def main(pred_file, split_type='valid', **kwargs):
if 'cfg' not in kwargs:
from extended_config import (
cfg as conf,
key_maps,
# CN,
update_from_dict,
# post_proc_config
)
cfg = conf
cfg = update_from_dict(cfg, kwargs, key_maps)
else:
cfg = kwargs['cfg']
cfg.freeze()
# grnd_eval = GroundEval_Corr(cfg)
# grnd_eval = GroundEvalDS4(cfg)
comm = Munch()
exp = cfg.ds.exp_setting
if exp == 'gt5':
comm.num_prop_per_frm = 5
elif exp == 'p100':
comm.num_prop_per_frm = 100
else:
raise NotImplementedError
conc_type = cfg.ds.conc_type
if conc_type == 'sep' or conc_type == 'svsq':
grnd_eval = GroundEval_SEP(cfg, comm)
elif conc_type == 'temp':
grnd_eval = GroundEval_TEMP(cfg, comm)
elif conc_type == 'spat':
grnd_eval = GroundEval_SPAT(cfg, comm)
else:
raise NotImplementedError
out = grnd_eval.eval_ground_acc(pred_file, split_type=split_type)
# to_print = ['avg1', 'avg2']
# print(Counter(grnd_eval.pcs))
met_keys = ['avg1', 'avg1_cons',
'avg1_vidf', 'avg1_strict']
print({k: out[k] for k in met_keys})
# print(Counter(grnd_eval.stuff))
# return out
return
def main_dist(uid: str, **kwargs):
"""
uid is a unique identifier for the experiment name
Can be kept same as a previous run, by default will start executing
from latest saved model
**kwargs: allows arbit arguments of cfg to be changed
"""
cfg = conf
num_gpus = torch.cuda.device_count()
cfg.num_gpus = num_gpus
if num_gpus > 1:
if 'local_rank' in kwargs:
# We are doing distributed parallel
cfg.do_dist = True
torch.cuda.set_device(kwargs['local_rank'])
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
else:
# We are doing data parallel
cfg.do_dist = False
# Update the config file depending on the command line args
cfg = update_from_dict(cfg, kwargs, key_maps)
# Freeze the cfg, can no longer be changed
cfg.freeze()
# print(cfg)
# Initialize learner
learn = learner_init(uid, cfg)
# Train or Test
if not (cfg.only_val or cfg.only_test):
learn.fit(epochs=cfg.epochs, lr=cfg.lr)
else:
if cfg.only_val:
learn.testing(learn.data.valid_dl)
if cfg.only_test:
learn.testing(learn.data.test_dl)
def main(pred_file, split_type='valid', **kwargs):
if 'cfg' not in kwargs:
from extended_config import (
cfg as conf,
key_maps,
# CN,
update_from_dict,
# post_proc_config
)
cfg = conf
cfg = update_from_dict(cfg, kwargs, key_maps)
else:
cfg = kwargs['cfg']
cfg.freeze()
# grnd_eval = GroundEval_Corr(cfg)
# grnd_eval = GroundEvalDS4(cfg)
comm = Munch()
exp = cfg.ds.exp_setting
if exp == 'gt5':
comm.num_prop_per_frm = 5
elif exp == 'p100':
comm.num_prop_per_frm = 100
else:
raise NotImplementedError
conc_type = cfg.ds.conc_type
if conc_type == 'sep' or conc_type == 'svsq':
avis = ASRL_Vis_SEP(cfg, comm)
elif conc_type == 'temp':
avis = ASRL_Vis_TEMP(cfg, comm)
elif conc_type == 'spat':
avis = ASRL_Vis_SPAT(cfg, comm)
else:
raise NotImplementedError
# avis.draw_boxes_all_indices(
# pred_file, split_type=split_type
# )
return avis
def main_dist(uid: str, **kwargs):
"""
uid is a unique identifier for the experiment name
Can be kept same as a previous run, by default will start executing
from latest saved model
**kwargs: allows arbit arguments of cfg to be changed
"""
cfg = conf
num_gpus = torch.cuda.device_count()
cfg.num_gpus = num_gpus
cfg.uid = uid
cfg.cmd = sys.argv
if num_gpus > 1:
if 'local_rank' in kwargs:
# We are doing distributed parallel
cfg.do_dist = True
torch.cuda.set_device(kwargs['local_rank'])
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
else:
# We are doing data parallel
cfg.do_dist = False
# cfg.do_dp = True
# Update the config file depending on the command line args
cfg = update_from_dict(cfg, kwargs, key_maps)
cfg = post_proc_config(cfg)
# Freeze the cfg, can no longer be changed
cfg.freeze()
# print(cfg)
# Initialize learner
learn = learner_init(uid, cfg)
# Train or Test
if not (cfg.only_val or cfg.only_test or cfg.overfit_batch):
learn.fit(epochs=cfg.train.epochs, lr=cfg.train.lr)
if cfg.run_final_val:
print('Running Final Validation using best model')
learn.load_model_dict(
resume_path=learn.model_file,
load_opt=False
)
val_loss, val_acc, _ = learn.validate(
db={'valid': learn.data.valid_dl},
write_to_file=True
)
print(val_loss)
print(val_acc)
else:
pass
else:
if cfg.overfit_batch:
learn.overfit_batch(1000, 1e-4)
if cfg.only_val:
val_loss, val_acc, _ = learn.validate(
db={'valid': learn.data.valid_dl},
write_to_file=True
)
print(val_loss)
print(val_acc)
# learn.testing(learn.data.valid_dl)
pass
if cfg.only_test:
# learn.testing(learn.data.test_dl)
test_loss, test_acc, _ = learn.validate(
db=learn.data.test_dl)
print(test_loss)
print(test_acc)
return