def __init__(self, args, init_train_info={}, sub_dir=None):
self.args = args
misc.ensure_dir(args.logdir)
sub_dir = args.continue_from or sub_dir or misc.datetimestr()
self.logdir = os.path.join(args.logdir, sub_dir)
misc.ensure_dir(self.logdir)
self._setup_log_file()
self._create_train_info(args, init_train_info)
def exploration(board,
models,
gpu_id,
tau_func=default_tau_func,
policy_noise_ratio=0,
resign=None,
logger=None):
import misc, json, sys
from ccboard import action_index
history = []
cur_node = Node(board)
winner = None
# from tqdm import trange
game_name = misc.datetimestr()
for step in range(args().max_game_steps):
# misc.progress_bar(step, args().max_game_steps, game_name)
policy, actions = _mcts_policy(cur_node, models[step % 2], gpu_id,
tau_func(step))
pos, action = _next_action(cur_node.s, policy, actions,
policy_noise_ratio)
# history.append( Experience(cur_node.s, policy, cur_node.v) ) # $v$ is here for resignation check later
history.append(
Experience(cur_node.s, action_index(pos, action), cur_node.v))
if resign is not None and v < resign:
winner = (step + 1) % 2 # current player losses
# sys.stderr.write('\n')
break
if logger is not None:
logger.log_game_action(game_name, pos, board.action_list()[action])
cur_node = cur_node.next_edges[(pos, action)].next_node
cur_node.prev_edge = None
if cur_node.s.is_terminated():
# history.append( Experience(cur_node.s, null, cur_node.v) )
# the current step loss if the next step wins
next_step_wins = cur_node.s.is_winner()
winner = (step + 1 if next_step_wins else step) % 2
# sys.stderr.write('\n')
break
if logger is not None:
logger.end_log_game_action(game_name)
# fill scores
min_winner_score = 1
for step in range(len(history)):
if winner is None:
history[step].v = 0
elif step % 2 == winner:
if min_winner_score > history[step].v:
min_winner_score = history[step].v
history[step].v = 1
else:
history[step].v = -1
return history, winner, min_winner_score
def try_reload_model(model_path, best_model, gpu_id, proc_id, iters):
import os
from log import Logger
cur_model_path = Logger.lastest_model_path(args().logdir)
if best_model is None or model_path != cur_model_path:
model_path = cur_model_path
best_model = Logger._load_model(model_path, gpu_id)
print('Exploration GPU%d-%d it:%d %s model updated' % (gpu_id, proc_id, iters, misc.datetimestr()))
return model_path, best_model
def compare_models(iters, eval_model, best_model, gpu_id, logger):
from mcts import exploration
wins = 0
has_cuda = torch.cuda.is_available()
iters2 = 0
for i in range(args().evaluation_games):
_, winner, _ = exploration(ccboard.ChessBoard(), [best_model, eval_model], gpu_id, policy_noise_ratio=args().policy_noise_ratio)
if winner is not None:
wins += winner
iters2 += 1
logger.info("Evaluation %d:%d %s New model %s" % (iters, iters2, misc.datetimestr(), 'wins' if winner == 1 else 'lose'))
return wins / float(args().evaluation_games)
def save_model(self, model, model_path=None):
if isinstance(model, torch.nn.DataParallel):
model = model.module
if model_path is None:
if 'model_path' in self.train_info and os.path.exists(
os.join(self.logdir, self.train_info['model_path'])):
print("Removing old model {}".format(
self.train_info['model_path']))
os.remove(os.join(self.logdir, self.train_info['model_path']))
model_path = misc.datetimestr() + '.model.pth'
self.train_info['model_path'] = model_path
print("Saving model to {}".format(model_path))
model.save(os.path.join(self.logdir, model_path))
def exploration_process_func(gpu_id, proc_id, queue):
import numpy
numpy.random.seed(gpu_id * 101 + proc_id)
from mcts import exploration
model_path = None
best_model = None
has_cuda = torch.cuda.is_available()
iters = 0
while True:
iters += 1
model_path, best_model = try_reload_model(model_path, best_model, gpu_id, proc_id, iters)
history, _, _ = exploration(ccboard.ChessBoard(), [best_model, best_model], gpu_id, policy_noise_ratio=args().policy_noise_ratio)
print('Exploration GPU%d-%d it:%d %s history_size:%d' % (gpu_id, proc_id, iters, misc.datetimestr(), len(history)))
queue.put(history)
def evaluation(iters, best_model, model, model_lock, logger):
has_cuda = torch.cuda.is_available()
gpu_id = args().train_gpu
model_lock.acquire()
eval_model = logger.clone_model(model, gpu_id)
model_lock.release()
wins = compare_models(iters, eval_model, best_model, gpu_id, logger)
logger.info(">>> Evaluation %d %s New model wins %.0f%%" % (iters, misc.datetimestr(), wins * 100))
if wins < 0.55: return best_model
global epoch
best_model = eval_model
logger.train_info['epoch'] = epoch
logger.save_train_info()
logger.save_model(best_model)
return best_model
def train(replay_buffer, model, queue):
parameters = model.parameters()
optimizer = torch.optim.SGD(parameters,
lr=1e-4,
momentum=0.9,
nesterov=True)
max_norm = 400
loss_history = []
from tqdm import tqdm
import matplotlib.pyplot as plt
eval_iters = 1
best_model = model
torch.save(best_model, 'best_model.pth')
for epoch in tqdm(range(500), desc="Training epoch"):
if (epoch + 1) % 100 == 0:
optim_state = optimizer.state_dict()
optim_state['param_groups'][0][
'lr'] = optim_state['param_groups'][0]['lr'] / 1.01
optimizer.load_state_dict(optim_state)
_history, _, _ = explore(ChessBoard(), [best_model, best_model],
max_steps, 0.25)
replay_buffer.extend(_history)
# while not queue.empty():
# replay_buffer.extend(queue.get())
while len(replay_buffer) > 100000:
replay_buffer.pop()
loss = train_epoch(replay_buffer, optimizer, model, max_norm)
import misc
print('\nTrain epoch: %d, Buffer:%d, Time: %s, Loss: %.5f' %
(epoch, len(replay_buffer), misc.datetimestr(), loss))
loss_history.append(loss)
plt.clf()
plt.plot(loss_history)
plt.savefig("loss.png")
if (epoch + 1) % evaluation_interval == 0:
eval_iters += 1
best_model = evaluation(eval_iters, best_model, model)
with open('loss.his', 'wb') as f:
import pickle
pickle.dump(loss, f)
def save_model(self, model, model_path=None):
to_remove = None
if model_path is None:
# remove old
if 'model_path' in self.train_info:
model_path = os.path.join(self.logdir,
self.train_info['model_path'])
if os.path.exists(model_path):
to_remove = model_path
print("Removing old model {}".format(
self.train_info['model_path']))
# new path name
model_path = misc.datetimestr() + '.model.pth'
self.train_info['model_path'] = model_path
print("Saving model to {}".format(model_path))
package = {
'config': self._model_config(),
'state_dict': model.state_dict(),
}
torch.save(package, os.path.join(self.logdir, model_path))
if to_remove is not None:
os.remove(to_remove)
def train(replay_buffer, queue, model, model_lock, logger):
global epoch
parameters = model.parameters()
optimizer = torch.optim.SGD(parameters, lr=args().lr, momentum=args().momentum, nesterov=True)
has_cuda = torch.cuda.is_available()
gpu_id = args().train_gpu
max_norm = args().max_norm
for ep in range(args().epochs):
epoch = ep
if (epoch + 1) % args().anneal_interval == 0:
anneal_lr(optimizer, logger)
if epoch < logger.train_info['epoch']: continue
while not queue.empty():
replay_buffer.extend(queue.get())
while len(replay_buffer) > args().replay_buffer_size: replay_buffer.pop()
loss = train_epoch(replay_buffer, has_cuda, gpu_id, optimizer, model, model_lock, max_norm)
logger.info('Train epoch: %d, Buffer:%d, Time: %s, Loss: %.5f' % (epoch, len(replay_buffer), misc.datetimestr(), loss))
logger.train_info['loss'].append(loss)
if args().plot:
logger.plot_progress()