def test_logsoftmax():
var = mx.symbol.Variable('var')
data = mx.symbol.Variable('data')
net = mx.symbol.FullyConnected(data=data, name='fc1', num_hidden=10)
net = mx.symbol.Activation(data=net, name='relu1', act_type='relu')
net = mx.symbol.FullyConnected(data=net, name='fc2', num_hidden=4)
net = mx.symbol.Custom(data=net, name='policy', op_type='LogSoftmaxPolicy')
ctx = mx.gpu()
minibatch_size = 100
data_shapes = {
'data': (minibatch_size, 4),
'policy_score': (minibatch_size, )
}
qnet = Base(data_shapes=data_shapes,
sym_gen=net,
name='PolicyNet',
initializer=mx.initializer.Xavier(factor_type="in",
magnitude=1.0),
ctx=ctx)
print qnet.internal_sym_names
lr = 0.00001
optimizer = mx.optimizer.create(name='sgd',
learning_rate=0.00001,
clip_gradient=None,
rescale_grad=1.0,
wd=0.)
updater = mx.optimizer.get_updater(optimizer)
total_iter = 1000000
stats = numpy.zeros((total_iter, 3), dtype=numpy.float32)
plt.ion()
fig, ax = plt.subplots()
lines, = ax.plot([], [])
ax.set_autoscaley_on(True)
baseline = 0
for i in range(total_iter):
data = numpy.random.randn(minibatch_size, 4)
outputs = qnet.forward(is_train=True, data=data)
action = outputs[0].asnumpy()
prob = outputs[1].asnumpy()
#print 'data=', data, 'action=', action, 'prob=', prob
#ch = raw_input()
score = simple_game_discrete(data, action)
baseline = baseline - 0.001 * (baseline - score.mean())
print 'score=', score.mean(), 'acc=', numpy.sum(
action == numpy.argmax(data *
data, axis=1)).mean(), 'baseline=', baseline
stats[i] = [
score.mean(),
numpy.sum(action == numpy.argmax(data * data, axis=1)).mean(),
baseline
]
qnet.backward(policy_score=score - baseline)
qnet.update(updater)
update_line(lines, fig, ax, i,
score.mean()) # numpy.square(means - data*data).mean())
def main():
parser = argparse.ArgumentParser(
description='Script to test the trained network on a game.')
parser.add_argument('-r',
'--rom',
required=False,
type=str,
default=os.path.join('arena', 'games', 'roms',
'breakout.bin'),
help='Path of the ROM File.')
parser.add_argument('-v',
'--visualization',
required=False,
type=int,
default=0,
help='Visualize the runs.')
parser.add_argument('--lr',
required=False,
type=float,
default=0.01,
help='Learning rate of the AdaGrad optimizer')
parser.add_argument('--eps',
required=False,
type=float,
default=0.01,
help='Eps of the AdaGrad optimizer')
parser.add_argument('--clip-gradient',
required=False,
type=float,
default=None,
help='Clip threshold of the AdaGrad optimizer')
parser.add_argument('--double-q',
required=False,
type=bool,
default=False,
help='Use Double DQN')
parser.add_argument('--wd',
required=False,
type=float,
default=0.0,
help='Weight of the L2 Regularizer')
parser.add_argument(
'-c',
'--ctx',
required=False,
type=str,
default='gpu',
help='Running Context. E.g `-c gpu` or `-c gpu1` or `-c cpu`')
parser.add_argument('-d',
'--dir-path',
required=False,
type=str,
default='',
help='Saving directory of model files.')
parser.add_argument(
'--start-eps',
required=False,
type=float,
default=1.0,
help='Eps of the epsilon-greedy policy at the beginning')
parser.add_argument('--replay-start-size',
required=False,
type=int,
default=50000,
help='The step that the training starts')
parser.add_argument(
'--kvstore-update-period',
required=False,
type=int,
default=1,
help='The period that the worker updates the parameters from the sever'
)
parser.add_argument(
'--kv-type',
required=False,
type=str,
default=None,
help=
'type of kvstore, default will not use kvstore, could also be dist_async'
)
parser.add_argument('--optimizer',
required=False,
type=str,
default="adagrad",
help='type of optimizer')
args = parser.parse_args()
if args.dir_path == '':
rom_name = os.path.splitext(os.path.basename(args.rom))[0]
args.dir_path = 'dqn-%s-lr%g' % (rom_name, args.lr)
replay_start_size = args.replay_start_size
max_start_nullops = 30
replay_memory_size = 1000000
history_length = 4
rows = 84
cols = 84
ctx = parse_ctx(args.ctx)
q_ctx = mx.Context(*ctx[0])
game = AtariGame(rom_path=args.rom,
resize_mode='scale',
replay_start_size=replay_start_size,
resized_rows=rows,
resized_cols=cols,
max_null_op=max_start_nullops,
replay_memory_size=replay_memory_size,
display_screen=args.visualization,
history_length=history_length)
##RUN NATURE
freeze_interval = 10000
epoch_num = 200
steps_per_epoch = 250000
update_interval = 4
discount = 0.99
eps_start = args.start_eps
eps_min = 0.1
eps_decay = (eps_start - eps_min) / 1000000
eps_curr = eps_start
freeze_interval /= update_interval
minibatch_size = 32
action_num = len(game.action_set)
data_shapes = {
'data': (minibatch_size, history_length) + (rows, cols),
'dqn_action': (minibatch_size, ),
'dqn_reward': (minibatch_size, )
}
dqn_sym = dqn_sym_nature(action_num)
qnet = Base(data_shapes=data_shapes,
sym_gen=dqn_sym,
name='QNet',
initializer=DQNInitializer(factor_type="in"),
ctx=q_ctx)
target_qnet = qnet.copy(name="TargetQNet", ctx=q_ctx)
use_easgd = False
if args.optimizer != "easgd":
optimizer = mx.optimizer.create(name=args.optimizer,
learning_rate=args.lr,
eps=args.eps,
clip_gradient=args.clip_gradient,
rescale_grad=1.0,
wd=args.wd)
else:
use_easgd = True
easgd_beta = 0.9
easgd_p = 4
easgd_alpha = easgd_beta / (args.kvstore_update_period * easgd_p)
server_optimizer = mx.optimizer.create(name="ServerEASGD",
learning_rate=easgd_alpha)
easgd_eta = 0.00025
local_optimizer = mx.optimizer.create(name='adagrad',
learning_rate=args.lr,
eps=args.eps,
clip_gradient=args.clip_gradient,
rescale_grad=1.0,
wd=args.wd)
central_weight = OrderedDict([(n, nd.zeros(v.shape, ctx=q_ctx))
for n, v in qnet.params.items()])
# Create KVStore
if args.kv_type != None:
kv = kvstore.create(args.kv_type)
#Initialize KVStore
for idx, v in enumerate(qnet.params.values()):
kv.init(idx, v)
# Set Server optimizer on KVStore
if not use_easgd:
kv.set_optimizer(optimizer)
else:
kv.set_optimizer(server_optimizer)
local_updater = mx.optimizer.get_updater(local_optimizer)
kvstore_update_period = args.kvstore_update_period
args.dir_path = args.dir_path + "-" + str(kv.rank)
else:
updater = mx.optimizer.get_updater(optimizer)
qnet.print_stat()
target_qnet.print_stat()
# Begin Playing Game
training_steps = 0
total_steps = 0
for epoch in xrange(epoch_num):
# Run Epoch
steps_left = steps_per_epoch
episode = 0
epoch_reward = 0
start = time.time()
game.start()
while steps_left > 0:
# Running New Episode
episode += 1
episode_loss = 0.0
episode_q_value = 0.0
episode_update_step = 0
episode_action_step = 0
time_episode_start = time.time()
game.begin_episode(steps_left)
while not game.episode_terminate:
# 1. We need to choose a new action based on the current game status
if game.state_enabled and game.replay_memory.sample_enabled:
do_exploration = (npy_rng.rand() < eps_curr)
eps_curr = max(eps_curr - eps_decay, eps_min)
if do_exploration:
action = npy_rng.randint(action_num)
else:
# TODO Here we can in fact play multiple gaming instances simultaneously and make actions for each
# We can simply stack the current_state() of gaming instances and give prediction for all of them
# We need to wait after calling calc_score(.), which makes the program slow
# TODO Profiling the speed of this part!
current_state = game.current_state()
state = nd.array(
current_state.reshape((1, ) + current_state.shape),
ctx=q_ctx) / float(255.0)
qval_npy = qnet.forward(is_train=False,
data=state)[0].asnumpy()
action = numpy.argmax(qval_npy)
episode_q_value += qval_npy[0, action]
episode_action_step += 1
else:
action = npy_rng.randint(action_num)
# 2. Play the game for a single mega-step (Inside the game, the action may be repeated for several times)
game.play(action)
total_steps += 1
# 3. Update our Q network if we can start sampling from the replay memory
# Also, we update every `update_interval`
if total_steps % update_interval == 0 and game.replay_memory.sample_enabled:
# 3.1 Draw sample from the replay_memory
training_steps += 1
episode_update_step += 1
states, actions, rewards, next_states, terminate_flags \
= game.replay_memory.sample(batch_size=minibatch_size)
states = nd.array(states, ctx=q_ctx) / float(255.0)
next_states = nd.array(next_states,
ctx=q_ctx) / float(255.0)
actions = nd.array(actions, ctx=q_ctx)
rewards = nd.array(rewards, ctx=q_ctx)
terminate_flags = nd.array(terminate_flags, ctx=q_ctx)
# 3.2 Use the target network to compute the scores and
# get the corresponding target rewards
if not args.double_q:
target_qval = target_qnet.forward(is_train=False,
data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(target_qval))\
* (1.0 - terminate_flags) * discount
else:
target_qval = target_qnet.forward(is_train=False,
data=next_states)[0]
qval = qnet.forward(is_train=False,
data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(qval))\
* (1.0 - terminate_flags) * discount
outputs = qnet.forward(is_train=True,
data=states,
dqn_action=actions,
dqn_reward=target_rewards)
qnet.backward()
if args.kv_type != None:
if use_easgd:
if total_steps % kvstore_update_period == 0:
for ind, k in enumerate(qnet.params.keys()):
kv.pull(ind,
central_weight[k],
priority=-ind)
qnet.params[k][:] -= easgd_alpha * \
(qnet.params[k] - central_weight[k])
kv.push(ind, qnet.params[k], priority=-ind)
qnet.update(updater=local_updater)
else:
update_on_kvstore(kv, qnet.params,
qnet.params_grad)
else:
qnet.update(updater=updater)
# 3.3 Calculate Loss
diff = nd.abs(
nd.choose_element_0index(outputs[0], actions) -
target_rewards)
quadratic_part = nd.clip(diff, -1, 1)
loss = 0.5 * nd.sum(nd.square(quadratic_part)).asnumpy()[0] +\
nd.sum(diff - quadratic_part).asnumpy()[0]
episode_loss += loss
# 3.3 Update the target network every freeze_interval
# (We can do annealing instead of hard copy)
if training_steps % freeze_interval == 0:
qnet.copy_params_to(target_qnet)
steps_left -= game.episode_step
time_episode_end = time.time()
# Update the statistics
epoch_reward += game.episode_reward
if args.kv_type != None:
info_str = "Node[%d]: " % kv.rank
else:
info_str = ""
info_str += "Epoch:%d, Episode:%d, Steps Left:%d/%d, Reward:%f, fps:%f, Exploration:%f" \
% (epoch, episode, steps_left, steps_per_epoch, game.episode_reward,
game.episode_step / (time_episode_end - time_episode_start), eps_curr)
if episode_update_step > 0:
info_str += ", Avg Loss:%f/%d" % (
episode_loss / episode_update_step, episode_update_step)
if episode_action_step > 0:
info_str += ", Avg Q Value:%f/%d" % (
episode_q_value / episode_action_step, episode_action_step)
logging.info(info_str)
end = time.time()
fps = steps_per_epoch / (end - start)
qnet.save_params(dir_path=args.dir_path, epoch=epoch)
if args.kv_type is not None:
logging.info(
"Node[%d]: Epoch:%d, FPS:%f, Avg Reward: %f/%d" %
(kv.rank, epoch, fps, epoch_reward / float(episode), episode))
else:
logging.info("Epoch:%d, FPS:%f, Avg Reward: %f/%d" %
(epoch, fps, epoch_reward / float(episode), episode))
def test_lognormal():
var = mx.symbol.Variable('var')
data = mx.symbol.Variable('data')
net_mean = mx.symbol.FullyConnected(data=data,
name='fc_mean_1',
num_hidden=20)
net_mean = mx.symbol.Activation(data=net_mean,
name='fc_mean_relu_1',
act_type='relu')
net_mean = mx.symbol.FullyConnected(data=data,
name='fc_mean_2',
num_hidden=20)
net_mean = mx.symbol.Activation(data=net_mean,
name='fc_mean_relu_2',
act_type='relu')
net_mean = mx.symbol.FullyConnected(data=net_mean,
name='fc_mean_3',
num_hidden=10)
net_var = mx.symbol.FullyConnected(data=data,
name='fc_var_1',
num_hidden=10)
net_var = mx.symbol.Activation(data=net_var,
name='fc_var_softplus_1',
act_type='softrelu')
net = mx.symbol.Custom(mean=net_mean,
var=net_var,
name='policy',
deterministic=False,
entropy_regularization=0.01,
op_type='LogNormalPolicy')
ctx = mx.gpu()
minibatch_size = 100
data_shapes = {
'data': (minibatch_size, 10),
'policy_score': (minibatch_size, )
} #, 'var':(minibatch_size,)}
qnet = Base(data_shapes=data_shapes,
sym_gen=net,
name='PolicyNet',
initializer=mx.initializer.Xavier(factor_type="in",
magnitude=1.0),
ctx=ctx)
print qnet.internal_sym_names
lr = 0.01
lr_scheduler = FactorScheduler(1000, 1.0 / 1.5)
optimizer = mx.optimizer.create(
name='sgd',
learning_rate=lr, #momentum=0.9,
clip_gradient=None,
lr_scheduler=lr_scheduler,
rescale_grad=1.0,
wd=0.)
updater = mx.optimizer.get_updater(optimizer)
total_iter = 1000000
stats = numpy.zeros((total_iter, 3), dtype=numpy.float32)
plt.ion()
fig, ax = plt.subplots()
lines, = ax.plot([], [])
ax.set_autoscaley_on(True)
baseline = 0
for i in range(total_iter):
# for k, v in qnet.params.items():
# print k, v.asnumpy()
data = numpy.random.randn(minibatch_size, 10)
means = qnet.compute_internal(sym_name="fc_mean_3_output",
data=data).asnumpy()
vars = qnet.compute_internal(sym_name="fc_var_softplus_1_output",
data=data).asnumpy()
outputs = qnet.forward(
is_train=True,
data=data) #, var=0.5*numpy.ones((minibatch_size, )))
action = outputs[0].asnumpy()
score = simple_game_multimodal(data, action, 1)
baseline = baseline - 0.01 * (baseline - score.mean())
print 'score=', score.mean(), 'err=', numpy.square(
means -
data * data).mean(), 'var=', vars.mean(), 'baseline=', baseline
stats[i] = [
score.mean(),
numpy.square(means - data * data).mean(),
vars.mean()
]
qnet.backward(policy_score=score - baseline)
norm_clipping(qnet.params_grad, 10)
qnet.update(updater)
if i % 10 == 0:
update_line(lines, fig, ax, i,
score.mean()) #numpy.square(means - data*data).mean())
clip_gradient=None,
rescale_grad=1.0 / minibatch_size,
wd=0.00001)
updater = mx.optimizer.get_updater(optimizer)
qnet.print_stat()
baseline = numpy.zeros((time_step, ))
decay_factor = 0.5
for epoch in range(10000):
data = [("data", numpy.random.rand(minibatch_size, 4))]
data_ndarray = {k: nd.array(v, ctx=mx.gpu()) for k, v in data}
outputs = qnet.forward(batch_size=minibatch_size, **data_ndarray)
actions = get_npy_list(outputs[:time_step])
means = get_npy_list(outputs[time_step:(time_step * 2)])
vars = get_npy_list(outputs[(time_step * 2):(time_step * 3)])
scores = simple_sequence_generation_game(dict(data), actions)
scores = [
score * pow(decay_factor, time_step - 1 - i)
for i, score in enumerate(scores)
]
q_estimation = numpy.cumsum(scores[::-1], axis=0)[::-1]
baseline = baseline - 0.01 * (baseline - q_estimation.mean(axis=1))
qnet.backward(
batch_size=minibatch_size,
**dict(data_ndarray.items() +
[("policy_t%d_score" % (i), score)
for i, score in enumerate(q_estimation -
baseline.reshape(time_step, 1))]))
qnet.update(updater)
print 'scores=', numpy.mean(scores), 'mean_score=', numpy.mean(simple_sequence_generation_game(dict(data), means)), \
'baseline=', baseline
def main():
parser = argparse.ArgumentParser(
description='Script to test the trained network on a game.')
parser.add_argument('-r',
'--rom',
required=False,
type=str,
default=os.path.join('arena', 'games', 'roms',
'breakout.bin'),
help='Path of the ROM File.')
parser.add_argument('-v',
'--visualization',
required=False,
type=int,
default=0,
help='Visualize the runs.')
parser.add_argument('--lr',
required=False,
type=float,
default=0.01,
help='Learning rate of the AdaGrad optimizer')
parser.add_argument('--eps',
required=False,
type=float,
default=0.01,
help='Eps of the AdaGrad optimizer')
parser.add_argument('--rms-decay',
required=False,
type=float,
default=0.95,
help='Decay rate of the RMSProp')
parser.add_argument('--clip-gradient',
required=False,
type=float,
default=None,
help='Clip threshold of the AdaGrad optimizer')
parser.add_argument('--double-q',
required=False,
type=bool,
default=False,
help='Use Double DQN')
parser.add_argument('--wd',
required=False,
type=float,
default=0.0,
help='Weight of the L2 Regularizer')
parser.add_argument(
'-c',
'--ctx',
required=False,
type=str,
default=None,
help='Running Context. E.g `-c gpu` or `-c gpu1` or `-c cpu`')
parser.add_argument('-d',
'--dir-path',
required=False,
type=str,
default='',
help='Saving directory of model files.')
parser.add_argument(
'--start-eps',
required=False,
type=float,
default=1.0,
help='Eps of the epsilon-greedy policy at the beginning')
parser.add_argument('--replay-start-size',
required=False,
type=int,
default=50000,
help='The step that the training starts')
parser.add_argument(
'--kvstore-update-period',
required=False,
type=int,
default=16,
help='The period that the worker updates the parameters from the sever'
)
parser.add_argument(
'--kv-type',
required=False,
type=str,
default=None,
help=
'type of kvstore, default will not use kvstore, could also be dist_async'
)
parser.add_argument('--optimizer',
required=False,
type=str,
default="adagrad",
help='type of optimizer')
parser.add_argument('--nactor',
required=False,
type=int,
default=16,
help='number of actor')
parser.add_argument('--exploration-period',
required=False,
type=int,
default=4000000,
help='length of annealing of epsilon greedy policy')
parser.add_argument('--replay-memory-size',
required=False,
type=int,
default=100,
help='size of replay memory')
parser.add_argument('--single-batch-size',
required=False,
type=int,
default=5,
help='batch size for every actor')
parser.add_argument('--symbol',
required=False,
type=str,
default="nature",
help='type of network, nature or nips')
parser.add_argument('--sample-policy',
required=False,
type=str,
default="recent",
help='minibatch sampling policy, recent or random')
parser.add_argument('--epoch-num',
required=False,
type=int,
default=50,
help='number of epochs')
parser.add_argument('--param-update-period',
required=False,
type=int,
default=5,
help='Parameter update period')
parser.add_argument('--resize-mode',
required=False,
type=str,
default="scale",
help='Resize mode, scale or crop')
parser.add_argument('--eps-update-period',
required=False,
type=int,
default=8000,
help='eps greedy policy update period')
parser.add_argument('--server-optimizer',
required=False,
type=str,
default="easgd",
help='type of server optimizer')
parser.add_argument('--nworker',
required=False,
type=int,
default=1,
help='number of kv worker')
parser.add_argument('--easgd-alpha',
required=False,
type=float,
default=0.01,
help='easgd alpha')
args, unknown = parser.parse_known_args()
logging.info(str(args))
if args.dir_path == '':
rom_name = os.path.splitext(os.path.basename(args.rom))[0]
time_str = time.strftime("%m%d_%H%M_%S", time.localtime())
args.dir_path = ('dqn-%s-%d_' % (rom_name,int(args.lr*10**5)))+time_str \
+ "_" + os.environ.get('DMLC_TASK_ID')
logging.info("saving to dir: " + args.dir_path)
if args.ctx == None:
args.ctx = os.environ.get('CTX')
logging.info("Context: %s" % args.ctx)
ctx = re.findall('([a-z]+)(\d*)', args.ctx)
ctx = [(device, int(num)) if len(num) > 0 else (device, 0)
for device, num in ctx]
# Async verision
nactor = args.nactor
param_update_period = args.param_update_period
replay_start_size = args.replay_start_size
max_start_nullops = 30
replay_memory_size = args.replay_memory_size
history_length = 4
rows = 84
cols = 84
q_ctx = mx.Context(*ctx[0])
games = []
for g in range(nactor):
games.append(
AtariGame(rom_path=args.rom,
resize_mode=args.resize_mode,
replay_start_size=replay_start_size,
resized_rows=rows,
resized_cols=cols,
max_null_op=max_start_nullops,
replay_memory_size=replay_memory_size,
display_screen=args.visualization,
history_length=history_length))
##RUN NATURE
freeze_interval = 40000 / nactor
freeze_interval /= param_update_period
epoch_num = args.epoch_num
steps_per_epoch = 4000000 / nactor
discount = 0.99
save_screens = False
eps_start = numpy.ones((3, )) * args.start_eps
eps_min = numpy.array([0.1, 0.01, 0.5])
eps_decay = (eps_start - eps_min) / (args.exploration_period / nactor)
eps_curr = eps_start
eps_id = numpy.zeros((nactor, ))
eps_update_period = args.eps_update_period
eps_update_count = numpy.zeros((nactor, ))
single_batch_size = args.single_batch_size
minibatch_size = nactor * single_batch_size
action_num = len(games[0].action_set)
data_shapes = {
'data': (minibatch_size, history_length) + (rows, cols),
'dqn_action': (minibatch_size, ),
'dqn_reward': (minibatch_size, )
}
if args.symbol == "nature":
dqn_sym = dqn_sym_nature(action_num)
elif args.symbol == "nips":
dqn_sym = dqn_sym_nips(action_num)
else:
raise NotImplementedError
qnet = Base(data_shapes=data_shapes,
sym=dqn_sym,
name='QNet',
initializer=DQNInitializer(factor_type="in"),
ctx=q_ctx)
target_qnet = qnet.copy(name="TargetQNet", ctx=q_ctx)
if args.optimizer == "adagrad":
optimizer = mx.optimizer.create(name=args.optimizer,
learning_rate=args.lr,
eps=args.eps,
clip_gradient=args.clip_gradient,
rescale_grad=1.0,
wd=args.wd)
elif args.optimizer == "rmsprop" or args.optimizer == "rmspropnoncentered":
optimizer = mx.optimizer.create(name=args.optimizer,
learning_rate=args.lr,
eps=args.eps,
clip_gradient=args.clip_gradient,
gamma1=args.rms_decay,
gamma2=0,
rescale_grad=1.0,
wd=args.wd)
lr_decay = (args.lr - 0) / (steps_per_epoch * epoch_num /
param_update_period)
# Create kvstore
use_easgd = False
if args.kv_type != None:
kvType = args.kv_type
kv = kvstore.create(kvType)
#Initialize kvstore
for idx, v in enumerate(qnet.params.values()):
kv.init(idx, v)
if args.server_optimizer == "easgd":
use_easgd = True
easgd_beta = 0.9
easgd_alpha = args.easgd_alpha
server_optimizer = mx.optimizer.create(name="ServerEasgd",
learning_rate=easgd_alpha)
easgd_eta = 0.00025
central_weight = OrderedDict([(n, v.copyto(q_ctx))
for n, v in qnet.params.items()])
kv.set_optimizer(server_optimizer)
updater = mx.optimizer.get_updater(optimizer)
else:
kv.set_optimizer(optimizer)
kvstore_update_period = args.kvstore_update_period
npy_rng = numpy.random.RandomState(123456 + kv.rank)
else:
updater = mx.optimizer.get_updater(optimizer)
qnet.print_stat()
target_qnet.print_stat()
states_buffer_for_act = numpy.zeros(
(nactor, history_length) + (rows, cols), dtype='uint8')
states_buffer_for_train = numpy.zeros(
(minibatch_size, history_length + 1) + (rows, cols), dtype='uint8')
next_states_buffer_for_train = numpy.zeros(
(minibatch_size, history_length) + (rows, cols), dtype='uint8')
actions_buffer_for_train = numpy.zeros((minibatch_size, ), dtype='uint8')
rewards_buffer_for_train = numpy.zeros((minibatch_size, ), dtype='float32')
terminate_flags_buffer_for_train = numpy.zeros((minibatch_size, ),
dtype='bool')
# Begin Playing Game
training_steps = 0
total_steps = 0
ave_fps = 0
ave_loss = 0
time_for_info = time.time()
parallel_executor = concurrent.futures.ThreadPoolExecutor(nactor)
for epoch in xrange(epoch_num):
# Run Epoch
steps_left = steps_per_epoch
episode = 0
epoch_reward = 0
start = time.time()
#
for g, game in enumerate(games):
game.start()
game.begin_episode()
eps_rand = npy_rng.rand()
if eps_rand < 0.4:
eps_id[g] = 0
elif eps_rand < 0.7:
eps_id[g] = 1
else:
eps_id[g] = 2
episode_stats = [EpisodeStat() for i in range(len(games))]
while steps_left > 0:
for g, game in enumerate(games):
if game.episode_terminate:
episode += 1
epoch_reward += game.episode_reward
if args.kv_type != None:
info_str = "Node[%d]: " % kv.rank
else:
info_str = ""
info_str += "Epoch:%d, Episode:%d, Steps Left:%d/%d, Reward:%f, fps:%f, Exploration:%f" \
% (epoch, episode, steps_left, steps_per_epoch, game.episode_reward,
ave_fps, (eps_curr[eps_id[g]]))
info_str += ", Avg Loss:%f" % ave_loss
if episode_stats[g].episode_action_step > 0:
info_str += ", Avg Q Value:%f/%d" % (
episode_stats[g].episode_q_value /
episode_stats[g].episode_action_step,
episode_stats[g].episode_action_step)
if g == 0: logging.info(info_str)
if eps_update_count[g] * eps_update_period < total_steps:
eps_rand = npy_rng.rand()
if eps_rand < 0.4:
eps_id[g] = 0
elif eps_rand < 0.7:
eps_id[g] = 1
else:
eps_id[g] = 2
eps_update_count[g] += 1
game.begin_episode(steps_left)
episode_stats[g] = EpisodeStat()
if total_steps > history_length:
for g, game in enumerate(games):
current_state = game.current_state()
states_buffer_for_act[g] = current_state
states = nd.array(states_buffer_for_act, ctx=q_ctx) / float(255.0)
qval_npy = qnet.forward(batch_size=nactor,
data=states)[0].asnumpy()
actions_that_max_q = numpy.argmax(qval_npy, axis=1)
actions = [0] * nactor
for g, game in enumerate(games):
# 1. We need to choose a new action based on the current game status
if games[g].state_enabled and games[
g].replay_memory.sample_enabled:
do_exploration = (npy_rng.rand() < eps_curr[eps_id[g]])
if do_exploration:
action = npy_rng.randint(action_num)
else:
# TODO Here we can in fact play multiple gaming instances simultaneously and make actions for each
# We can simply stack the current_state() of gaming instances and give prediction for all of them
# We need to wait after calling calc_score(.), which makes the program slow
# TODO Profiling the speed of this part!
action = actions_that_max_q[g]
episode_stats[g].episode_q_value += qval_npy[g, action]
episode_stats[g].episode_action_step += 1
else:
action = npy_rng.randint(action_num)
actions[g] = action
# t0=time.time()
for ret in parallel_executor.map(play_game, zip(games, actions)):
pass
# t1=time.time()
# logging.info("play time: %f" % (t1-t0))
eps_curr = numpy.maximum(eps_curr - eps_decay, eps_min)
total_steps += 1
steps_left -= 1
if total_steps % 100 == 0:
this_time = time.time()
ave_fps = (100 / (this_time - time_for_info))
time_for_info = this_time
# 3. Update our Q network if we can start sampling from the replay memory
# Also, we update every `update_interval`
if total_steps > minibatch_size and \
total_steps % (param_update_period) == 0 and \
games[-1].replay_memory.sample_enabled:
if use_easgd and training_steps % kvstore_update_period == 0:
for paramIndex in range(len(qnet.params)):
k = qnet.params.keys()[paramIndex]
kv.pull(paramIndex,
central_weight[k],
priority=-paramIndex)
qnet.params[k][:] -= easgd_alpha * (qnet.params[k] -
central_weight[k])
kv.push(paramIndex,
qnet.params[k],
priority=-paramIndex)
# 3.1 Draw sample from the replay_memory
for g, game in enumerate(games):
episode_stats[g].episode_update_step += 1
nsample = single_batch_size
i0 = (g * nsample)
i1 = (g + 1) * nsample
if args.sample_policy == "recent":
action, reward, terminate_flag=game.replay_memory.sample_last(batch_size=nsample,\
states=states_buffer_for_train,offset=i0)
elif args.sample_policy == "random":
action, reward, terminate_flag=game.replay_memory.sample_inplace(batch_size=nsample,\
states=states_buffer_for_train,offset=i0)
actions_buffer_for_train[i0:i1] = action
rewards_buffer_for_train[i0:i1] = reward
terminate_flags_buffer_for_train[i0:i1] = terminate_flag
states = nd.array(states_buffer_for_train[:, :-1],
ctx=q_ctx) / float(255.0)
next_states = nd.array(states_buffer_for_train[:, 1:],
ctx=q_ctx) / float(255.0)
actions = nd.array(actions_buffer_for_train, ctx=q_ctx)
rewards = nd.array(rewards_buffer_for_train, ctx=q_ctx)
terminate_flags = nd.array(terminate_flags_buffer_for_train,
ctx=q_ctx)
# 3.2 Use the target network to compute the scores and
# get the corresponding target rewards
if not args.double_q:
target_qval = target_qnet.forward(
batch_size=minibatch_size, data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(target_qval))\
* (1.0 - terminate_flags) * discount
else:
target_qval = target_qnet.forward(
batch_size=minibatch_size, data=next_states)[0]
qval = qnet.forward(batch_size=minibatch_size,
data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(qval))\
* (1.0 - terminate_flags) * discount
outputs = qnet.forward(batch_size=minibatch_size,
is_train=True,
data=states,
dqn_action=actions,
dqn_reward=target_rewards)
qnet.backward(batch_size=minibatch_size)
if args.kv_type is None or use_easgd:
qnet.update(updater=updater)
else:
update_on_kvstore(kv, qnet.params, qnet.params_grad)
# 3.3 Calculate Loss
diff = nd.abs(
nd.choose_element_0index(outputs[0], actions) -
target_rewards)
quadratic_part = nd.clip(diff, -1, 1)
loss = (0.5 * nd.sum(nd.square(quadratic_part)) +
nd.sum(diff - quadratic_part)).asscalar()
if ave_loss == 0:
ave_loss = loss
else:
ave_loss = 0.95 * ave_loss + 0.05 * loss
# 3.3 Update the target network every freeze_interval
# (We can do annealing instead of hard copy)
if training_steps % freeze_interval == 0:
qnet.copy_params_to(target_qnet)
if args.optimizer == "rmsprop" or args.optimizer == "rmspropnoncentered":
optimizer.lr -= lr_decay
if save_screens and training_steps % (
60 * 60 * 2 / param_update_period) == 0:
logging.info("saving screenshots")
for g in range(nactor):
screen = states_buffer_for_train[(
g * single_batch_size), -2, :, :].reshape(
states_buffer_for_train.shape[2:])
cv2.imwrite("screen_" + str(g) + ".png", screen)
training_steps += 1
end = time.time()
fps = steps_per_epoch / (end - start)
qnet.save_params(dir_path=args.dir_path, epoch=epoch)
if args.kv_type != None:
logging.info(
"Node[%d]: Epoch:%d, FPS:%f, Avg Reward: %f/%d" %
(kv.rank, epoch, fps, epoch_reward / float(episode), episode))
else:
logging.info("Epoch:%d, FPS:%f, Avg Reward: %f/%d" %
(epoch, fps, epoch_reward / float(episode), episode))
def main():
parser = argparse.ArgumentParser(description='Script to test the trained network on a game.')
parser.add_argument('-r', '--rom', required=False, type=str,
default=os.path.join('arena', 'games', 'roms', 'breakout.bin'),
help='Path of the ROM File.')
parser.add_argument('-v', '--visualization', required=False, type=int, default=0,
help='Visualize the runs.')
parser.add_argument('--lr', required=False, type=float, default=0.01,
help='Learning rate of the AdaGrad optimizer')
parser.add_argument('--eps', required=False, type=float, default=0.01,
help='Eps of the AdaGrad optimizer')
parser.add_argument('--clip-gradient', required=False, type=float, default=None,
help='Clip threshold of the AdaGrad optimizer')
parser.add_argument('--double-q', required=False, type=bool, default=False,
help='Use Double DQN')
parser.add_argument('--wd', required=False, type=float, default=0.0,
help='Weight of the L2 Regularizer')
parser.add_argument('-c', '--ctx', required=False, type=str, default='gpu',
help='Running Context. E.g `-c gpu` or `-c gpu1` or `-c cpu`')
parser.add_argument('-d', '--dir-path', required=False, type=str, default='',
help='Saving directory of model files.')
parser.add_argument('--start-eps', required=False, type=float, default=1.0,
help='Eps of the epsilon-greedy policy at the beginning')
parser.add_argument('--replay-start-size', required=False, type=int, default=50000,
help='The step that the training starts')
parser.add_argument('--kvstore-update-period', required=False, type=int, default=1,
help='The period that the worker updates the parameters from the sever')
parser.add_argument('--kv-type', required=False, type=str, default=None,
help='type of kvstore, default will not use kvstore, could also be dist_async')
args, unknown = parser.parse_known_args()
if args.dir_path == '':
rom_name = os.path.splitext(os.path.basename(args.rom))[0]
args.dir_path = 'dqn-%s' % rom_name
ctx = re.findall('([a-z]+)(\d*)', args.ctx)
ctx = [(device, int(num)) if len(num) >0 else (device, 0) for device, num in ctx]
replay_start_size = args.replay_start_size
max_start_nullops = 30
replay_memory_size = 1000000
history_length = 4
rows = 84
cols = 84
q_ctx = mx.Context(*ctx[0])
game = AtariGame(rom_path=args.rom, resize_mode='scale', replay_start_size=replay_start_size,
resized_rows=rows, resized_cols=cols, max_null_op=max_start_nullops,
replay_memory_size=replay_memory_size, display_screen=args.visualization,
history_length=history_length)
##RUN NATURE
freeze_interval = 10000
epoch_num = 200
steps_per_epoch = 250000
update_interval = 4
discount = 0.99
eps_start = args.start_eps
eps_min = 0.1
eps_decay = (eps_start - 0.1) / 1000000
eps_curr = eps_start
freeze_interval /= update_interval
minibatch_size = 32
action_num = len(game.action_set)
data_shapes = {'data': (minibatch_size, history_length) + (rows, cols),
'dqn_action': (minibatch_size,), 'dqn_reward': (minibatch_size,)}
#optimizer = mx.optimizer.create(name='sgd', learning_rate=args.lr,wd=args.wd)
optimizer = mx.optimizer.Nop()
dqn_output_op = DQNOutputNpyOp()
dqn_sym = dqn_sym_nature(action_num, dqn_output_op)
qnet = Base(data_shapes=data_shapes, sym=dqn_sym, name='QNet',
initializer=DQNInitializer(factor_type="in"),
ctx=q_ctx)
target_qnet = qnet.copy(name="TargetQNet", ctx=q_ctx)
# Create kvstore
testShape = (1,1686180*100)
testParam = nd.ones(testShape,ctx=q_ctx)
testGrad = nd.zeros(testShape,ctx=q_ctx)
# Create kvstore
if args.kv_type != None:
kvType = args.kv_type
kvStore = kvstore.create(kvType)
#Initialize kvstore
for idx,v in enumerate(qnet.params.values()):
kvStore.init(idx,v);
# Set optimizer on kvstore
kvStore.set_optimizer(optimizer)
kvstore_update_period = args.kvstore_update_period
else:
updater = mx.optimizer.get_updater(optimizer)
# if args.kv_type != None:
# kvType = args.kv_type
# kvStore = kvstore.create(kvType)
# kvStore.init(0,testParam)
# testOptimizer = mx.optimizer.Nop()
# kvStore.set_optimizer(testOptimizer)
# kvstore_update_period = args.kvstore_update_period
qnet.print_stat()
target_qnet.print_stat()
# Begin Playing Game
training_steps = 0
total_steps = 0
while(1):
time_before_wait = time.time()
# kvStore.push(0,testGrad,priority=0)
# kvStore.pull(0,testParam,priority=0)
# testParam.wait_to_read()
for paramIndex in range(len(qnet.params)):#range(6):#
k=qnet.params.keys()[paramIndex]
kvStore.push(paramIndex,qnet.params_grad[k],priority=-paramIndex)
kvStore.pull(paramIndex,qnet.params[k],priority=-paramIndex)
for v in qnet.params.values():
v.wait_to_read()
logging.info("wait time %f" %(time.time()-time_before_wait))
for epoch in xrange(epoch_num):
# Run Epoch
steps_left = steps_per_epoch
episode = 0
epoch_reward = 0
start = time.time()
game.start()
while steps_left > 0:
# Running New Episode
episode += 1
episode_loss = 0.0
episode_q_value = 0.0
episode_update_step = 0
episode_action_step = 0
time_episode_start = time.time()
game.begin_episode(steps_left)
while not game.episode_terminate:
# 1. We need to choose a new action based on the current game status
if game.state_enabled and game.replay_memory.sample_enabled:
do_exploration = (npy_rng.rand() < eps_curr)
eps_curr = max(eps_curr - eps_decay, eps_min)
if do_exploration:
action = npy_rng.randint(action_num)
else:
# TODO Here we can in fact play multiple gaming instances simultaneously and make actions for each
# We can simply stack the current_state() of gaming instances and give prediction for all of them
# We need to wait after calling calc_score(.), which makes the program slow
# TODO Profiling the speed of this part!
current_state = game.current_state()
state = nd.array(current_state.reshape((1,) + current_state.shape),
ctx=q_ctx) / float(255.0)
qval_npy = qnet.forward(batch_size=1, data=state)[0].asnumpy()
action = numpy.argmax(qval_npy)
episode_q_value += qval_npy[0, action]
episode_action_step += 1
else:
action = npy_rng.randint(action_num)
# 2. Play the game for a single mega-step (Inside the game, the action may be repeated for several times)
game.play(action)
total_steps += 1
# 3. Update our Q network if we can start sampling from the replay memory
# Also, we update every `update_interval`
if total_steps % update_interval == 0 and game.replay_memory.sample_enabled:
# 3.1 Draw sample from the replay_memory
training_steps += 1
episode_update_step += 1
states, actions, rewards, next_states, terminate_flags \
= game.replay_memory.sample(batch_size=minibatch_size)
states = nd.array(states, ctx=q_ctx) / float(255.0)
next_states = nd.array(next_states, ctx=q_ctx) / float(255.0)
actions = nd.array(actions, ctx=q_ctx)
rewards = nd.array(rewards, ctx=q_ctx)
terminate_flags = nd.array(terminate_flags, ctx=q_ctx)
# 3.2 Use the target network to compute the scores and
# get the corresponding target rewards
if not args.double_q:
target_qval = target_qnet.forward(batch_size=minibatch_size,
data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(target_qval))\
* (1.0 - terminate_flags) * discount
else:
target_qval = target_qnet.forward(batch_size=minibatch_size,
data=next_states)[0]
qval = qnet.forward(batch_size=minibatch_size, data=next_states)[0]
target_rewards = rewards + nd.choose_element_0index(target_qval,
nd.argmax_channel(qval))\
* (1.0 - terminate_flags) * discount
outputs = qnet.forward(batch_size=minibatch_size,is_train=True, data=states,
dqn_action=actions,
dqn_reward=target_rewards)
qnet.backward(batch_size=minibatch_size)
nd.waitall()
time_before_update = time.time()
if args.kv_type != None:
if total_steps % kvstore_update_period == 0:
update_to_kvstore(kvStore,qnet.params,qnet.params_grad)
else:
qnet.update(updater=updater)
logging.info("update time %f" %(time.time()-time_before_update))
time_before_wait = time.time()
nd.waitall()
logging.info("wait time %f" %(time.time()-time_before_wait))
'''nd.waitall()
time_before_wait = time.time()
kvStore.push(0,testGrad,priority=0)
kvStore.pull(0,testParam,priority=0)
nd.waitall()
logging.info("wait time %f" %(time.time()-time_before_wait))'''
# 3.3 Calculate Loss
diff = nd.abs(nd.choose_element_0index(outputs[0], actions) - target_rewards)
quadratic_part = nd.clip(diff, -1, 1)
loss = (0.5 * nd.sum(nd.square(quadratic_part)) + nd.sum(diff - quadratic_part)).asscalar()
episode_loss += loss
# 3.3 Update the target network every freeze_interval
# (We can do annealing instead of hard copy)
if training_steps % freeze_interval == 0:
qnet.copy_params_to(target_qnet)
steps_left -= game.episode_step
time_episode_end = time.time()
# Update the statistics
epoch_reward += game.episode_reward
info_str = "Epoch:%d, Episode:%d, Steps Left:%d/%d, Reward:%f, fps:%f, Exploration:%f" \
% (epoch, episode, steps_left, steps_per_epoch, game.episode_reward,
game.episode_step / (time_episode_end - time_episode_start), eps_curr)
if episode_update_step > 0:
info_str += ", Avg Loss:%f/%d" % (episode_loss / episode_update_step,
episode_update_step)
if episode_action_step > 0:
info_str += ", Avg Q Value:%f/%d" % (episode_q_value / episode_action_step,
episode_action_step)
logging.info(info_str)
end = time.time()
fps = steps_per_epoch / (end - start)
qnet.save_params(dir_path=args.dir_path, epoch=epoch)
logging.info("Epoch:%d, FPS:%f, Avg Reward: %f/%d"
% (epoch, fps, epoch_reward / float(episode), episode))
advantages = np.concatenate([p['advantages'] for p in paths])
cur_batch_size = observations.shape[0]
outputs = net.forward(
is_train=True,
data=observations,
var=1. * np.ones((cur_batch_size, 1)),
)
policy_actions = outputs[0].asnumpy()
critics = outputs[3].asnumpy()
variance = outputs[2].asnumpy()
action_mean = outputs[1].asnumpy()
net.backward(
policy_score=advantages,
policy_backward_action=actions,
critic_label=q_estimations.reshape(q_estimations.size, ),
)
for grad in net.params_grad.values():
grad[:] = grad[:] / cur_batch_size
if args.clip_gradient:
norm_clipping(net.params_grad, 10)
net.update(updater)
print 'Epoch:%d, Average Return:%f, Max Return:%f, Min Return:%f, Num Traj:%d\n, Mean:%f, Var:%f, Average Baseline:%f' \
%(itr, np.mean([sum(p["rewards"]) for p in paths]),
np.max([sum(p["rewards"]) for p in paths]),
np.min([sum(p["rewards"]) for p in paths]),
N, action_mean.mean(), variance.mean(), critics.mean()
)
if args.save_model:
net.save_params(dir_path='./', epoch=itr)
