FM / optim

/usr/local/lib64/python3.6/site-packages/torch/distributed/optim/

/usr/local/lib64/python3.6/site-packages/torch/distributed/optim

Name	Size	Mode	Actions
__pycache__/	-	0755	rm
functional_adadelta.py	3224	0644	edit dl rm
functional_adagrad.py	3634	0644	edit dl rm
functional_adam.py	6569	0644	edit dl rm
functional_adamax.py	4222	0644	edit dl rm
functional_adamw.py	6686	0644	edit dl rm
functional_rmsprop.py	3937	0644	edit dl rm
functional_rprop.py	3175	0644	edit dl rm
functional_sgd.py	4849	0644	edit dl rm
optimizer.py	9694	0644	edit dl rm
post_localSGD_optimizer.py	3402	0644	edit dl rm
zero_redundancy_optimizer.py	68601	0644	edit dl rm
__init__.py	1537	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/distributed/optim/functional_rmsprop.py (3937B)

from typing import List, Dict, Optional
import torch
import torch.optim._functional as F

from torch import Tensor

# Define a TorchScript compatible Functional RMSprop Optimizer
# where we use these optimizer in a functional way.
# Instead of using the `param.grad` when updating parameters,
# we explicitly allow the distributed optimizer pass gradients to
# the `step` function. In this way, we could separate the gradients
# and parameters and allow multithreaded trainer to update the
# parameters without data traces on accumulating to the same .grad.
# NOTE: This should be only used by distributed optimizer internals
# and not meant to expose to the user.
@torch.jit.script
class _FunctionalRMSprop(object):
    def __init__(
        self,
        params: List[Tensor],
        lr: float = 1e-2,
        alpha: float = 0.99,
        eps: float = 1e-8,
        weight_decay: float = 0.0,
        momentum: float = 0.0,
        centered: bool = False,
        _allow_empty_param_list: bool = False,
    ):
        self.defaults = {
            "lr": lr,
            "alpha": alpha,
            "eps": eps,
            "weight_decay": weight_decay,
            "momentum": momentum,
        }
        self.centered = centered

if len(params) == 0 and not _allow_empty_param_list:
            raise ValueError("optimizer got an empty parameter list")

# NOTE: we only have one param_group and don't allow user to add additional
        # param group as it's not a common use case.
        self.param_group = {"params": params}

self.state = torch.jit.annotate(Dict[torch.Tensor, Dict[str, torch.Tensor]], {})

def step(self, gradients: List[Optional[Tensor]]):
        params = self.param_group['params']
        params_with_grad = []
        grads = []
        square_avgs = []
        grad_avgs = []
        momentum_buffer_list = []
        lr = self.defaults['lr']
        alpha = self.defaults['alpha']
        eps = self.defaults['eps']
        momentum = self.defaults['momentum']
        weight_decay = self.defaults['weight_decay']

if len(params) != len(gradients):
            raise ValueError(
                "the gradients passed in does not equal to the size of the parameters!"
                + f"Params length: {len(params)}. "
                + f"Gradients length: {len(gradients)}"
            )

for param, gradient in zip(params, gradients):
            if gradient is not None:
                params_with_grad.append(param)
                grads.append(gradient)
                # Lazy state initialization
                if param not in self.state:
                    self.state[param] = {}
                    state = self.state[param]
                    state['step'] = torch.tensor(0.0)
                    state['square_avg'] = torch.zeros_like(param, memory_format=torch.preserve_format)
                    if momentum > 0:
                        state['momentum_buffer'] = torch.zeros_like(param, memory_format=torch.preserve_format)
                    if self.centered:
                        state['grad_avg'] = torch.zeros_like(param, memory_format=torch.preserve_format)

state = self.state[param]
                square_avgs.append(state['square_avg'])
                if momentum > 0:
                    momentum_buffer_list.append(state['momentum_buffer'])
                if self.centered:
                    grad_avgs.append(state['grad_avg'])

state['step'] += 1

with torch.no_grad():
            F.rmsprop(params_with_grad,
                      grads,
                      square_avgs,
                      grad_avgs,
                      momentum_buffer_list,
                      lr=lr,
                      alpha=alpha,
                      eps=eps,
                      weight_decay=weight_decay,
                      momentum=momentum,
                      centered=self.centered)

from typing import List, Dict, Optional import torch import torch.optim._functional as F from torch import Tensor # Define a TorchScript compatible Functional RMSprop Optimizer # where we use these optimizer in a functional way. # Instead of using the `param.grad` when updating parameters, # we explicitly allow the distributed optimizer pass gradients to # the `step` function. In this way, we could separate the gradients # and parameters and allow multithreaded trainer to update the # parameters without data traces on accumulating to the same .grad. # NOTE: This should be only used by distributed optimizer internals # and not meant to expose to the user. @torch.jit.script class _FunctionalRMSprop(object): def __init__( self, params: List[Tensor], lr: float = 1e-2, alpha: float = 0.99, eps: float = 1e-8, weight_decay: float = 0.0, momentum: float = 0.0, centered: bool = False, _allow_empty_param_list: bool = False, ): self.defaults = { "lr": lr, "alpha": alpha, "eps": eps, "weight_decay": weight_decay, "momentum": momentum, } self.centered = centered if len(params) == 0 and not _allow_empty_param_list: raise ValueError("optimizer got an empty parameter list") # NOTE: we only have one param_group and don't allow user to add additional # param group as it's not a common use case. self.param_group = {"params": params} self.state = torch.jit.annotate(Dict[torch.Tensor, Dict[str, torch.Tensor]], {}) def step(self, gradients: List[Optional[Tensor]]): params = self.param_group['params'] params_with_grad = [] grads = [] square_avgs = [] grad_avgs = [] momentum_buffer_list = [] lr = self.defaults['lr'] alpha = self.defaults['alpha'] eps = self.defaults['eps'] momentum = self.defaults['momentum'] weight_decay = self.defaults['weight_decay'] if len(params) != len(gradients): raise ValueError( "the gradients passed in does not equal to the size of the parameters!" + f"Params length: {len(params)}. " + f"Gradients length: {len(gradients)}" ) for param, gradient in zip(params, gradients): if gradient is not None: params_with_grad.append(param) grads.append(gradient) # Lazy state initialization if param not in self.state: self.state[param] = {} state = self.state[param] state['step'] = torch.tensor(0.0) state['square_avg'] = torch.zeros_like(param, memory_format=torch.preserve_format) if momentum > 0: state['momentum_buffer'] = torch.zeros_like(param, memory_format=torch.preserve_format) if self.centered: state['grad_avg'] = torch.zeros_like(param, memory_format=torch.preserve_format) state = self.state[param] square_avgs.append(state['square_avg']) if momentum > 0: momentum_buffer_list.append(state['momentum_buffer']) if self.centered: grad_avgs.append(state['grad_avg']) state['step'] += 1 with torch.no_grad(): F.rmsprop(params_with_grad, grads, square_avgs, grad_avgs, momentum_buffer_list, lr=lr, alpha=alpha, eps=eps, weight_decay=weight_decay, momentum=momentum, centered=self.centered)