def learner_init(uid: str, cfg: CN) -> Learner:
device = torch.device('cuda')
data = get_data(cfg)
# Ugly hack because I wanted ratios, scales
# in fractional formats
if type(cfg['ratios']) != list:
ratios = eval(cfg['ratios'], {})
else:
ratios = cfg['ratios']
if type(cfg['scales']) != list:
scales = cfg['scale_factor'] * np.array(eval(cfg['scales'], {}))
else:
scales = cfg['scale_factor'] * np.array(cfg['scales'])
num_anchors = len(ratios) * len(scales)
mdl = get_default_net(num_anchors=num_anchors, cfg=cfg)
mdl.to(device)
if cfg.do_dist:
mdl = torch.nn.parallel.DistributedDataParallel(
mdl,
device_ids=[cfg.local_rank],
output_device=cfg.local_rank,
broadcast_buffers=True,
find_unused_parameters=True)
elif not cfg.do_dist and cfg.num_gpus:
# Use data parallel
mdl = torch.nn.DataParallel(mdl)
loss_fn = get_default_loss(ratios, scales, cfg)
loss_fn.to(device)
eval_fn = get_default_eval(ratios, scales, cfg)
# eval_fn.to(device)
opt_fn = partial(torch.optim.Adam, betas=(0.9, 0.99))
learn = Learner(uid=uid,
data=data,
mdl=mdl,
loss_fn=loss_fn,
opt_fn=opt_fn,
eval_fn=eval_fn,
device=device,
cfg=cfg)
return learn
def learner_init(uid, cfg):
device_count = torch.cuda.device_count()
device = torch.device('cuda')
if type(cfg['ratios']) != list:
ratios = eval(cfg['ratios'], {})
else:
ratios = cfg['ratios']
if type(cfg['scales']) != list:
scales = cfg['scale_factor'] * np.array(eval(cfg['scales'], {}))
else:
scales = cfg['scale_factor'] * np.array(cfg['scales'])
num_anchors = len(ratios) * len(scales)
qnet = get_default_net(num_anchors=num_anchors, cfg=cfg)
qnet = qnet.to(device)
qnet = torch.nn.DataParallel(qnet)
qlos = get_default_loss(ratios, scales, cfg)
qlos = qlos.to(device)
qeval = Evaluator(ratios, scales, cfg)
# db = get_data(bs=cfg['bs'] * device_count, nw=cfg['nw'], bsv=cfg['bsv'] * device_count,
# nwv=cfg['nwv'], devices=cfg['devices'], do_tfms=cfg['do_tfms'],
# cfg=cfg, data_cfg=data_cfg)
# db = get_data(cfg, ds_name=cfg['ds_to_use'])
db = get_data(cfg)
opt_fn = partial(torch.optim.Adam, betas=(0.9, 0.99))
# Note: Currently using default optimizer
learn = Learner(uid=uid,
data=db,
mdl=qnet,
loss_fn=qlos,
opt_fn=opt_fn,
eval_fn=qeval,
device=device,
cfg=cfg)
return learn