print("load ... %d" % start_from)
for model in models:
model.load(savedir, start_from)
else:
start_from = 0
# print debug info
print(args)
print("view_space", env.get_view_space(handles[0]))
print("feature_space", env.get_feature_space(handles[0]))
# play
start = time.time()
for k in range(start_from, start_from + args.n_round):
tic = time.time()
eps = buffer.piecewise_decay(k, [0, 700, 1400],
[1, 0.2, 0.05]) if not args.greedy else 0
loss, num, reward, value = play_a_round(
env,
args.map_size,
handles,
models,
train=args.train,
print_every=50,
render=args.render or (k + 1) % args.render_every == 0,
eps=eps) # for e-greedy
log.info("round %d\t loss: %s\t num: %s\t reward: %s\t value: %s" %
(k, loss, num, reward, value))
print("round time %.2f total time %.2f\n" %
(time.time() - tic, time.time() - start))
print("load ... %d" % start_from)
models[0].load(savedir, start_from)
else:
start_from = 0
# print debug info
print(args)
print("view_size", env.get_view_space(handles[0]))
print("feature_size", env.get_feature_space(handles[0]))
# play
start = time.time()
for k in range(start_from, start_from + args.n_round):
tic = time.time()
start = 1 if args.opponent != -1 else 0.1
train_eps = buffer.piecewise_decay(
k, [0, 100, 250], [start, 0.1, 0.05]) if not args.greedy else 0
opponent_eps = train_eps if k < 100 else 0.05 # can use curriculum learning in first 100 steps
loss, num, reward, value = play_a_round(
env,
args.map_size,
handles,
models,
eps=[opponent_eps, train_eps],
step_batch_size=step_batch_size,
train=args.train,
print_every=50,
render=args.render
or (k + 1) % args.render_every == 0) # for e-greedy
log.info("round %d\t loss: %s\t num: %s\t reward: %s\t value: %s" %
print("load ... %d" % start_from)
for model in models:
model.load(savedir, start_from)
else:
start_from = 0
# print state info
print(args)
print("view_space", env.get_view_space(handles[0]))
print("feature_space", env.get_feature_space(handles[0]))
# play
start = time.time()
for k in range(start_from, start_from + args.n_round):
tic = time.time()
eps = buffer.piecewise_decay(k, [0, 600, 1200],
[1, 0.2, 0.1]) if not args.greedy else 0
loss, num, reward, value = play_a_round(
env,
args.map_size,
handles,
models,
train=args.train,
print_every=50,
render=args.render or (k + 1) % args.render_every == 0,
eps=eps) # for e-greedy
log.info("round %d\t loss: %s\t num: %s\t reward: %s\t value: %s" %
(k, loss, num, reward, value))
print("round time %.2f total time %.2f\n" %
(time.time() - tic, time.time() - start))
args.map_size,
food_handle,
player_handles,
models,
-1,
record=True,
render=False,
print_every=args.print_every,
eps=eps)
else:
# play
start = time.time()
train_id = 0 if args.train else -1
for k in range(start_from, start_from + args.n_round):
tic = time.time()
eps = buffer.piecewise_decay(
k, [0, 400, 1200], [1.0, 0.2, 0.10]) if not args.greedy else 0
loss, reward, value, pos_reward_ct, fill_rate = \
play_a_round(env, args.map_size, food_handle, player_handles, models,
train_id, record=False,
render=args.render or (k+1) % args.render_every == 0,
print_every=args.print_every, eps=eps)
log.info(
"round %d\t loss: %.3f\t reward: %.2f\t value: %.3f\t pos_reward_ct: %d\t fill: %.2f"
% (k, loss, reward, value, pos_reward_ct, fill_rate))
print("round time %.2f total time %.2f\n" %
(time.time() - tic, time.time() - start))
if (k + 1) % args.save_every == 0 and args.train:
print("save models...")
for model in models:
model.save(save_dir, k)