def backward(ctx, *grad_output): """ In the backward pass we receive a Tensor containing the gradient of the loss with respect to the output, and we need to compute the gradient of the loss with respect to the input. """ *param_list, scale = ctx.saved_tensors dscale_dL = sum( grad.sum() for grad in torch._foreach_mul(param_list, grad_output)) dparam_dL = torch._foreach_mul(grad_output, float(scale)) return (dscale_dL, ) + dparam_dL
def _multi_tensor_rprop(params: List[Tensor], grads: List[Tensor], prevs: List[Tensor], step_sizes: List[Tensor], *, step_size_min: float, step_size_max: float, etaminus: float, etaplus: float): if len(params) == 0: return signs = torch._foreach_mul(grads, prevs) signs = [s.sign() for s in signs] for sign in signs: sign[sign.gt(0)] = etaplus sign[sign.lt(0)] = etaminus sign[sign.eq(0)] = 1 # update stepsizes with step size updates torch._foreach_mul_(step_sizes, signs) for step_size in step_sizes: step_size.clamp_(step_size_min, step_size_max) # for dir<0, dfdx=0 # for dir>=0 dfdx=dfdx for i in range(len(grads)): grads[i] = grads[i].clone(memory_format=torch.preserve_format) grads[i][signs[i].eq(etaminus)] = 0 # update parameters grad_signs = [grad.sign() for grad in grads] torch._foreach_addcmul_(params, grad_signs, step_sizes, value=-1) for i in range(len(prevs)): prevs[i].copy_(grads[i])
def adagrad(params: List[Tensor], grads: List[Tensor], state_sums: List[Tensor], state_steps: List[int], has_sparse_grad: bool, *, lr: float, weight_decay: float, lr_decay: float, eps: float): r"""Functional API that performs Adagrad algorithm computation. See :class:`~torch.optim.Adagrad` for details. """ if weight_decay != 0: if has_sparse_grad: raise RuntimeError( "weight_decay option is not compatible with sparse gradients") torch._foreach_add_(grads, params, alpha=weight_decay) minus_clr = [-lr / (1 + (step - 1) * lr_decay) for step in state_steps] if has_sparse_grad: # sparse is not supported by multi_tensor. Fall back to optim.adagrad # implementation for sparse gradients for i, (param, grad, state_sum, step) in enumerate(zip(params, grads, state_sums, state_steps)): grad = grad.coalesce( ) # the update is non-linear so indices must be unique grad_indices = grad._indices() grad_values = grad._values() size = grad.size() state_sum.add_(_make_sparse(grad, grad_indices, grad_values.pow(2))) std_sparse = state_sum.sparse_mask(grad) std_sparse_values = std_sparse._values().sqrt_().add_(eps) param.add_( _make_sparse(grad, grad_indices, grad_values / std_sparse_values), alpha=minus_clr[i], ) else: grads = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in grads ] state_sums = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in state_sums ] torch._foreach_addcmul_(state_sums, grads, grads, value=1) std = torch._foreach_add(torch._foreach_sqrt(state_sums), eps) toAdd = torch._foreach_div(torch._foreach_mul(grads, minus_clr), std) toAdd = [ torch.view_as_complex(x) if torch.is_complex(params[i]) else x for i, x in enumerate(toAdd) ] torch._foreach_add_(params, toAdd) state_sums = [ torch.view_as_complex(x) if torch.is_complex(params[i]) else x for i, x in enumerate(state_sums) ]
def forward(ctx, scale, *param_list): """ In the forward pass we receive a Tensor containing the input and return a Tensor containing the output. ctx is a context object that can be used to stash information for backward computation. You can cache arbitrary objects for use in the backward pass using the ctx.save_for_backward method. """ ctx.save_for_backward(*param_list, scale.detach()) return torch._foreach_mul(param_list, float(scale))
def _multi_tensor_rprop(params: List[Tensor], grads: List[Tensor], prevs: List[Tensor], step_sizes: List[Tensor], *, step_size_min: float, step_size_max: float, etaminus: float, etaplus: float, maximize: bool): if len(params) == 0: return # Handle complex params def _view_complex_as_real(tensor_list): return [torch.view_as_real(t) if torch.is_complex(t) else t for t in tensor_list] grads = _view_complex_as_real(grads) prevs = _view_complex_as_real(prevs) params = _view_complex_as_real(params) step_sizes = _view_complex_as_real(step_sizes) if maximize: grads = torch._foreach_neg(grads) signs = torch._foreach_mul(grads, prevs) signs = [s.sign() for s in signs] for sign in signs: sign[sign.gt(0)] = etaplus sign[sign.lt(0)] = etaminus sign[sign.eq(0)] = 1 # update stepsizes with step size updates torch._foreach_mul_(step_sizes, signs) for step_size in step_sizes: step_size.clamp_(step_size_min, step_size_max) # for dir<0, dfdx=0 # for dir>=0 dfdx=dfdx grads = list(grads) for i in range(len(grads)): grads[i] = grads[i].clone(memory_format=torch.preserve_format) grads[i][signs[i].eq(etaminus)] = 0 # update parameters grad_signs = [grad.sign() for grad in grads] torch._foreach_addcmul_(params, grad_signs, step_sizes, value=-1) for i in range(len(prevs)): prevs[i].copy_(grads[i])
def _multi_tensor_adagrad(params: List[Tensor], grads: List[Tensor], state_sums: List[Tensor], state_steps: List[Tensor], *, lr: float, weight_decay: float, lr_decay: float, eps: float, has_sparse_grad: bool): # Foreach functions will throw errors if given empty lists if len(params) == 0: return if has_sparse_grad is None: has_sparse_grad = any([grad.is_sparse for grad in grads]) if has_sparse_grad: return _single_tensor_adagrad(params, grads, state_sums, state_steps, lr=lr, weight_decay=weight_decay, lr_decay=lr_decay, eps=eps, has_sparse_grad=has_sparse_grad) # Update steps torch._foreach_add_(state_steps, 1) if weight_decay != 0: torch._foreach_add_(grads, params, alpha=weight_decay) minus_clr = [-lr / (1 + (step - 1) * lr_decay) for step in state_steps] grads = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in grads ] state_sums = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in state_sums ] torch._foreach_addcmul_(state_sums, grads, grads, value=1) std = torch._foreach_add(torch._foreach_sqrt(state_sums), eps) toAdd = torch._foreach_div(torch._foreach_mul(grads, minus_clr), std) toAdd = [ torch.view_as_complex(x) if torch.is_complex(params[i]) else x for i, x in enumerate(toAdd) ] torch._foreach_add_(params, toAdd) state_sums = [ torch.view_as_complex(x) if torch.is_complex(params[i]) else x for i, x in enumerate(state_sums) ]
def multi_tensor_scale( self, src: Sequence[torch.Tensor], dst: Sequence[torch.Tensor], scale: float, ) -> None: with torch.no_grad(): # type: ignore[no-untyped-call] # _foreach_zero for long type is not supported in CUDA if self._enable_foreach and src[0].is_floating_point(): # scale val = torch._foreach_mul(tuple(src), scale) # copy tensor torch._foreach_zero_(tuple(dst)) torch._foreach_add_(tuple(dst), val) else: for s, d in zip(src, dst): d.copy_(s * scale)
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: with torch.enable_grad(): loss = closure() grads = [] states = [] params_with_grad = [] step_sizes = [] for group in self.param_groups: for p in group['params']: etaminus, etaplus = group['etas'] step_size_min, step_size_max = group['step_sizes'] if p.grad is not None: if p.grad.is_sparse: raise RuntimeError( 'RMSprop does not support sparse gradients') grads.append(p.grad) params_with_grad.append(p) state = self.state[p] # State initialization if len(state) == 0: state['step'] = 0 state['prev'] = torch.zeros_like( p, memory_format=torch.preserve_format) state['step_size'] = p.grad.new().resize_as_( p.grad).fill_(group['lr']) state['step'] += 1 states.append(state) step_sizes.append(state['step_size']) signs = torch._foreach_mul(grads, [s['prev'] for s in states]) signs = [s.sign() for s in signs] for sign in signs: sign[sign.gt(0)] = etaplus sign[sign.lt(0)] = etaminus sign[sign.eq(0)] = 1 # update stepsizes with step size updates torch._foreach_mul_(step_sizes, signs) for step_size in step_sizes: step_size.clamp_(step_size_min, step_size_max) # for dir<0, dfdx=0 # for dir>=0 dfdx=dfdx for i in range(len(grads)): grads[i] = grads[i].clone(memory_format=torch.preserve_format) grads[i][signs[i].eq(etaminus)] = 0 # update parameters grad_signs = [grad.sign() for grad in grads] torch._foreach_addcmul_(params_with_grad, grad_signs, step_sizes, value=-1) for i in range(len(states)): states[i]['prev'].copy_(grads[i]) return loss
def _multi_tensor_adamw(params: List[Tensor], grads: List[Tensor], exp_avgs: List[Tensor], exp_avg_sqs: List[Tensor], max_exp_avg_sqs: List[Tensor], state_steps: List[Tensor], *, amsgrad: bool, beta1: float, beta2: float, lr: float, weight_decay: float, eps: float, maximize: bool, capturable: bool): if len(params) == 0: return if capturable: assert all(p.is_cuda and step.is_cuda for p, step in zip(params, state_steps)), \ "If capturable=True, params and state_steps must be CUDA tensors." if maximize: grads = torch._foreach_neg(tuple(grads)) # type: ignore[assignment] grads = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in grads ] exp_avgs = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in exp_avgs ] exp_avg_sqs = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in exp_avg_sqs ] params = [ torch.view_as_real(x) if torch.is_complex(x) else x for x in params ] # update steps torch._foreach_add_(state_steps, 1) # Perform stepweight decay torch._foreach_mul_(params, 1 - lr * weight_decay) # Decay the first and second moment running average coefficient torch._foreach_mul_(exp_avgs, beta1) torch._foreach_add_(exp_avgs, grads, alpha=1 - beta1) torch._foreach_mul_(exp_avg_sqs, beta2) torch._foreach_addcmul_(exp_avg_sqs, grads, grads, 1 - beta2) if capturable: # TODO: use foreach_pow if/when foreach_pow is added bias_correction1 = [torch.pow(beta1, step) for step in state_steps] bias_correction2 = [torch.pow(beta2, step) for step in state_steps] # foreach_sub doesn't allow a scalar as the first arg torch._foreach_sub_(bias_correction1, 1) torch._foreach_sub_(bias_correction2, 1) torch._foreach_neg_(bias_correction1) torch._foreach_neg_(bias_correction2) # foreach_div doesn't allow a scalar as the first arg step_size = torch._foreach_div(bias_correction1, lr) torch._foreach_reciprocal_(step_size) torch._foreach_neg_(step_size) bias_correction2_sqrt = torch._foreach_sqrt(bias_correction2) if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch._foreach_maximum_(max_exp_avg_sqs, exp_avg_sqs) # Use the max. for normalizing running avg. of gradient max_exp_avg_sq_sqrt = torch._foreach_sqrt(max_exp_avg_sqs) # Folds in (admittedly ugly) 1-elem step_size math here to avoid extra param-set-sized read+write # (can't fold it into addcdiv_ below because addcdiv_ requires value is a Number, not a Tensor) torch._foreach_div_( max_exp_avg_sq_sqrt, torch._foreach_mul(bias_correction2_sqrt, step_size)) eps_over_step_size = torch._foreach_div(step_size, eps) torch._foreach_reciprocal_(eps_over_step_size) denom = torch._foreach_add(max_exp_avg_sq_sqrt, eps_over_step_size) else: exp_avg_sq_sqrt = torch._foreach_sqrt(exp_avg_sqs) torch._foreach_div_( exp_avg_sq_sqrt, torch._foreach_mul(bias_correction2_sqrt, step_size)) eps_over_step_size = torch._foreach_div(step_size, eps) torch._foreach_reciprocal_(eps_over_step_size) denom = torch._foreach_add(exp_avg_sq_sqrt, eps_over_step_size) torch._foreach_addcdiv_(params, exp_avgs, denom) else: bias_correction1 = [1 - beta1**step.item() for step in state_steps] bias_correction2 = [1 - beta2**step.item() for step in state_steps] step_size = [(lr / bc) * -1 for bc in bias_correction1] bias_correction2_sqrt = [math.sqrt(bc) for bc in bias_correction2] if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch._foreach_maximum_(max_exp_avg_sqs, exp_avg_sqs) # Use the max. for normalizing running avg. of gradient max_exp_avg_sq_sqrt = torch._foreach_sqrt(max_exp_avg_sqs) torch._foreach_div_(max_exp_avg_sq_sqrt, bias_correction2_sqrt) denom = torch._foreach_add(max_exp_avg_sq_sqrt, eps) else: exp_avg_sq_sqrt = torch._foreach_sqrt(exp_avg_sqs) torch._foreach_div_(exp_avg_sq_sqrt, bias_correction2_sqrt) denom = torch._foreach_add(exp_avg_sq_sqrt, eps) torch._foreach_addcdiv_(params, exp_avgs, denom, step_size)