d646a4a8fc315ee0146365ba86c4f8afd9cfdb2c,fairseq/multiprocessing_trainer.py,MultiprocessingTrainer,_async_init,#MultiprocessingTrainer#Any#Any#Any#Any#Any#Any#,54

Before Change

        nccl.initialize(self.num_replicas, nccl_uid, device_id)

        // copy model and criterion to current device
        self.model = model.cuda()
        self.criterion = criterion.cuda()

        // initialize optimizer
        self.optimizer = NAG(self.model.parameters(), lr=self.args.lr,
                             momentum=self.args.momentum,
                             weight_decay=self.args.weight_decay)
        self.flat_grads = None
        self.loss = None

        // initialize LR scheduler

After Change

        self.criterion = criterion.cuda()

        // initialize optimizer
        self.optimizer = self._build_optimizer()
        self.flat_grads = None
        self.loss = None

        // initialize LR scheduler

Instances