optimizer = torch.optim.SGD(params_list,
lr=base_lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
# config lr policy
total_iteration = config.nepochs * config.niters_per_epoch
lr_policy = PolyLR(base_lr, config.lr_power, total_iteration)
if engine.distributed:
if torch.cuda.is_available():
model.cuda()
model = DistributedDataParallel(model)
else:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DataParallelModel(model, device_ids=engine.devices)
model.to(device)
engine.register_state(dataloader=train_loader,
model=model,
optimizer=optimizer)
if engine.continue_state_object:
engine.restore_checkpoint()
model.train()
for epoch in range(engine.state.epoch, config.nepochs):
if engine.distributed:
train_sampler.set_epoch(epoch)
bar_format = '{desc}[{elapsed}<{remaining},{rate_fmt}]'
pbar = tqdm(range(config.niters_per_epoch),
# config lr policy
total_iteration = config.nepochs * config.niters_per_epoch
lr_policy = PolyLR(base_lr, config.lr_power, total_iteration)
optimizer = torch.optim.SGD(params_list,
lr=base_lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
if engine.distributed:
if torch.cuda.is_available():
model.cuda()
model = DistributedDataParallel(model)
else:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DataParallelModel(model, engine.devices)
model.to(device)
engine.register_state(dataloader=train_loader,
model=model,
optimizer=optimizer)
if engine.continue_state_object:
engine.restore_checkpoint()
optimizer.zero_grad()
model.train()
for epoch in range(engine.state.epoch, config.nepochs):
if engine.distributed:
train_sampler.set_epoch(epoch)
bar_format = '{desc}[{elapsed}<{remaining},{rate_fmt}]'