def main(cfg: default.ModelNet40Config) -> None:
# ifx paths
cfg = monitor.fix_path_in_configs(CW_DIR, cfg, [["dataset", "root"]])
# set a seed
pytorch_tools.set_seed(cfg.general.seed, cfg.general.device,
cfg.general.reproducibility)
# set a device
cfg.general.device = pytorch_tools.select_device(cfg.general.device)
# training env
## Get model.
model = get_model(cfg)
## Get dataset.
dataset, train_collate_fn = get_dataset(cfg)
## Get loader
train_dataset_loader = get_loader(cfg,
dataset["train"],
shuffle=cfg.loader.shuffle,
collate_fn=train_collate_fn)
test_dataset_loader = get_loader(cfg, dataset["test"])
## Get loss functions.
criterion = get_losses(cfg)
## Get optimizer.
optimizer = get_optimizer(cfg, model)
## Get scheduler.
scheduler = get_scheduler(cfg, optimizer)
## Get logger.
writer = get_writer(cfg)
# progress bar
loader = tqdm(range(cfg.general.start_epoch, cfg.general.epochs),
desc="Training",
ncols=70)
# training start
for epoch in loader:
# training
train_log = train(cfg, model, train_dataset_loader, optimizer,
criterion, scheduler)
# save training log
monitor.dict2logger(train_log, writer, epoch, cfg.writer.name)
# test
eval_log = eval(cfg, model, test_dataset_loader, criterion)
# save eval log
monitor.dict2logger(eval_log, writer, epoch, cfg.writer.name)
# save params and model
if (epoch+1) % cfg.general.save_epoch == 0 and \
cfg.general.save_epoch != -1:
save_params("model.path.tar", epoch + 1, cfg, model, optimizer,
scheduler)
save_params("f_model.path.tar", cfg.general.epochs, cfg, model, optimizer,
scheduler)
print("Finish training.")
def main(cfg: configs.ModelNet40Config) -> None:
# ifx paths
cfg = monitor.fix_path_in_configs(
CW_DIR, cfg, [["dataset", "root"], ["model", "resume"]])
# set a seed
pytorch_tools.set_seed(cfg.general.seed, cfg.general.device,
cfg.general.reproducibility)
# set a device
cfg.general.device = pytorch_tools.select_device(cfg.general.device)
# get a trained model
checkpoint, checkpoint_cfg = get_checkpoint(cfg.model.resume)
# test env
## Get model.
model = get_model(checkpoint_cfg)
## Get dataset and loader.
dataset = get_dataset(checkpoint_cfg)
datset_loader = get_loader(checkpoint_cfg, dataset)
## Get loss functions.
criterion = get_losses(checkpoint_cfg)
# set trained params
model.load_state_dict(checkpoint["model"])
# test start
test_log = test(cfg, model, datset_loader["test"], criterion)
# show results
print(test_log)
print("Finish test.")
def main(cfg: configs.KITTIConfig) -> None:
# ifx paths
cfg = monitor.fix_path_in_configs(CW_DIR, cfg, [["dataset", "dataset_path"]])
# set a seed
pytorch_tools.set_seed(
cfg.general.seed,
cfg.general.device,
cfg.general.reproducibility
)
# set a device
cfg.general.device = pytorch_tools.select_device(cfg.general.device)
## I will remove this code.
log_file = os.path.join("./", 'log_train.txt')
logger = create_logger(log_file)
# training env
## Get model.
model = get_model(cfg)
## Get dataset and loader.
dataset = get_dataset(cfg, logger)
datset_loader = get_loader(cfg, dataset)
## Get loss functions.
criterion = get_losses(cfg)
## Get optimizer.
optimizer = get_optimizer(cfg, model)
## Get scheduler.
scheduler = get_scheduler(cfg, optimizer)
## Get logger.
writer = get_writer(cfg)
# progress bar
loader = tqdm(range(cfg.general.start_epoch, cfg.general.epochs),
desc="Training", ncols=70)
# training start
for epoch in loader:
# training
train_log = train(cfg, model, datset_loader, optimizer, criterion,
scheduler)
# save training log
monitor.dict2logger(train_log, writer, epoch, cfg.writer.name)
# save params and model
if (epoch+1) % cfg.general.save_epoch == 0 and \
cfg.general.save_epoch != -1:
save_params("model.path.tar", epoch+1, cfg, model, optimizer,
scheduler)
print('Epoch {}/{}:'.format(cfg.general.epochs, cfg.general.epochs))
save_params("f_model.path.tar", cfg.general.epochs, cfg, model, optimizer,
scheduler)
print("Finish training.")