Method start

tasks/base_task.py:220–258

(cls)

    @classmethod
    def start(cls):
        set_hparams()
        os.environ['MASTER_PORT'] = str(random.randint(15000, 30000))
        random.seed(hparams['seed'])
        np.random.seed(hparams['seed'])
        task = cls()
        work_dir = hparams['work_dir']
        trainer = BaseTrainer(checkpoint_callback=LatestModelCheckpoint(
                                  filepath=work_dir,
                                  verbose=True,
                                  monitor='val_loss',
                                  mode='min',
                                  num_ckpt_keep=hparams['num_ckpt_keep'],
                                  save_best=hparams['save_best'],
                                  period=1 if hparams['save_ckpt'] else 100000
                              ),
                              logger=TensorBoardLogger(
                                  save_dir=work_dir,
                                  name='lightning_logs',
                                  version='lastest'
                              ),
                              gradient_clip_val=hparams['clip_grad_norm'],
                              val_check_interval=hparams['val_check_interval'],
                              row_log_interval=hparams['log_interval'],
                              max_updates=hparams['max_updates'],
                              num_sanity_val_steps=hparams['num_sanity_val_steps'] if not hparams[
                                  'validate'] else 10000,
                              accumulate_grad_batches=hparams['accumulate_grad_batches'])
        if not hparams['infer']:  # train
            t = datetime.now().strftime('%Y%m%d%H%M%S')
            code_dir = f'{work_dir}/codes/{t}'
            subprocess.check_call(f'mkdir -p "{code_dir}"', shell=True)
            for c in hparams['save_codes']:
                subprocess.check_call(f'cp -r "{c}" "{code_dir}/"', shell=True)
            print(f"| Copied codes to {code_dir}.")
            trainer.checkpoint_callback.task = task
            trainer.fit(task)
        else:
            trainer.test(task)

    def configure_ddp(self, model, device_ids):
        model = DDP(