Method step

bing_bert/nvidia/optimizers.py:170–276

Performs a single optimization step. Arguments: closure (callable, optional): A closure that reevaluates the model and returns the loss.

(self, closure=None)

                        )

    def step(self, closure=None):
        """Performs a single optimization step.

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            loss = closure()
        check = 1#torch.norm(all_grads, 2)

        grad_list = []
        param_list = []
        per_param_decay = []
        momentum = []
        velocity = []

        fp16_grad_list = []
        fp16_from_fp32_param_list = []
        fp32_param_list = []
        fp16_per_param_decay = []
        fp16_momentum = []
        fp16_velocity = []
        
        if not self.updates_created:
            self.update = []
            self.fp16_update = []
        for group in self.param_groups:
            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad.data
                if grad.is_sparse:
                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    # Keep step here for compatibility with earlier resume from checkpoint
                    state['step'] = 0
                    # Exponential moving average of gradient values
                    state['momentum'] = torch.zeros_like(p.data, dtype=torch.float32)
                    # Exponential moving average of squared gradient values
                    state['velocity'] = torch.zeros_like(p.data, dtype=torch.float32)
                    # fp32 master weights
                if 'master_param' not in state.keys() and p.type() == 'torch.cuda.HalfTensor':
                    state['master_param'] = p.detach().clone().float()

                # ensure these 3 are float tensors
                if state['momentum'].type() != 'torch.cuda.FloatTensor':
                    state['momentum'] = state['momentum'].float()
                if state['velocity'].type() != 'torch.cuda.FloatTensor':
                    state['velocity'] = state['velocity'].float()
                if 'master_param' in state.keys() and state['master_param'].type() != 'torch.cuda.FloatTensor':
                    state['master_param'] = state['master_param'].float()