep, tournament(eval_env, args.num_eval_games)[0])
# Get the paths
csv_path, fig_path = logger.csv_path, logger.fig_path
# Plot the learning curve
plot_curve(csv_path, fig_path, 'cfr')
def load_model(model_path, env=None, position=None, device=None):
if os.path.isfile(model_path): # Torch model
import torch
agent = torch.load(model_path, map_location=device)
agent.set_device(device)
elif os.path.isdir(model_path): # CFR model
from rlcard.agents import CFRAgent
agent = CFRAgent(env, model_path)
agent.load()
elif model_path == 'random': # Random model
from rlcard.agents import RandomAgent
agent = RandomAgent(num_actions=env.num_actions)
else: # A model in the model zoo
from rlcard import models
agent = models.load(model_path).agents[position]
return agent
def train_uno():
# Make environment and enable human mode
env = rlcard.make('uno', config={'seed': 0, 'allow_step_back':True})
eval_env = rlcard.make('uno', config={'seed': 0})
# Set the iterations numbers and how frequently we evaluate the performance and save model
evaluate_every = 100
save_plot_every = 1000
evaluate_num = 10000
episode_num = 10000
# The paths for saving the logs and learning curves
log_dir = './experiments/uno_cfr_result/'
# Set a global seed
set_global_seed(0)
model_path = 'models/uno_cfr'
# Initilize CFR Agent
agent = CFRAgent(env,model_path = model_path)
agent.load() # If we have saved model, we first load the model
# Evaluate CFR against pre-trained NFSP
random_agent = RandomAgent(action_num=eval_env.action_num)
eval_env.set_agents([agent, random_agent])
# 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:
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 LeducHoldemCFRModel(Model):
''' A pretrained model on Leduc Holdem with CFR (chance sampling)
'''
def __init__(self):
''' Load pretrained model
'''
env = rlcard.make('leduc-holdem')
self.agent = CFRAgent(env, model_path=os.path.join(ROOT_PATH, 'leduc_holdem_cfr'))
self.agent.load()
@property
def agents(self):
''' Get a list of agents for each position in a the game
Returns:
agents (list): A list of agents
Note: Each agent should be just like RL agent with step and eval_step
functioning well.
'''
return [self.agent, self.agent]
# Set the iterations numbers and how frequently we evaluate the performance and save model
evaluate_every = 100
save_plot_every = 1000
evaluate_num = 10000
episode_num = 10000
# 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])