def do_train(cfg, train_loader, valid_dict, tr_comp: TrainComponent, saver):
# tb_log = TensorBoardXLog(cfg, saver.save_dir)
trainer = create_supervised_trainer(tr_comp.model,
tr_comp.optimizer,
tr_comp.loss,
device=cfg.MODEL.DEVICE,
apex=cfg.APEX.IF_ON)
evaler = Eval(valid_dict, cfg.MODEL.DEVICE)
evaler.get_valid_eval_map(cfg, tr_comp.model)
run(cfg, train_loader, tr_comp, saver, trainer, evaler)
def fine_tune_current_model(cfg, train_loader, valid_dict, tr_comp, saver):
for param in tr_comp.model.base.parameters():
param.requires_grad = False
trainer = create_supervised_trainer(tr_comp.model,
tr_comp.optimizer,
tr_comp.loss,
device=cfg.MODEL.DEVICE,
apex=cfg.APEX.IF_ON)
evaler = Eval(valid_dict, cfg.MODEL.DEVICE)
evaler.get_valid_eval_map(cfg, tr_comp.model)
copy_cfg = copy.deepcopy(cfg)
copy_cfg["TRAIN"]["MAX_EPOCHS"] = 60
run(copy_cfg, train_loader, tr_comp, saver, trainer, evaler)
def train_autoencoder(cfg, train_loader, valid_dict,
source_tr_comp: TrainComponent,
current_tr_comp: TrainComponent, saver):
trainer = create_autoencoder_trainer(source_tr_comp.model,
current_tr_comp.model,
current_tr_comp.optimizer,
current_tr_comp.loss,
device=cfg.MODEL.DEVICE,
apex=cfg.APEX.IF_ON)
evaler = Eval(valid_dict, cfg.MODEL.DEVICE)
evaler.get_valid_eval_map_autoencoder(cfg, source_tr_comp.model,
current_tr_comp.model)
copy_cfg = copy.deepcopy(cfg)
copy_cfg["TRAIN"]["MAX_EPOCHS"] = 90
run(copy_cfg, train_loader, current_tr_comp, saver, trainer, evaler)
def ebll_train(cfg, train_loader, valid_dict, source_tr_comp, current_tr_comp,
autoencoder_tr, saver):
for param in current_tr_comp.model.base.parameters():
param.requires_grad = True
trainer = create_ebll_trainer(source_tr_comp.model,
autoencoder_tr.model,
current_tr_comp.model,
current_tr_comp.optimizer,
current_tr_comp.loss,
apex=cfg.APEX.IF_ON,
device=cfg.MODEL.DEVICE)
evaler = Eval(valid_dict, cfg.MODEL.DEVICE)
evaler.get_valid_eval_map_ebll(cfg, source_tr_comp.model,
current_tr_comp.model)
run(cfg, train_loader, current_tr_comp, saver, trainer, evaler)