def create(cls):
if not hasattr(cls, 'length_to_eps'):
# Maps episode length to list of episodes
cls.length_to_eps = {}
if not hasattr(cls, 'ep_indices'):
# Set of episode indices already in the cache
cls.ep_indices = set()
if not hasattr(cls, 'batches'):
# List of batches if popping batches
cls.batches = []
if not hasattr(cls, 'load_complete'):
# If all episodes have been loaded into memory
cls.load_complete = Value(ctypes.c_bool, False)
if not hasattr(cls, 'batches_lock'):
# Lock to access batches
cls.batches_lock = Lock()
if not hasattr(cls, 'cache_lock'):
# Lock to access length_to_eps
cls.cache_lock = Lock()
if not hasattr(cls, 'fill_cache_lock'):
# Lock for condition variables
cls.fill_cache_lock = RLock()
if not hasattr(cls, 'add_to_cache_cv'):
# Condition notifying Loader to add to cache
cls.add_to_cache_cv = Condition(lock=cls.fill_cache_lock)
if not hasattr(cls, 'cache_filled_cv'):
# Condition notifying teacher that cache has episodes
cls.cache_filled_cv = Condition(lock=cls.fill_cache_lock)
def __init__(self,
keys: List[Union[int, str]],
zero_obs: dict,
device: torch.device,
out_queue: Queue,
recorder: Recorder,
trajectory_length: int = 128):
# ATTRIBUTES
self.out_queue = out_queue
self.device = device
self.recorder = recorder
self._trajectory_length = trajectory_length
self.zero_obs = {k: v.to(self.device) for k, v in zero_obs.items()}
self.zero_obs['episode_id'] = torch.ones(
(1, 1), device=self.device) * -1
self.zero_obs['prev_episode_id'] = torch.ones(
(1, 1), device=self.device) * -1
self._reset_states(self.zero_obs)
# Counters
self._trajectory_counter = 0
self._episode_counter = len(keys)
# Setup storage
self._internal_store = {k: self._new_trajectory() for k in keys}
self._episode_id_store = {k: i for i, k in enumerate(keys)}
self._locks_trajectories = {k: Lock() for k in keys}
self._lock_episode_counter = Lock()
def __init__(self, config: ParamDict, environment: Environment,
policy: Policy, filter_op: Filter):
threads, gpu = config.require("threads", "gpu")
threads_gpu = config["gpu threads"] if "gpu threads" in config else 2
super(Agent_async, self).__init__(config, environment, policy,
filter_op)
# sync signal, -1: terminate, 0: normal running, >0 restart and waiting for parameter update
self._sync_signal = Value('i', 0)
# environment sub-process list
self._environment_proc = []
# policy sub-process list
self._policy_proc = []
# used for synchronize policy parameters
self._param_pipe = None
self._policy_lock = Lock()
# used for synchronize roll-out commands
self._control_pipe = None
self._environment_lock = Lock()
step_pipe = []
cmd_pipe_child, cmd_pipe_parent = Pipe(duplex=True)
param_pipe_child, param_pipe_parent = Pipe(duplex=False)
self._control_pipe = cmd_pipe_parent
self._param_pipe = param_pipe_parent
for i_envs in range(threads):
child_name = f"environment_{i_envs}"
step_pipe_pi, step_pipe_env = Pipe(duplex=True)
step_lock = Lock()
worker_cfg = ParamDict({
"seed": self.seed + 1024 + i_envs,
"gpu": gpu
})
child = Process(target=Agent_async._environment_worker,
name=child_name,
args=(worker_cfg, cmd_pipe_child, step_pipe_env,
self._environment_lock, step_lock,
self._sync_signal, deepcopy(environment),
deepcopy(filter_op)))
self._environment_proc.append(child)
step_pipe.append((step_pipe_pi, step_lock))
child.start()
for i_policies in range(threads_gpu):
child_name = f"policy_{i_policies}"
worker_cfg = ParamDict({
"seed": self.seed + 2048 + i_policies,
"gpu": gpu
})
child = Process(target=Agent_async._policy_worker,
name=child_name,
args=(worker_cfg, param_pipe_child, step_pipe,
self._policy_lock, self._sync_signal,
deepcopy(policy)))
self._policy_proc.append(child)
child.start()
sleep(5)
def __init__(self,
config,
share_batches=True,
manager=None,
new_process=True):
if new_process == True and manager is None:
manager = Manager()
self.knows = Semaphore(0) # > 0 if we know if any are coming
# == 0 if DatasetReader is processing a command
self.working = Semaphore(1 if new_process else 100)
self.finished_reading = Lock(
) # locked if we're still reading from file
# number of molecules that have been sent to the pipe:
self.in_pipe = Value('i', 0)
# Tracking what's already been sent through the pipe:
self._example_number = Value('i', 0)
# The final kill switch:
self._close = Value('i', 0)
self.command_queue = manager.Queue(10)
self.molecule_pipeline = None
self.batch_queue = Queue(config.data.batch_queue_cap
) #manager.Queue(config.data.batch_queue_cap)
self.share_batches = share_batches
self.dataset_reader = DatasetReader("dataset_reader",
self,
config,
new_process=new_process)
if new_process:
self.dataset_reader.start()
def __init__(self,
data_path,
sizes_filename,
seq_length,
initial_seed,
max_epochs=100):
# Input parameters.
self.data_path = data_path
self.sizes_filename = sizes_filename
self.seq_length = seq_length
self.initial_seed = initial_seed
self.max_epochs = max_epochs
# Lock for building the dataset.
self.lock = Lock()
# Shard stuff.
# Dictionary from shard nameto its size (number of element).
self.master_shard_size_dict = None
# Dictionary from shard name to modified size so it is
# divisible by self.seq_length.
self.shard_size_dict = None
# Long array (self.max_epochs * num-shards) populated
# randomly with shard names.
self.shards_name = None
# Start index of the data for a shard.
self.shards_start_index = None
self.build_shard_mappings_()
self.data_length = self.shards_start_index[-1]
# Data.
self.shards_data = [None] * self.shards_name.size
self.shards_sample_index = [None] * self.shards_name.size
def __init__(self, config: ParamDict, environment: Environment,
policy: Policy, filter_op: Filter):
threads, gpu = config.require("threads", "gpu")
super(Agent_sync, self).__init__(config, environment, policy,
filter_op)
# sync signal, -1: terminate, 0: normal running, >0 restart and waiting for parameter update
self._sync_signal = Value('i', 0)
# sampler sub-process list
self._sampler_proc = []
# used for synchronize commands
self._cmd_pipe = None
self._param_pipe = None
self._cmd_lock = Lock()
cmd_pipe_child, cmd_pipe_parent = Pipe(duplex=True)
param_pipe_child, param_pipe_parent = Pipe(duplex=False)
self._cmd_pipe = cmd_pipe_parent
self._param_pipe = param_pipe_parent
for i_thread in range(threads):
child_name = f"sampler_{i_thread}"
worker_cfg = ParamDict({
"seed": self.seed + 1024 + i_thread,
"gpu": gpu
})
child = Process(target=Agent_sync._sampler_worker,
name=child_name,
args=(worker_cfg, cmd_pipe_child, param_pipe_child,
self._cmd_lock, self._sync_signal,
deepcopy(policy), deepcopy(environment),
deepcopy(filter_op)))
self._sampler_proc.append(child)
child.start()
def per_step(valLoader,
model,
criterion,
downsamplingFactor):
model.eval()
criterion.eval()
avgPER = 0
varPER = 0
nItems = 0
print("Starting the PER computation through beam search")
bar = progressbar.ProgressBar(maxval=len(valLoader))
bar.start()
for index, data in enumerate(valLoader):
bar.update(index)
with torch.no_grad():
seq, sizeSeq, phone, sizePhone = prepare_data(data)
c_feature = model(seq)
sizeSeq = sizeSeq / downsamplingFactor
predictions = torch.nn.functional.softmax(criterion.getPrediction(c_feature),
dim=2).cpu()
c_feature = c_feature
phone = phone.cpu()
sizeSeq = sizeSeq.cpu()
sizePhone = sizePhone.cpu()
mutex = Lock()
manager = Manager()
poolData = manager.list()
processes = []
for b in range(sizeSeq.size(0)):
l_ = min(sizeSeq[b] // 4, predictions.size(1))
s_ = sizePhone[b]
p = torch.multiprocessing.Process(target=get_local_per,
args=(poolData, mutex, predictions[b, :l_].view(l_, -1).numpy(),
phone[b, :s_].view(-1).numpy().astype(np.int32), criterion.BLANK_LABEL))
p.start()
processes.append(p)
for p in processes:
p.join()
avgPER += sum([x for x in poolData])
varPER += sum([x*x for x in poolData])
nItems += len(poolData)
bar.finish()
avgPER /= nItems
varPER /= nItems
varPER -= avgPER**2
print(f"Average PER {avgPER}")
print(f"Standard deviation PER {math.sqrt(varPER)}")
def init_data(self):
self.is_working = False
self.semaphore = True
self.is_change_bar = Value(
c_bool, False) #whether user has dragged the slider,default: False
self.frame_index = Value('i', 0)
self.share_lock = Lock() #shared lock for frame_index
self.share_lock2 = Lock() # shared lock for frame_index
self.mutex = threading.Lock()
self.timer = QTimer(self) # used for the updating of progress bar
self.temp_timer = QTimer(
self) #used for detecting whether the frame_total is given.
self.frame_total = Value('i', -1)
self.playable = Value(c_bool, True)
self.is_working = Value(c_bool, False)
manager = Manager()
self.play_src = manager.Value(c_char_p, '0') #用于记录播放的视频地址
self.mode = None # 'online' or 'offline'
def __init__(self,
chk_dir,
chk,
keep_epoch_chk=True,
overwrite=True,
mode=CFMode.AUTO,
chk_prefix='model_v_'):
self.logger = logging.getLogger(__name__)
self.chk_dir = chk_dir
self.chk = chk
self.keep_epoch_chk = keep_epoch_chk
self.overwrite = overwrite
self.chk_prefix = chk_prefix
self.mode = mode
self.chk_epoch_subdir = 'epoch'
self.mp_manager = Manager()
self.snapshot_copy = None
self.cpu_side = False
# Active snapshot, if true, don't snapshot again
self.active_snapshot = Value('i', 0)
self.lock = Lock()
self.in_progress_snapshot = Value('i', 0)
# Handle to the process performing checkpoint
# Can be only one at any instant. A new checkpoint
# cannot start unless the previous one completes
self.chk_process = None
# `overwrite` supersedes if False
if self.overwrite is False and self.keep_epoch_chk is False:
self.keep_epoch_chk = True
# Global ID of checkpoints being written
# Used to format the checkpoint path
# Instantiate from chk when restoring
self.chk_global_id = -1
# Sorted List of available checkpoints (fnames)
self.available_chk_iters = self.mp_manager.list()
self.available_chk_epochs = self.mp_manager.list()
self.initalize_chk_dir()
self.logger.info("Available checkpoints : ")
for item in self.available_chk_iters:
self.logger.info(item)
def __init__(self, log_directory: str, hyperparams: HyperParams):
self.lock = Lock()
self.episode_lengths = []
self.episode_scores = []
self.episode_values = []
self.frame_counter = None
self.queue = Queue(100)
self.log_directory = log_directory
if not os.path.exists(log_directory + '/hyperparams.txt'):
with open(log_directory + '/hyperparams.txt', 'w') as param_file:
for parameter in hyperparams._fields:
value = getattr(hyperparams, parameter)
print(f'{parameter}: {value}', file=param_file)
def __init__(self, path, data_type='data', mem_map=False, map_fn=None):
lazypath = get_lazy_path(path)
datapath = os.path.join(lazypath, data_type)
#get file where array entries are concatenated into one big string
self._file = open(datapath, 'rb')
self.file = self._file
#memory map file if necessary
self.mem_map = mem_map
if self.mem_map:
self.file = mmap.mmap(self.file.fileno(), 0, prot=mmap.PROT_READ)
lenpath = os.path.join(lazypath, data_type+'.len.pkl')
self.lens = pkl.load(open(lenpath, 'rb'))
self.ends = list(accumulate(self.lens))
self.dumb_ends = list(self.ends)
self.read_lock = Lock()
self.process_fn = map_fn
self.map_fn = map_fn
self._tokenizer = None
def __init__(self,
path,
data_type='data',
mem_map=False,
map_fn=None,
is_array=False,
array_data_type=np.int32):
lazypath = get_lazy_path(path)
datapath = os.path.join(lazypath, data_type)
#get file where array entries are concatenated into one big string
self._file = open(datapath, 'rb')
self.file = self._file
self.is_array = is_array
self.array_data_type = array_data_type
#memory map file if necessary
lenpath = os.path.join(lazypath, data_type + '.len.pkl')
self.lens = pkl.load(open(lenpath, 'rb'))
self.ends = list(accumulate(self.lens))
self.dumb_ends = list(self.ends)
self.mem_map = mem_map
if self.mem_map:
if is_array:
if self.ends[-1] == 0:
self.file = np.array([], dtype=array_data_type)
else:
self.file = np.memmap(self.file,
dtype=array_data_type,
mode='r',
order='C')
else:
if self.ends[-1] == 0:
self.file = bytearray()
else:
self.file = mmap.mmap(self.file.fileno(),
0,
prot=mmap.PROT_READ)
self.read_lock = Lock()
self.process_fn = map_fn
self.map_fn = map_fn
self._tokenizer = None
self.is_lazy = True
def run_sac(env,
obs_state,
num_actions,
hyperps,
device=torch.device("cpu"),
render=True):
mp.set_start_method('spawn')
print(torch.multiprocessing.get_start_method())
#import pudb; pudb.set_trace()
shared_model = ActorSimpleMem(412, 2).share_memory().to(device)
shared_optimizer = SharedAdam(shared_model.parameters(),
lr=hyperps['q_lr'])
lock = Lock()
shared_buffer = BasicBuffer(hyperps['maxmem'])
processes = []
summary_model = ActorSimpleMem(412, 2)
summary_model.load_state_dict(shared_model.state_dict())
writer.add_graph(summary_model,
[(torch.randn(64, 412), torch.randn(64, 256))])
print('Spwaning training processes')
for rank in range(3):
p = mp.Process(target=train_on_env,
args=(env, rank, lock, shared_model, shared_optimizer,
obs_state, num_actions, hyperps, shared_buffer,
device, render))
time.sleep(10)
p.start()
processes.append(p)
for p in processes:
p.join()
#train_on_env(env, rank, lock, shared_model, shared_optimizer, obs_state, num_actions, hyperps, shared_buffer, device, render)
writer.close()
def main():
args = parse_args()
categories = parse_categories(parse_data(args.data)['names'])
cap = cv2.VideoCapture(0)
frame_queue = Queue()
preds_queue = Queue()
cur_dets = None
frame_lock = Lock()
proc = Process(target=detect,
args=(frame_queue, preds_queue, frame_lock, args))
proc.start()
try:
while (True):
ret, frame = cap.read()
frame_lock.acquire()
while not frame_queue.empty():
frame_queue.get()
frame_queue.put(frame)
frame_lock.release()
if not preds_queue.empty():
cur_dets = preds_queue.get()
if cur_dets is not None and len(cur_dets) > 0:
frame = draw_detections_opencv(frame, cur_dets[0], categories)
cv2.imshow('frame', frame)
cv2.waitKey(1)
except KeyboardInterrupt:
print('Interrupted')
proc.join()
cap.release()
cv2.destroyAllWindows()
def __init__(self, bot_id, cfg, task_factory, encoder, Actor, Critic, goal_encoder):
self.cfg = cfg
self.bot = Bot(
cfg,
bot_id,
encoder,
goal_encoder,
Actor,
Critic,
task_factory.state_size,
task_factory.action_size,
task_factory.wrap_action,
task_factory.wrap_value)
self.bot.share_memory() # !! must be done from main process !!
self.iter = 0
self.freezed = 0
self.counter = 1
self.tau = LinearSchedule(cfg['tau_replay_counter'], cfg['tau_base'], cfg['tau_final'])
self.lock = Lock()
self.bot = BotProxy(self.lock, cfg, self.bot, cfg['device'])
def __init__(self,
cache_dir,
dataset_dir,
dataset_list,
cuda,
batch_size=500,
num_workers=3,
renew_frequency=5,
rejection_radius_position=0,
numpatches=900,
numneg=3,
pos_thr=50.0,
reject=True,
mode='train',
rejection_radius=3000,
dist_type='3D',
patch_radius=None,
use_depth=False,
use_normals=False,
use_silhouettes=False,
color_jitter=False,
greyscale=False,
maxres=4096,
scale_jitter=False,
photo_jitter=False,
uniform_negatives=False,
needles=0,
render_only=False,
maxitems=200,
cache_once=False):
super(MultimodalPatchesCache, self).__init__()
self.cache_dir = cache_dir
self.dataset_dir = dataset_dir
#self.images_path = images_path
self.dataset_list = dataset_list
self.cuda = cuda
self.batch_size = batch_size
self.num_workers = num_workers
self.renew_frequency = renew_frequency
self.rejection_radius_position = rejection_radius_position
self.numpatches = numpatches
self.numneg = numneg
self.pos_thr = pos_thr
self.reject = reject
self.mode = mode
self.rejection_radius = rejection_radius
self.dist_type = dist_type
self.patch_radius = patch_radius
self.use_depth = use_depth
self.use_normals = use_normals
self.use_silhouettes = use_silhouettes
self.color_jitter = color_jitter
self.greyscale = greyscale
self.maxres = maxres
self.scale_jitter = scale_jitter
self.photo_jitter = photo_jitter
self.uniform_negatives = uniform_negatives
self.needles = needles
self.render_only = render_only
self.cache_done_lock = Lock()
self.all_done = Value('B', 0) # 0 is False
self.cache_done = Value('B', 0) # 0 is False
self.wait_for_cache_builder = Event()
# prepare for wait until initial cache is built
self.wait_for_cache_builder.clear()
self.cache_builder_resume = Event()
self.maxitems = maxitems
self.cache_once = cache_once
if self.mode == 'eval':
self.maxitems = -1
self.cache_builder = Process(target=self.buildCache,
args=[self.maxitems])
self.current_cache_build = Value('B', 0) # 0th cache
self.current_cache_use = Value('B', 1) # 1th cache
self.cache_names = ["cache1", "cache2"] # constant
rebuild_cache = True
if self.mode == 'eval':
validation_dir = os.path.join(
self.cache_dir,
self.cache_names[self.current_cache_build.value])
if os.path.isdir(validation_dir):
# we don't need to rebuild validation cache
# TODO: check if cache is VALID
rebuild_cache = False
elif cache_once:
build_dataset_dir = os.path.join(
self.cache_dir,
self.cache_names[self.current_cache_build.value])
if os.path.isdir(build_dataset_dir):
# we don't need to rebuild training cache if we are training
# on limited subset of the training set
rebuild_cache = False
if rebuild_cache:
# clear the caches if they already exist
build_dataset_dir = os.path.join(
self.cache_dir,
self.cache_names[self.current_cache_build.value])
if os.path.isdir(build_dataset_dir):
shutil.rmtree(build_dataset_dir)
use_dataset_dir = os.path.join(
self.cache_dir, self.cache_names[self.current_cache_use.value])
if os.path.isdir(use_dataset_dir):
shutil.rmtree(use_dataset_dir)
os.makedirs(build_dataset_dir)
self.cache_builder_resume.set()
self.cache_builder.start()
# wait until initial cache is built
# print("before wait to build")
# print("wait for cache builder state",
# self.wait_for_cache_builder.is_set())
self.wait_for_cache_builder.wait()
# print("after wait to build")
# we have been resumed
if self.mode != 'eval' and (not self.cache_once):
# for training, we can set up the cache builder to build
# the second cache
self.restart()
else:
# else for validation we don't need second cache
# we just need to switch the built cache to the use cache in order
# to use it
tmp = self.current_cache_build.value
self.current_cache_build.value = self.current_cache_use.value
self.current_cache_use.value = tmp
else:
playouts = []
# Load the replay memory
replay_memory = ReplayMemory(16384)
for playout in playouts:
for move_datapoint in playout:
replay_memory.push(move_datapoint)
# Setup the network
board_size = len(playouts[0][0].board_list)
net = Resnet(2, board_size).to(DEVICE)
if args.init_model is not None:
net.Load(args.init_model)
mem_lock = Lock()
param_queue = None
server = None
shutdown_val = None
receiver_proc = None
if args.server_config is not None:
# Setup the handling of workers and the parameter server
with open(args.server_config) as server_file:
config = json.load(server_file)
if ('addr' not in config) or ('port' not in config):
print("IP address (addr) and port number required in config")
address = config['addr']
port = int(config['port'])
entropy_layout = dict(title="Entropies",
xaxis={'title': 'n-step iter'},
yaxis={'title': 'entropy'})
MAX_EPISODES = 2000
DISCOUNT_FACTOR = 0.99
STEPS = 10
GlobalModel = Model()
GlobalModel.share_memory()
CriticOptimizer = torch.optim.Adam(GlobalModel.CriticParameters(), lr=0.01)
ActorOptimizer = torch.optim.Adam(GlobalModel.ActorParameters(), lr=0.001)
Optimizer = torch.optim.Adam(GlobalModel.parameters(), lr=0.001)
lock = Lock()
num_cpu = 4
agents = []
for cpu in range(num_cpu):
agents.append(Agent(cpu))
receiver, sender = Pipe()
agent_threads = []
for agent in agents:
thread = Process(target=agent.letsgo,
args=(
GlobalModel,
CriticOptimizer,
ActorOptimizer,
'''
Maps episode length to dictionary with following keys:
current_idx: which episode in the list are we at (if simply indexing
into list)
ep_list: list of episodes of the length of the key
bucket_complete: if there are no more episodes left to consider in
the bucket
'''
# Maps episode length to list of episodes
length_to_eps = {}
# List of batches if popping batches
batches = []
# If all episodes have been loaded into memory
load_complete = Value(ctypes.c_bool, False)
# Lock to access batches
batches_lock = Lock()
# Lock to access length_to_eps
cache_lock = Lock()
# Lock for condition variables
fill_cache_lock = RLock()
# Condition notifying Loader to add to cache
add_to_cache_cv = Condition(lock=fill_cache_lock)
# Condition notifying teacher that cache has episodes
cache_filled_cv = Condition(lock=fill_cache_lock)
def batch_cache(function):
max_cache_size = 10000 # Max unseen eps
min_cache_size = 1000 # Min unseen eps
def get_cache_size():
'''
Maps episode length to dictionary with following keys:
current_idx: which episode in the list are we at (if simply indexing
into list)
ep_list: list of episodes of the length of the key
bucket_complete: if there are no more episodes left to consider in
the bucket
'''
length_to_eps = {} # Maps episode length to list
# of episodes
batches = [] # List of batches if popping
# batches
load_complete = Value(ctypes.c_bool, False) # If all episodes have been
# loaded into memory
batches_lock = Lock() # Lock to access batches
cache_lock = Lock() # Lock to access length_to_eps
fill_cache_lock = RLock() # Lock for condition variables
add_to_cache_cv = Condition(lock=fill_cache_lock) # Condition notifying Loader
# to add to cache
cache_filled_cv = Condition(lock=fill_cache_lock) # Condition notifying teacher
# that cache has episodes
def batch_cache(function):
max_cache_size = 10000 # Max unseen eps
min_cache_size = 1000 # Min unseen eps
def get_cache_size():
'''Returns number of available episodes '''
return sum(len(v['ep_list']) - v['current_idx']for k, v in length_to_eps.items())
def crop_face(args):
for k, v in default_args.items():
setattr(args, k, v)
assert osp.exists(args.data_dir), "The input dir not exist"
root_folder_name = args.data_dir.split('/')[-1]
src_folder = args.data_dir
dst_folder = args.data_dir.replace(root_folder_name, root_folder_name + '_OPPOFaces')
lz.mkdir_p(dst_folder, delete=False)
ds = TestData(src_folder)
loader = torch.utils.data.DataLoader(ds, batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False,
pin_memory=True,
drop_last=False
)
# 1. load pre-tained model
checkpoint_fp = 'models/phase1_wpdc_vdc.pth.tar'
arch = 'mobilenet_1'
checkpoint = torch.load(checkpoint_fp, map_location=lambda storage, loc: storage)['state_dict']
model = getattr(mobilenet_v1, arch)(num_classes=62) # 62 = 12(pose) + 40(shape) +10(expression)
model_dict = model.state_dict()
# because the model is trained by multiple gpus, prefix module should be removed
for k in checkpoint.keys():
model_dict[k.replace('module.', '')] = checkpoint[k]
model.load_state_dict(model_dict)
if args.mode == 'gpu':
cudnn.benchmark = True
model = model.cuda()
model.eval()
# 2. load dlib model for face detection and landmark used for face cropping
queue = Queue()
lock = Lock()
consumers = []
for i in range(args.num_consumers):
p = Process(target=consumer, args=(queue, lock))
p.daemon = True
consumers.append(p)
for c in consumers:
c.start()
# 3. forward
ttl_nimgs = 0
ttl_imgs = []
data_meter = lz.AverageMeter()
model_meter = lz.AverageMeter()
post_meter = lz.AverageMeter()
lz.timer.since_last_check('start crop face')
for ind, data in enumerate(loader):
data_meter.update(lz.timer.since_last_check(verbose=False))
if (data['finish'] == 1).all().item():
logging.info('finish')
break
if ind % 10 == 0:
logging.info(
f'proc batch {ind}, data time: {data_meter.avg:.2f}, model: {model_meter.avg:.2f}, post: {post_meter.avg:.2f}')
mask = data['finish'] == 0
input = data['img'][mask]
input_np = input.numpy()
roi_box = data['roi_box'][mask].numpy()
imgfn = np.asarray(data['imgfn'])[mask.numpy().astype(bool)]
dst_imgfn = [img_fp.replace(root_folder_name, root_folder_name + '_OPPOFaces') for img_fp in imgfn]
ttl_imgs.extend(dst_imgfn)
ttl_nimgs += mask.sum().item()
with torch.no_grad():
if args.mode == 'gpu':
input = input.cuda()
param = model(input)
param = param.squeeze().cpu().numpy().astype(np.float32)
model_meter.update(lz.timer.since_last_check(verbose=False))
queue.put((imgfn, param, roi_box, dst_imgfn))
# pts68 = [predict_68pts(param[i], roi_box[i]) for i in range(param.shape[0])]
# pts68_proc = [predict_68pts(param[i], [0, 0, STD_SIZE, STD_SIZE]) for i in range(param.shape[0])]
# for img_fp, pts68_, pts68_proc_, img_, dst in zip(imgfn, pts68, pts68_proc, input_np, dst_imgfn):
# ## this may need opt to async read write
# img_ori = cvb.read_img(img_fp)
# pts5 = to_landmark5(pts68_[:2, :].transpose())
# warped = preprocess(img_ori, landmark=pts5)
# # plt_imshow(warped, inp_mode='bgr'); plt.show()
# lz.mkdir_p(osp.dirname(dst), delete=False)
# cvb.write_img(warped, dst)
#
# ## this may cause black margin
# # pts5 = to_landmark5(pts68_proc_[:2, :].transpose())
# # warped = preprocess(to_img(img_), landmark=pts5)
# # # plt_imshow(warped, inp_mode='bgr'); plt.show()
# # dst = img_fp.replace(root_folder_name, root_folder_name + '_OPPOFaces')
# # cvb.write_img(warped, dst)
# if args.dump_res:
# img_ori = cvb.read_img(img_fp)
# pts_res = [pts68_]
# dst = img_fp.replace(root_folder_name, root_folder_name + '_kpts.demo')
# lz.mkdir_p(osp.dirname(dst), delete=False)
# draw_landmarks(img_ori, pts_res,
# wfp=dst,
# show_flg=args.show_flg)
post_meter.update(lz.timer.since_last_check(verbose=False))
lz.msgpack_dump(ttl_imgs, dst_folder + '/' + 'all_imgs.pk')
del model, input
torch.cuda.empty_cache()
while not queue.empty():
time.sleep(1)
def sim_games(N_games, N_MCTS, model, number_of_processes, v_resign, model2 = None, duel=False, batch_size = 8, board_size = 9):
#### Function for generating games
print("Starting sim games")
process_workers = []
torch.multiprocessing.set_start_method('spawn', force=True)
# Make queues for sending data
gpu_Q = Queue()
if (duel==False):
data_Q = Queue()
# Also make pipe for receiving v_resign
conn_rec, conn_send = Pipe(False)
p_data = Process(target=data_handler, args=(data_Q, N_games, conn_send))
process_workers.append(p_data)
else:
winner_Q = Queue()
gpu_Q2 = Queue()
process_workers.append(Process(target=gpu_worker, args=(gpu_Q2, batch_size, board_size, model2)))
# Make counter and lock
game_counter = Value('i', 0)
lock = Lock()
# Make process for gpu worker and data_loader
process_workers.append(Process(target=gpu_worker, args=(gpu_Q, batch_size, board_size, model)))
# Start gpu and data_loader worker
print("GPU processes")
for p in process_workers:
p.start()
# Construct tasks for workers
procs = []
torch.multiprocessing.set_start_method('fork', force=True)
print("defining worker processes")
for i in range(number_of_processes):
seed = np.random.randint(int(2**31))
if (duel==True):
procs.append(Process(target=sim_duel_game_worker, args=(gpu_Q, gpu_Q2, N_MCTS, winner_Q, N_games, lock, game_counter, seed)))
else:
procs.append(Process(target=sim_game_worker, args=(gpu_Q, N_MCTS, data_Q, v_resign, N_games, lock, game_counter, seed)))
print("Starting worker processes")
# Begin running games
for p in procs:
p.start()
# Join processes
if (duel==False):
# Receive new v_resign
v_resign = conn_rec.recv()
else:
player1_wins = 0
player2_wins = 0
for i in range(N_games):
player1_won = winner_Q.get(True)
if (player1_won==1):
player1_wins += 1
else:
player2_wins += 1
for p in procs:
p.join()
# Close processes
for p in process_workers:
p.terminate()
# Returns v_resign if training else winrate when dueling
if (duel==False):
return v_resign
else:
return player1_wins, player2_wins
def __init__(self, val=0):
self.val = Value('i', val)
self.lock = Lock()
def train():
np.random.seed(random_seed)
torch.manual_seed(random_seed)
writer = SummaryWriter()
s2 = S2(latent_num, cnn_chanel_num, stat_dim).to(device).share_memory()
writer.add_graph(
s2, (torch.zeros([1, 1, img_shape[0], img_shape[1]]).to(device),
torch.zeros([1, stat_dim]).to(device)))
optim = GlobalAdam([{
'params': s2.encode_img.parameters()
}, {
'params': s2.encode_stat.parameters()
}, {
'params': s2.pi.parameters()
}, {
'params': s2.actor.parameters()
}],
lr=1e-2,
weight_decay=0.01)
if os.path.exists('S2_state_dict.pt'):
s2.load_state_dict(torch.load('S2_state_dict.pt'))
optim.load_state_dict(torch.load('S2_Optim_state_dict.pt'))
pair_queue = Queue(10000)
validate_queue = Queue()
optimizer_lock = Lock()
process = []
data_list = [
'A8888.XDCE', 'AL8888.XSGE', 'AU8888.XSGE', 'C8888.XDCE', 'M8888.XDCE',
'RU8888.XSGE', 'SR8888.XZCE'
]
for no in range(mp.cpu_count() - 1):
data = pd.read_csv(f"../data/{data_list[no]}_5m.csv")
worker = Worker_Generator(no, data, pair_queue)
worker.start()
process.append(worker)
validater = Validate(s2, optimizer_lock, validate_queue)
validater.start()
epochs = 0
while True:
imgs = []
stats = []
cates = []
seen = 0
while seen < minibatch:
img, stat, cate = pair_queue.get()
imgs.append(img)
stats.append(stat)
cates.append(cate)
seen += 1
imgs = torch.tensor(imgs).float().to(device)
stats = torch.tensor(stats).float().to(device)
g_t = torch.tensor(cates).long().to(device)
pred = s2(imgs, stats)
loss = F.cross_entropy(pred, g_t)
accr = (pred.argmax(1) == g_t).sum().item() / minibatch
with optimizer_lock:
optim.zero_grad()
loss.backward()
optim.step()
if not validate_queue.empty():
val_reward, val_money, val_win = validate_queue.get()
writer.add_scalar('Validate/reward', val_reward, epochs)
writer.add_scalar('Validate/money', val_money, epochs)
writer.add_scalar('Validate/win_rate', val_win, epochs)
writer.add_scalar('Train/Loss', loss.item(), epochs)
writer.add_scalar('Train/Accr', accr, epochs)
epochs += 1
if epochs % save_every == 0:
torch.save(s2.state_dict(), 'S2_state_dict.pt')
torch.save(optim.state_dict(), 'S2_Optim_state_dict.pt')
for worker in process:
worker.join()
import fasttext
import functools
import numpy as np
import pytext.utils.cuda_utils as cuda_utils
from abc import abstractmethod
from collections import OrderedDict
from sentencepiece import SentencePieceProcessor
from typing import List
from pytorch_pretrained_bert import BertModel, BertTokenizer
from pytext.config.field_config import ConfigBase
from torch.multiprocessing import Lock
MODEL_DOWNLOAD_LOCK = Lock()
def run_model(model, inputs, layer):
layers, _ = model(inputs)
return layers[layer].cpu()
class EmbedderInterface:
__REGISTRY = dict()
class Config(ConfigBase):
max_pieces: int = -1
preproc_dir: str = "."
use_cuda_if_available: bool = True
embed_type: str = "BERTEmbed"
def get_best_model(
log_path: str,
mss: ModelSearchSet,
train_data: Dataset,
validn_data: Dataset,
test_data: Dataset,
get_dataset_kwargs: Dict[str, Any],
exp_dir: str = '../out',
devices_info: List[Tuple[int, int]] = [
(-1, 1)
], # (device_id, max_proc_cnt), -1 for cpu
show_hpsearch_stats: bool = True):
logger = Logger(log_path)
if LOG_DATASET_STATS:
logger.log(train_data.get_stats(title='Train Data'))
logger.log(validn_data.get_stats(title='Validation Data'))
logger.log(test_data.get_stats(title='Test Data'))
# Prepare process-specific device ids, hyperparams and datasets
proc_device_id: List[int]
nprocs: int
# Prep proc_device_id
proc_device_id = []
device_ids_helper = {}
for info in devices_info:
device_ids_helper[info[0]] = info[1]
rem = True
while rem:
rem = False
for device_id in device_ids_helper:
if device_ids_helper[device_id] > 0:
rem = True
proc_device_id.append(device_id)
device_ids_helper[device_id] -= 1
# Prep proc_hps
proc_hps = []
if mss.model_class not in CPU_MODELS:
total_proc_cnt = len(proc_device_id)
proc_hp_cnt = np.full(total_proc_cnt, len(mss.hps) // total_proc_cnt)
proc_hp_cnt[:len(mss.hps) % total_proc_cnt] += 1
agg = 0
for cnt in proc_hp_cnt:
if cnt > 0:
proc_hps.append(mss.hps[agg:agg + cnt])
agg += cnt
else:
proc_hps = [mss.hps]
nprocs = len(proc_hps)
proc_device_id = proc_device_id[:nprocs]
# Run the processes
l: 'multiprocessing.synchronize.Lock' = Lock()
cmlock: 'multiprocessing.synchronize.Lock' = Lock()
started_hps = Value('i', 0)
started_hps_lk: 'multiprocessing.synchronize.Lock' = Lock()
finished_hps = Value('i', 0)
total_hps = len(mss.hps)
start_time = time.time()
assert (train_data.name
== validn_data.name) and (validn_data.name == test_data.name) and (
test_data.name == get_dataset_kwargs['dataset_name'])
assert (train_data.shuffle_seed == validn_data.shuffle_seed) and (
validn_data.shuffle_seed
== test_data.shuffle_seed) and (test_data.shuffle_seed
== get_dataset_kwargs['shuffle_seed'])
model_search_summary_path = '{}/{}/{}-search-summary{}{}.csv'.format(
exp_dir, EXP_LOGS_DIR, get_dataset_kwargs['dataset_name'], DESC_SEP,
exp_dir.split(DESC_SEP)[-1])
assert not os.path.exists(model_search_summary_path)
logger.log(f'\nExperiment dir: {os.path.abspath(exp_dir)}\n')
logger.log(f'Hyperparam configs: {total_hps}\n')
logger.log(f'Number of processes: {nprocs}\n')
args = (l, cmlock, started_hps, started_hps_lk, finished_hps, total_hps,
start_time, mss.model_class, proc_hps, train_data, validn_data,
test_data, get_dataset_kwargs, exp_dir, proc_device_id, log_path,
model_search_summary_path, mss.use_lforb)
if mss.model_class not in CPU_MODELS:
spawn(get_best_model_aux, args=args, nprocs=nprocs)
else:
get_best_model_aux(0, *args)
if show_hpsearch_stats:
hpsearch_stats(model_search_summary_path,
get_dataset_kwargs['dataset_name'])
logger.log('\n==================================\n')
logger.log('Model search summary saved to: {}\n'.format(
os.path.abspath(model_search_summary_path)))
logger.log('==================================\n\n')
logger.close()
def train():
np.random.seed(random_seed)
torch.manual_seed(random_seed)
writer = SummaryWriter()
ac = AC(latent_num, cnn_chanel_num, stat_dim)
writer.add_graph(ac, (torch.zeros([1, 1, img_shape[0], img_shape[1]
]), torch.zeros([1, stat_dim])))
optim = GlobalAdam([{
'params': ac.encode_img.parameters(),
'lr': 2.5e-5
}, {
'params': ac.encode_stat.parameters(),
'lr': 2.5e-5
}, {
'params': ac.pi.parameters(),
'lr': 2.5e-5
}, {
'params': ac.actor.parameters(),
'lr': 2.5e-5
}, {
'params': ac.f.parameters()
}, {
'params': ac.V.parameters()
}],
lr=5e-3,
weight_decay=weight_decay)
if os.path.exists('S3_state_dict.pt'):
ac.load_state_dict(torch.load('S3_state_dict.pt'))
optim.load_state_dict(torch.load('S3_Optim_state_dict.pt'))
else:
ac.load_state_dict(torch.load('../stage2/S2_state_dict.pt'),
strict=False)
result_queue = Queue()
validate_queue = Queue()
gradient_queue = Queue()
loss_queue = Queue()
ep_cnt = Value('i', 0)
optimizer_lock = Lock()
processes = []
ac.share_memory()
optimizer_worker = Process(target=update_shared_model,
args=(gradient_queue, optimizer_lock, optim,
ac))
optimizer_worker.start()
for no in range(mp.cpu_count() - 3):
worker = Worker(no, ac, ep_cnt, optimizer_lock, result_queue,
gradient_queue, loss_queue)
worker.start()
processes.append(worker)
validater = Validate(ac, ep_cnt, optimizer_lock, validate_queue)
validater.start()
best_reward = 0
while True:
with ep_cnt.get_lock():
if not result_queue.empty():
ep_cnt.value += 1
reward, money, win_rate = result_queue.get()
objective_actor, loss_critic, loss_f = loss_queue.get()
writer.add_scalar('Interaction/Reward', reward, ep_cnt.value)
writer.add_scalar('Interaction/Money', money, ep_cnt.value)
writer.add_scalar('Interaction/win_rate', win_rate,
ep_cnt.value)
writer.add_scalar('Update/objective_actor', objective_actor,
ep_cnt.value)
writer.add_scalar('Update/loss_critic', loss_critic,
ep_cnt.value)
writer.add_scalar('Update/loss_f', loss_f, ep_cnt.value)
with optimizer_lock:
if reward > best_reward:
best_reward = reward
torch.save(ac.state_dict(), 'S3_BEST_state_dict.pt')
if ep_cnt.value % save_every == 0:
torch.save(ac.state_dict(), 'S3_state_dict.pt')
torch.save(optim.state_dict(),
'S3_Optim_state_dict.pt')
if not validate_queue.empty():
val_reward, val_money, val_win_rate = validate_queue.get()
writer.add_scalar('Validation/reward', val_reward,
ep_cnt.value)
writer.add_scalar('Validation/money', val_money, ep_cnt.value)
writer.add_scalar('Validation/win_rate', val_win_rate,
ep_cnt.value)
for worker in processes:
worker.join()
optimizer_worker.kill()
def __init__(self):
self.value = Value('i', 0)
self.lock = Lock()
self.local = 0