def get_args():
parser = HyperOptArgumentParser()
parser.add_argument('--work_dir',
metavar='DIR',
default="./work_dir",
type=str,
help='path to save output')
parser.add_argument('--proj_name', type=str)
parser.add_argument('--name', type=str)
parser.add_argument('--gpus', type=str, default='7', help='how many gpus')
parser.add_argument('--dist_bd',
type=str,
default='dp',
choices=('dp', 'ddp', 'ddp2'),
help='supports three options dp, ddp, ddp2')
parser.add_argument('--use_16bit',
dest='use_16bit',
action='store_true',
help='if true uses 16 bit precision')
parser.add_argument('--eval',
'--evaluate',
dest='evaluate',
action='store_true',
help='evaluate model on validation set')
parser.add_argument('--seed', default=1, type=int)
parser.add_argument('--load_mem', action='store_true')
parser.add_argument('--track_grad_norm', action='store_true')
parser = CGCNModel.add_model_specific_args(parser)
return parser.parse_args()
def main(params, gpus=None, results_dict=None):
init_seed(params)
params.gpu_num = len(params.gpus.split(','))
params.dataset = 'gqa_graph'
if params.work_dir == './work_dir':
params.work_dir = params.work_dir + f'/{params.dataset}'
if not params.track_grad_norm:
params.track_grad_norm = -1
else:
params.track_grad_norm = 1
params.lr = params.lr * params.gpu_num
params.batch_size = params.batch_size * params.gpu_num
auto_set_name(params)
logger = TestTubeLogger(params.work_dir, name=params.name)
model = CGCNModel(params)
trainer = pl.Trainer(
logger=logger,
default_save_path=params.work_dir,
gpus=params.gpus,
max_nb_epochs=params.epochs,
distributed_backend=params.dist_bd,
use_amp=params.use_16bit,
# nb_sanity_val_steps=0,
# val_check_interval=0.01,
# val_percent_check=0.001,
# train_percent_check=0.001,
early_stop_callback=False,
max_epochs=params.epochs,
track_grad_norm=params.track_grad_norm,
# log_gpu_memory='all',
)
def main(params):
if params.seed is not None:
random.seed(params.seed)
torch.manual_seed(params.seed)
random.seed(params.seed)
np.random.seed(params.seed)
torch.manual_seed(params.seed)
torch.cuda.manual_seed(params.seed)
cudnn.deterministic = True
logger = TestTubeLogger(params.save_path, name=params.dataset)
model = CGCNModel(params)
trainer = pl.Trainer(
logger=logger,
default_save_path=params.save_path,
gpus=params.gpus,
max_nb_epochs=params.epochs,
distributed_backend=params.dist_backend,
use_amp=params.use_16bit,
# nb_sanity_val_steps=0,
# val_check_interval=0.01,
# val_percent_check=0.01,
# train_percent_check=0.002,
early_stop_callback=False,
)
if params.evaluate:
trainer.run_evaluation()
else:
trainer.fit(model)
def main(params, gpus=None, results_dict=None):
init_seed(params)
params.padding_idx = 0
if params.work_dir == './work_dir':
params.work_dir = params.work_dir + f'/{params.dataset}'
auto_set_name(params)
if not params.track_grad_norm:
params.track_grad_norm = -1
else:
params.track_grad_norm = 1
logger = TestTubeLogger(params.work_dir, name=params.name)
model = CGCNModel(params)
idx2name = {}
for idx, (p_nam, p) in enumerate(model.named_parameters()):
idx2name[idx] = p_nam
gpu_num = len(params.gpus.split(','))
params.lr = params.lr * gpu_num
trainer = pl.Trainer(
logger=logger,
default_save_path=params.work_dir,
gpus=params.gpus,
max_nb_epochs=params.epochs,
distributed_backend=params.dist_bd,
use_amp=params.use_16bit,
# nb_sanity_val_steps=0,
# val_check_interval=0.01,
# val_percent_check=0.001,
# train_percent_check=0.001,
early_stop_callback=False,
max_epochs=params.epochs,
track_grad_norm=params.track_grad_norm,
# log_gpu_memory='all',
)
if params.evaluate:
trainer.run_evaluation()
else:
trainer.fit(model)
def main(params, gpus=None, results_dict=None):
init_seed(params)
params.gpu_num = len(params.gpus.split(','))
params.dataset = 'gqa_graph'
if params.proj_name is None:
params.proj_name = params.dataset
params.work_dir = params.work_dir + f'/{params.proj_name}'
if not params.track_grad_norm:
params.track_grad_norm = -1
else:
params.track_grad_norm = 1
params.lr = params.lr * params.gpu_num * (params.batch_size / 64.)
params.batch_size = params.batch_size * params.gpu_num
auto_set_name(params)
logger = TestTubeLogger(params.work_dir, name=params.name)
model = CGCNModel(params)
# checkpoint = ModelCheckpoint()
trainer = pl.Trainer(
logger=logger,
default_save_path=params.work_dir,
gpus=params.gpus,
distributed_backend=params.dist_bd,
use_amp=params.use_16bit,
# nb_sanity_val_steps=0,
# val_check_interval=0.01,
# val_percent_check=0.001,
# train_percent_check=0.001,
early_stop_callback=False,
max_epochs=params.epochs,
# max_epochs=1,
track_grad_norm=params.track_grad_norm,
# log_gpu_memory='all',
# checkpoint_callback=False,
row_log_interval=100,
gradient_clip_val=params.grad_clip)
if params.evaluate:
trainer.run_evaluation()
else:
trainer.fit(model)