def dump_config(config):
cfg = copy.deepcopy(config)
alloc = cfg.get('eth', {}).get('block', {}).get('GENESIS_INITIAL_ALLOC', {})
if len(alloc) > 100:
log.info('omitting reporting of %d accounts in genesis' % len(alloc))
del cfg['eth']['block']['GENESIS_INITIAL_ALLOC']
konfig.dump_config(cfg)
def run(ctx, dev, nodial, fake):
"""Start the client ( --dev to stop on error)"""
config = ctx.obj['config']
if nodial:
# config['deactivated_services'].append(PeerManager.name)
# config['deactivated_services'].append(NodeDiscovery.name)
config['discovery']['bootstrap_nodes'] = []
config['discovery']['listen_port'] = 29873
config['p2p']['listen_port'] = 29873
config['p2p']['min_peers'] = 0
if fake:
from ethereum import blocks
blocks.GENESIS_DIFFICULTY = 1024
blocks.BLOCK_DIFF_FACTOR = 16
blocks.MIN_GAS_LIMIT = blocks.GENESIS_GAS_LIMIT / 2
# workaround for genesis.json hack
blocks.GENESIS_JSON["difficulty"] = blocks.int_to_hex(blocks.GENESIS_DIFFICULTY)
# create app
app = EthApp(config)
# development mode
if dev:
gevent.get_hub().SYSTEM_ERROR = BaseException
try:
config['client_version'] += '/' + os.getlogin()
except:
log.warn("can't get and add login name to client_version")
pass
# dump config
konfig.dump_config(config)
# register services
for service in services:
assert issubclass(service, BaseService)
if service.name not in app.config['deactivated_services']:
assert service.name not in app.services
service.register_with_app(app)
assert hasattr(app.services, service.name)
unlock_accounts(ctx.obj['unlock'], app.services.accounts, password=ctx.obj['password'])
# start app
log.info('starting')
app.start()
if config['post_app_start_callback'] is not None:
config['post_app_start_callback'](app)
# wait for interrupt
evt = Event()
gevent.signal(signal.SIGQUIT, evt.set)
gevent.signal(signal.SIGTERM, evt.set)
gevent.signal(signal.SIGINT, evt.set)
evt.wait()
# finally stop
app.stop()
def logger_init():
logging.basicConfig(level=logging.DEBUG, format='%(module)15s %(asctime)s %(message)s', datefmt='%H:%M:%S')
if config().log.to_file:
log_filename = os.path.join(config().log.dir,
config().log.prefix + datetime.datetime.now().strftime("%m%d%H%M%S"))
logging.getLogger().addHandler(logging.FileHandler(log_filename))
if config().log.dump_config:
dump_config()
def run(ctx, dev):
"""Start the client ( --dev to stop on error)"""
# create app
app = EthApp(ctx.obj['config'])
# development mode
if dev:
gevent.get_hub().SYSTEM_ERROR = BaseException
try:
ctx.obj['config']['client_version'] += '/' + os.getlogin()
except:
log.warn("can't get and add login name to client_version")
pass
# dump config
konfig.dump_config(ctx.obj['config'])
# register services
for service in services:
assert issubclass(service, BaseService)
if service.name not in app.config['deactivated_services']:
assert service.name not in app.services
service.register_with_app(app)
assert hasattr(app.services, service.name)
# start app
app.start()
# wait for interrupt
evt = Event()
gevent.signal(signal.SIGQUIT, evt.set)
gevent.signal(signal.SIGTERM, evt.set)
gevent.signal(signal.SIGINT, evt.set)
evt.wait()
# finally stop
app.stop()
def main():
config = c.get_default_config()
parser = ArgumentParser(version=__version__)
config = create_config(parser)
# Logging
configure_logging(config.get('misc', 'logging') or '', verbosity=config.getint('misc', 'verbosity'))
logger.info('PyEPM %s', __version__)
logger.info('=====')
logger.debug(c.dump_config(config))
args = parser.parse_args()
for filename in args.filename:
if not os.path.exists(filename):
logger.warn("File does not exist: %s" % filename)
else:
logger.info("Deploying %s..." % filename)
deployment = deploy.Deploy(filename, config)
deployment.deploy()
def main():
config = c.get_default_config()
parser = ArgumentParser(version=__version__)
config = create_config(parser)
# Logging
configure_logging(config.get('misc', 'logging') or '', verbosity=config.getint('misc', 'verbosity'))
logger.info(colors.HEADER + '=====' + colors.ENDC)
logger.info(colors.OKGREEN + 'PyEPM ' + colors.ENDC + '%s', __version__)
logger.info(colors.HEADER + '=====' + colors.ENDC)
logger.debug(c.dump_config(config))
args = parser.parse_args()
for filename in args.filename:
if not os.path.exists(filename):
logger.warn("\nFile does not exist: %s" % filename)
else:
logger.info("\nDeploying " + colors.BOLD + "%s" % filename + colors.ENDC + "...")
deployment = deploy.Deploy(filename, config)
deployment.deploy()
def main():
config = c.get_default_config()
parser = ArgumentParser(version=__version__)
config = create_config(parser)
# Logging
configure_logging(config.get('misc', 'logging') or '',
verbosity=config.getint('misc', 'verbosity'))
logger.info('PyEPM %s', __version__)
logger.info('=====')
logger.debug(c.dump_config(config))
args = parser.parse_args()
for filename in args.filename:
if not os.path.exists(filename):
logger.warn("File does not exist: %s" % filename)
else:
logger.info("Deploying %s..." % filename)
deployment = deploy.Deploy(filename, config)
deployment.deploy()
def config(ctx):
"""Show the config"""
konfig.dump_config(ctx.obj['config'])
def run_training(cfg):
"""
run the training loops
"""
# torch gpu setup
os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.gpu_idx)
if cfg.model_zoo is not None:
os.environ["TORCH_MODEL_ZOO"] = cfg.model_zoo
# make the exp dir
os.makedirs(cfg.exp_dir, exist_ok=True)
# set the seeds
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
# set cudnn to reproducibility mode
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# dump the exp config to the exp dir
dump_config(cfg)
# setup datasets
dset_train, dset_val, dset_test = dataset_zoo(**cfg.DATASET)
# init loaders
trainloader = torch.utils.data.DataLoader(dset_train,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=True)
if dset_val is not None:
valloader = torch.utils.data.DataLoader(dset_val,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=False)
else:
valloader = None
# test loaders
if dset_test is not None:
testloader = torch.utils.data.DataLoader(dset_test,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=False)
_, _, eval_vars = eval_zoo(cfg.DATASET.dataset_name)
else:
testloader = None
eval_vars = None
# init the model
model, stats, optimizer_state = init_model(cfg, add_log_vars=eval_vars)
start_epoch = stats.epoch + 1
# move model to gpu
if torch.cuda.is_available():
model.cuda()
# init the optimizer
optimizer, scheduler = init_optimizer(model,
optimizer_state=optimizer_state,
**cfg.SOLVER)
# loop through epochs
scheduler.last_epoch = start_epoch
for epoch in range(start_epoch, cfg.SOLVER.max_epochs):
with stats: # automatic new_epoch and plotting at every epoch start
print("scheduler lr = %1.2e" % float(scheduler.get_lr()[-1]))
# train loop
trainvalidate(model,
stats,
epoch,
trainloader,
optimizer,
False,
visdom_env_root=get_visdom_env(cfg),
**cfg)
# val loop
if valloader is not None:
trainvalidate(model,
stats,
epoch,
valloader,
optimizer,
True,
visdom_env_root=get_visdom_env(cfg),
**cfg)
# eval loop (optional)
if testloader is not None:
eval_result = run_eval(cfg, model, testloader, stats=stats)
dump_eval_result(cfg, eval_result)
assert stats.epoch == epoch, "inconsistent stats!"
# delete previous models if required
if cfg.store_checkpoints_purge > 0 and cfg.store_checkpoints:
for prev_epoch in range(epoch - cfg.store_checkpoints_purge):
purge_epoch(cfg.exp_dir, prev_epoch)
# save model
if cfg.store_checkpoints:
outfile = get_checkpoint(cfg.exp_dir, epoch)
save_model(model, stats, outfile, optimizer=optimizer)
scheduler.step()
# the final eval
if testloader is not None:
eval_result = run_eval(cfg, model, testloader, stats=None)
dump_eval_result(cfg, eval_result)
return eval_result
else:
return None
def run_training(cfg):
# run the training loops
# make the exp dir
os.makedirs(cfg.exp_dir, exist_ok=True)
# set the seed
np.random.seed(cfg.seed)
# dump the exp config to the exp dir
dump_config(cfg)
# setup datasets
dset_train, dset_val, dset_test = dataset_zoo(**cfg.DATASET)
# init loaders
if cfg.batch_sampler == 'default':
trainloader = torch.utils.data.DataLoader(dset_train,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=False)
elif cfg.batch_sampler == 'sequence':
trainloader = torch.utils.data.DataLoader(
dset_train,
num_workers=cfg.num_workers,
pin_memory=True,
batch_sampler=SceneBatchSampler(
torch.utils.data.SequentialSampler(dset_train),
cfg.batch_size,
True,
))
else:
raise BaseException()
if dset_val is not None:
if cfg.batch_sampler == 'default':
valloader = torch.utils.data.DataLoader(
dset_val,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=False)
elif cfg.batch_sampler == 'sequence':
valloader = torch.utils.data.DataLoader( dset_val,
num_workers=cfg.num_workers, pin_memory=True,
batch_sampler=SceneBatchSampler( \
torch.utils.data.SequentialSampler(dset_val),
cfg.batch_size,
True,
) )
else:
raise BaseException()
else:
valloader = None
# test loaders
if dset_test is not None:
testloader = torch.utils.data.DataLoader(
dset_test,
num_workers=cfg.num_workers,
pin_memory=True,
batch_size=cfg.batch_size,
shuffle=False,
)
_, _, eval_vars = eval_zoo(cfg.DATASET.dataset_name)
else:
testloader = None
eval_vars = None
# init the model
model, stats, optimizer_state = init_model(cfg, add_log_vars=eval_vars)
start_epoch = stats.epoch + 1
# annotate dataset with c3dpo outputs
if cfg.annotate_with_c3dpo_outputs:
for dset in dset_train, dset_val, dset_test:
if dset is not None:
run_c3dpo_model_on_dset(dset, cfg.MODEL.nrsfm_exp_path)
# move model to gpu
model.cuda(0)
# init the optimizer
optimizer, scheduler = init_optimizer(\
model, optimizer_state=optimizer_state, **cfg.SOLVER)
# loop through epochs
scheduler.last_epoch = start_epoch
for epoch in range(start_epoch, cfg.SOLVER.max_epochs):
with stats: # automatic new_epoch and plotting of stats at every epoch start
print("scheduler lr = %1.2e" % float(scheduler.get_lr()[-1]))
# train loop
trainvalidate(model, stats, epoch, trainloader, optimizer, False, \
visdom_env_root=get_visdom_env(cfg), **cfg )
# val loop
if valloader is not None:
trainvalidate(model, stats, epoch, valloader, optimizer, True, \
visdom_env_root=get_visdom_env(cfg), **cfg )
# eval loop (optional)
if testloader is not None:
if cfg.eval_interval >= 0:
if cfg.eval_interval == 0 or \
((epoch % cfg.eval_interval)==0 and epoch > 0):
torch.cuda.empty_cache(
) # we have memory heavy eval ...
with torch.no_grad():
run_eval(cfg, model, stats, testloader)
assert stats.epoch == epoch, "inconsistent stats!"
# delete previous models if required
if cfg.store_checkpoints_purge > 0 and cfg.store_checkpoints:
for prev_epoch in range(epoch - cfg.store_checkpoints_purge):
period = cfg.store_checkpoints_purge_except_every
if (period > 0 and prev_epoch % period == period - 1):
continue
purge_epoch(cfg.exp_dir, prev_epoch)
# save model
if cfg.store_checkpoints:
outfile = get_checkpoint(cfg.exp_dir, epoch)
save_model(model, stats, outfile, optimizer=optimizer)
scheduler.step()
def config(ctx):
"""Show the config"""
konfig.dump_config(ctx.obj['config'])
def set_repo_version(config, version):
config['REPO_VERSION'] = version
dump_config(config, config_path)
from config import config, overwrite_config_with_args, dump_config
from read_data import index_ent_rel, graph_size, read_data
from data_utils import heads_tails, inplace_shuffle, batch_by_num
from trans_e import TransE
from trans_d import TransD
from distmult import DistMult
from compl_ex import ComplEx
from logger_init import logger_init
from select_gpu import select_gpu
from corrupter import BernCorrupterMulti
logger_init()
torch.cuda.set_device(select_gpu())
overwrite_config_with_args()
dump_config()
task_dir = config().task.dir
kb_index = index_ent_rel(os.path.join(task_dir, 'train.txt'),
os.path.join(task_dir, 'valid.txt'),
os.path.join(task_dir, 'test.txt'))
n_ent, n_rel = graph_size(kb_index)
models = {'TransE': TransE, 'TransD': TransD, 'DistMult': DistMult, 'ComplEx': ComplEx}
gen_config = config()[config().g_config]
dis_config = config()[config().d_config]
gen = models[config().g_config](n_ent, n_rel, gen_config)
dis = models[config().d_config](n_ent, n_rel, dis_config)
gen.load(os.path.join(task_dir, gen_config.model_file))
dis.load(os.path.join(task_dir, dis_config.model_file))
def dump_config(self, conf):
config.dump_config(conf, os.path.join(self.data_dir, 'config.cnf'));
