Shortcuts

Source code for ding.torch_utils.optimizer_helper

import torch
import math
from torch.nn.utils import clip_grad_norm_, clip_grad_value_
from typing import Union, Iterable, Tuple, Callable, List
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import pdb
import numpy as np
import copy
import random

inf = math.inf


[docs]def calculate_grad_norm(model: torch.nn.Module, norm_type=2) -> float: """ Overview: calculate grad norm of the parameters whose grad norms are not None in the model. Arguments: - model: torch.nn.Module - norm_type (:obj:`int` or `inf`) """ parameters = list(filter(lambda p: p.grad is not None, model.parameters())) if parameters == []: parameters = 0 return 0 if norm_type == 'inf': total_norm = max(p.grad.data.abs().max() for p in parameters) return float(total_norm) else: total_norm = 0 for p in parameters: param_norm = p.grad.data.norm(norm_type) total_norm += param_norm.item() ** norm_type total_norm = total_norm ** (1. / norm_type) return float(total_norm)
[docs]def calculate_grad_norm_without_bias_two_norm(model: torch.nn.Module) -> float: """ Overview: calculate grad norm of the parameters whose grad norms are not None in the model. Arguments: - model: torch.nn.Module """ _list = [] for name, param in model.named_parameters(): if 'bias' not in name and param.requires_grad: if param.grad is None: return 0 _list.append(param.grad.data.norm(2).item() ** 2) return float(sum(_list) ** (1. / 2))
[docs]def grad_ignore_norm(parameters, max_norm, norm_type=2): """ Overview: Clip the gradient norm of an iterable of parameters. Arguments: - parameters (:obj:`Iterable`): an iterable of torch.Tensor - max_norm (:obj:`float`): the max norm of the gradients - norm_type (:obj:`float`): 2.0 means use norm2 to clip """ if isinstance(parameters, torch.Tensor): parameters = [parameters] parameters = list(filter(lambda p: p.grad is not None, parameters)) max_norm = float(max_norm) norm_type = float(norm_type) if norm_type == inf: total_norm = max(p.grad.data.abs().max() for p in parameters) else: total_norm = 0 for p in parameters: param_norm = p.grad.data.norm(norm_type) total_norm += param_norm.item() ** norm_type total_norm = total_norm ** (1. / norm_type) clip_coef = max_norm / (total_norm + 1e-6) if clip_coef < 1: for p in parameters: p.grad.zero_() return total_norm
[docs]def grad_ignore_value(parameters, clip_value): """ Overview: Clip the gradient value of an iterable of parameters. Arguments: - parameters (:obj:`Iterable`): an iterable of torch.Tensor - clip_value (:obj:`float`): the value to start clipping """ if isinstance(parameters, torch.Tensor): parameters = [parameters] clip_value = float(clip_value) flag = False for p in filter(lambda p: p.grad is not None, parameters): val = p.grad.data.abs().max() if val >= clip_value: flag = True break if flag: for p in filter(lambda p: p.grad is not None, parameters): p.grad.data.zero_()
[docs]class Adam(torch.optim.Adam): """ Overview: Rewrited Adam optimizer to support more features. Interfaces: ``__init__``, ``step``, ``_state_init``, ``get_grad`` """
[docs] def __init__( self, params: Iterable, lr: float = 1e-3, betas: Tuple[float, float] = (0.9, 0.999), eps: float = 1e-8, weight_decay: float = 0, amsgrad: bool = False, optim_type: str = 'adam', grad_clip_type: str = None, clip_value: Union[float, None] = None, clip_coef: float = 5, clip_norm_type: float = 2.0, clip_momentum_timestep: int = 100, grad_norm_type: str = None, grad_ignore_type: str = None, ignore_value: Union[float, None] = None, ignore_coef: float = 5, ignore_norm_type: float = 2.0, ignore_momentum_timestep: int = 100, ): """ Overview: init method of refactored Adam class Arguments: - params (:obj:`iterable`): – an iterable of torch.Tensor s or dict s. \ Specifies what Tensors should be optimized - lr (:obj:`float`): learning rate, default set to 1e-3 - betas (:obj:`Tuple[float, float]`): coefficients used for computing running averages of gradient and its\ square, default set to (0.9, 0.999)) - eps (:obj:`float`): term added to the denominator to improve numerical stability, default set to 1e-8 - weight_decay (:obj:`float`): weight decay coefficient, deault set to 0 - amsgrad (:obj:`bool`): whether to use the AMSGrad variant of this algorithm from the paper\ On the Convergence of Adam and Beyond <https://arxiv.org/abs/1904.09237> - optim_type (:obj:str): support ["adam", "adamw"] - grad_clip_type (:obj:`str`): support [None, 'clip_momentum', 'clip_value', 'clip_norm', \ 'clip_momentum_norm'] - clip_value (:obj:`float`): the value to start clipping - clip_coef (:obj:`float`): the cliping coefficient - clip_norm_type (:obj:`float`): 2.0 means use norm2 to clip - clip_momentum_timestep (:obj:`int`): after how many step should we start the momentum clipping - grad_ignore_type (:obj:`str`): support [None, 'ignore_momentum', 'ignore_value', 'ignore_norm', \ 'ignore_momentum_norm'] - ignore_value (:obj:`float`): the value to start ignoring - ignore_coef (:obj:`float`): the ignoreing coefficient - ignore_norm_type (:obj:`float`): 2.0 means use norm2 to ignore - ignore_momentum_timestep (:obj:`int`): after how many step should we start the momentum ignoring """ self._support_type = { 'optim': ['adam', 'adamw'], 'grad_clip': [None, 'clip_momentum', 'clip_value', 'clip_norm', 'clip_momentum_norm'], 'grad_norm': [None], 'grad_ignore': [None, 'ignore_momentum', 'ignore_value', 'ignore_norm', 'ignore_momentum_norm'], } assert optim_type in self._support_type['optim'] assert grad_clip_type in self._support_type['grad_clip'] assert grad_norm_type in self._support_type['grad_norm'] assert grad_ignore_type in self._support_type['grad_ignore'] if grad_clip_type: assert clip_value is not None if grad_ignore_type: assert ignore_value is not None self._optim_type = optim_type self._grad_clip_type = grad_clip_type self._grad_norm_type = grad_norm_type self._grad_ignore_type = grad_ignore_type self._clip_value = clip_value self._clip_norm_type = clip_norm_type self._clip_coef = clip_coef self._ignore_value = ignore_value self._ignore_norm_type = ignore_norm_type self._ignore_coef = ignore_coef self._clip_momentum_timestep = clip_momentum_timestep self._ignore_momentum_timestep = ignore_momentum_timestep if self._optim_type == 'adamw': self._weight_decay = weight_decay super(Adam, self).__init__(params, lr=lr, betas=betas, eps=eps, weight_decay=0, amsgrad=amsgrad) elif self._optim_type == 'adam': super(Adam, self).__init__(params, lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, amsgrad=amsgrad) else: raise NotImplementedError( "optimizer type {} is not implemented, support type is {}".format( self._optim_type, self._support_type['optim'] ) )
[docs] def _state_init(self, p, amsgrad): """ Overview: Initialize the state of the optimizer Arguments: - p (:obj:`torch.Tensor`): the parameter to be optimized - amsgrad (:obj:`bool`): whether to use the AMSGrad variant of this algorithm from the paper\ On the Convergence of Adam and Beyond <https://arxiv.org/abs/1904.09237> """ state = self.state[p] state['thre_exp_avg_sq'] = torch.zeros_like(p.data, device=p.data.device) # others if torch.__version__ < "1.12.0": state['step'] = 0 # TODO # wait torch upgrad to 1.4, 1.3.1 didn't support memory format state['step'] = 0 else: state['step'] = torch.zeros((1,), dtype=torch.float, device=p.device) \ if self.defaults['capturable'] else torch.tensor(0.) state['exp_avg'] = torch.zeros_like(p.data) # Exponential moving average of squared gradient values state['exp_avg_sq'] = torch.zeros_like(p.data) if amsgrad: # Maintains max of all exp. moving avg. of sq. grad. values state['max_exp_avg_sq'] = torch.zeros_like(p.data)
[docs] def step(self, closure: Union[Callable, None] = None): """ Overview: Performs a single optimization step Arguments: - closure (:obj:`callable`): A closure that reevaluates the model and returns the loss, default set to None """ # clipping new_params = [ t for group in self.param_groups for t in group['params'] if t.requires_grad and t.grad is not None ] if self._grad_clip_type == 'clip_value': clip_grad_value_(new_params, self._clip_value) elif self._grad_clip_type == 'clip_norm': clip_grad_norm_(new_params, self._clip_value, self._clip_norm_type) elif self._grad_clip_type == 'clip_momentum': ''' This is the implimentation mimic the clip used in OPENAI, quote: 'Gradients are additionally clipped per parameter to be within between ±5√v where v is the running estimate of the second moment of the (unclipped) gradient' ''' for group in self.param_groups: for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['amsgrad']) grad = p.grad.data # should we use same beta group? beta1, beta2 = group['betas'] bias_correction2 = 1 - beta2 ** state['step'] state['thre_exp_avg_sq'].mul_(beta2).addcmul_(1 - beta2, grad, grad) if state['step'] >= self._clip_momentum_timestep: # initial value is inaccurate flag = grad.abs( ) > (state['thre_exp_avg_sq'].sqrt() / math.sqrt(bias_correction2)) * self._clip_coef grad.mul_(~flag).add_( ((state['thre_exp_avg_sq'].sqrt() / math.sqrt(bias_correction2)) * self._clip_coef).mul_(flag) ) elif self._grad_clip_type == 'clip_momentum_norm': # might have multi param_group, we should calculate each group differently. for group in self.param_groups: total_norm = 0 total_momentum_norm = 0 step = inf for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['amsgrad']) grad = p.grad.data # should we use same beta group? beta1, beta2 = group['betas'] bias_correction2 = 1 - beta2 ** state['step'] state['thre_exp_avg_sq'].mul_(beta2).addcmul_(1 - beta2, grad, grad) # sum total_norm param_norm = grad.norm(self._clip_norm_type) total_norm += param_norm.item() ** self._clip_norm_type # sum momentum_norm momentum = ((state['thre_exp_avg_sq'].sqrt() / math.sqrt(bias_correction2)) * self._clip_coef).norm(self._clip_norm_type) total_momentum_norm += momentum.item() ** self._clip_norm_type step = min(step, state['step']) if step > self._clip_momentum_timestep: total_norm = total_norm ** (1. / self._clip_norm_type) total_momentum_norm = total_momentum_norm ** (1. / self._clip_norm_type) clip_coef = total_momentum_norm / (total_norm + 1e-6) if clip_coef < 1: for p in group['params']: p.grad.data.mul_(clip_coef) if self._grad_ignore_type == 'ignore_value': grad_ignore_value(new_params, self._ignore_value) elif self._grad_ignore_type == 'ignore_norm': grad_ignore_norm(new_params, self._ignore_value, self._ignore_norm_type) elif self._grad_ignore_type == 'ignore_momentum': flag = False for group in self.param_groups: for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['amsgrad']) grad = p.grad.data # should we use same beta group? beta1, beta2 = group['betas'] bias_correction2 = 1 - beta2 ** state['step'] state['thre_exp_avg_sq'].mul_(beta2).addcmul_(1 - beta2, grad, grad) if state['step'] >= self._ignore_momentum_timestep: # initial value is inaccurate if grad.abs() > (state['thre_exp_avg_sq'].sqrt() / math.sqrt(bias_correction2)) * self._ignore_coef: flag = True break else: continue break if flag: for group in self.param_groups: for p in group['params']: if p.grad is None: continue p.grad.zero_() elif self._grad_ignore_type == 'ignore_momentum_norm': # might have multi param_group, we should calculate each group differently. step = inf for group in self.param_groups: total_norm = 0 total_momentum_norm = 0 for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['amsgrad']) grad = p.grad.data # should we use same beta group? beta1, beta2 = group['betas'] bias_correction2 = 1 - beta2 ** state['step'] state['thre_exp_avg_sq'].mul_(beta2).addcmul_(1 - beta2, grad, grad) # sum total_norm param_norm = grad.norm(self._ignore_norm_type) total_norm += param_norm.item() ** self._ignore_norm_type # sum momentum_norm momentum = ((state['thre_exp_avg_sq'].sqrt() / math.sqrt(bias_correction2)) * self._ignore_coef).norm(self._ignore_norm_type) total_momentum_norm += momentum.item() ** self._ignore_norm_type step = min(step, state['step']) if step > self._ignore_momentum_timestep: total_norm = total_norm ** (1. / self._ignore_norm_type) total_momentum_norm = total_momentum_norm ** (1. / self._ignore_norm_type) ignore_coef = total_momentum_norm / (total_norm + 1e-6) if ignore_coef < 1: for p in group['params']: p.grad.zero_() # Adam optim type if self._optim_type == 'adamw': for group in self.param_groups: for p in group['params']: if p.grad is None: continue p.data = p.data.add(-self._weight_decay * group['lr'], p.data) return super().step(closure=closure) elif self._optim_type == 'adam': return super().step(closure=closure)
[docs] def get_grad(self) -> float: total_norm = 0. params = [t for group in self.param_groups for t in group['params'] if t.requires_grad and t.grad is not None] for p in params: param_norm = p.grad.data.norm(self._clip_norm_type) total_norm += param_norm.item() ** self._clip_norm_type return total_norm
[docs]class RMSprop(torch.optim.RMSprop): r""" Overview: Rewrited RMSprop optimizer to support more features. Interfaces: ``__init__``, ``step``, ``_state_init``, ``get_grad`` """
[docs] def __init__( self, params: Iterable, lr: float = 1e-2, alpha: float = 0.99, eps: float = 1e-8, weight_decay: float = 0, momentum: float = 0, centered: bool = False, grad_clip_type: str = None, clip_value: Union[float, None] = None, clip_coef: float = 5, clip_norm_type: float = 2.0, clip_momentum_timestep: int = 100, grad_norm_type: str = None, grad_ignore_type: str = None, ignore_value: Union[float, None] = None, ignore_coef: float = 5, ignore_norm_type: float = 2.0, ignore_momentum_timestep: int = 100, ): """ Overview: init method of refactored Adam class Arguments: - params (:obj:`iterable`): – an iterable of torch.Tensor s or dict s. \ Specifies what Tensors should be optimized - lr (:obj:`float`): learning rate, default set to 1e-3 - alpha (:obj:`float`): smoothing constant, default set to 0.99 - eps (:obj:`float`): term added to the denominator to improve numerical stability, default set to 1e-8 - weight_decay (:obj:`float`): weight decay coefficient, deault set to 0 - centred (:obj:`bool`): if True, compute the centered RMSprop, \ the gradient is normalized by an estimation of its variance - grad_clip_type (:obj:`str`): support [None, 'clip_momentum', 'clip_value', 'clip_norm', \ 'clip_momentum_norm'] - clip_value (:obj:`float`): the value to start clipping - clip_coef (:obj:`float`): the cliping coefficient - clip_norm_type (:obj:`float`): 2.0 means use norm2 to clip - clip_momentum_timestep (:obj:`int`): after how many step should we start the momentum clipping - grad_ignore_type (:obj:`str`): support [None, 'ignore_momentum', 'ignore_value', 'ignore_norm', \ 'ignore_momentum_norm'] - ignore_value (:obj:`float`): the value to start ignoring - ignore_coef (:obj:`float`): the ignoreing coefficient - ignore_norm_type (:obj:`float`): 2.0 means use norm2 to ignore - ignore_momentum_timestep (:obj:`int`): after how many step should we start the momentum ignoring """ self._support_type = { 'grad_clip': [None, 'clip_momentum', 'clip_value', 'clip_norm', 'clip_momentum_norm'], 'grad_norm': [None], 'grad_ignore': [None, 'ignore_momentum', 'ignore_value', 'ignore_norm', 'ignore_momentum_norm'], } assert grad_clip_type in self._support_type['grad_clip'] assert grad_norm_type in self._support_type['grad_norm'] assert grad_ignore_type in self._support_type['grad_ignore'] if grad_clip_type: assert clip_value is not None if grad_ignore_type: assert ignore_value is not None self._grad_clip_type = grad_clip_type self._grad_norm_type = grad_norm_type self._grad_ignore_type = grad_ignore_type self._clip_value = clip_value self._clip_norm_type = clip_norm_type self._clip_coef = clip_coef self._ignore_value = ignore_value self._ignore_norm_type = ignore_norm_type self._ignore_coef = ignore_coef self._clip_momentum_timestep = clip_momentum_timestep self._ignore_momentum_timestep = ignore_momentum_timestep super(RMSprop, self).__init__( params, lr=lr, alpha=alpha, eps=eps, weight_decay=weight_decay, momentum=momentum, centered=centered )
[docs] def _state_init(self, p, momentum, centered): """ Overview: Initialize the state of the optimizer Arguments: - p (:obj:`torch.Tensor`): the parameter to be optimized - momentum (:obj:`float`): the momentum coefficient - centered (:obj:`bool`): if True, compute the centered RMSprop, \ the gradient is normalized by an estimation of its variance """ state = self.state[p] state['step'] = 0 state['thre_square_avg'] = torch.zeros_like(p.data, device=p.data.device) state['square_avg'] = torch.zeros_like(p.data, device=p.data.device) if momentum: state['momentum_buffer'] = torch.zeros_like(p.data, device=p.data.device) if centered: state['grad_avg'] = torch.zeros_like(p.data, device=p.data.device)
[docs] def step(self, closure: Union[Callable, None] = None): """ Overview: Performs a single optimization step Arguments: - closure (:obj:`callable`): A closure that reevaluates the model and returns the loss, default set to None """ # clipping new_params = [ t for group in self.param_groups for t in group['params'] if t.requires_grad and t.grad is not None ] if self._grad_clip_type == 'clip_value': clip_grad_value_(new_params, self._clip_value) elif self._grad_clip_type == 'clip_norm': clip_grad_norm_(new_params, self._clip_value, self._clip_norm_type) elif self._grad_clip_type == 'clip_momentum': ''' This implementation mimics the clip used in OPENAI, quote: 'Gradients are additionally clipped per parameter to be within between ±5√v where v is the running estimate of the second moment of the (unclipped) gradient' ''' for group in self.param_groups: for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['momentum'], group['centered']) grad = p.grad.data # beta1, beta2 = group['betas'] alpha = group['alpha'] state['thre_square_avg'].mul_(alpha).addcmul_(1 - alpha, grad, grad) if state['step'] >= self._clip_momentum_timestep: # initial value is inaccurate flag = grad.abs() > state['thre_square_avg'].sqrt() * self._clip_coef grad.mul_(~flag).add_((state['thre_square_avg'].sqrt() * self._clip_coef).mul_(flag)) elif self._grad_clip_type == 'clip_momentum_norm': # might have multi param_group, we should calculate each group differently. for group in self.param_groups: total_norm = 0 total_momentum_norm = 0 step = inf for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['momentum'], group['centered']) grad = p.grad.data alpha = group['alpha'] state['thre_square_avg'].mul_(alpha).addcmul_(1 - alpha, grad, grad) # sum total_norm param_norm = grad.norm(self._clip_norm_type) total_norm += param_norm.item() ** self._clip_norm_type # sum momentum_norm momentum = (state['thre_square_avg'].sqrt() * self._clip_coef).norm(self._clip_norm_type) total_momentum_norm += momentum.item() ** self._clip_norm_type step = min(step, state['step']) if step > self._clip_momentum_timestep: total_norm = total_norm ** (1. / self._clip_norm_type) total_momentum_norm = total_momentum_norm ** (1. / self._clip_norm_type) clip_coef = total_momentum_norm / (total_norm + 1e-6) if clip_coef < 1: for p in group['params']: p.grad.data.mul_(clip_coef) if self._grad_ignore_type == 'ignore_value': grad_ignore_value(new_params, self._ignore_value) elif self._grad_ignore_type == 'ignore_norm': grad_ignore_norm(new_params, self._ignore_value, self._ignore_norm_type) elif self._grad_ignore_type == 'ignore_momentum': flag = False for group in self.param_groups: for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['momentum'], group['centered']) grad = p.grad.data alpha = group['alpha'] state['thre_square_avg'].mul_(alpha).addcmul_(1 - alpha, grad, grad) if state['step'] >= self._ignore_momentum_timestep: # initial value is inaccurate if grad.abs() > state['thre_square_avg'].sqrt() * self._ignore_coef: flag = True break else: continue break if flag: for group in self.param_groups: for p in group['params']: if p.grad is None: continue p.grad.zero_() elif self._grad_ignore_type == 'ignore_momentum_norm': # might have multi param_group, we should calculate each group differently. step = inf for group in self.param_groups: total_norm = 0 total_momentum_norm = 0 for p in group['params']: if p.grad is None: continue state = self.state[p] if len(state) == 0: self._state_init(p, group['momentum'], group['centered']) grad = p.grad.data alpha = group['alpha'] state['thre_square_avg'].mul_(alpha).addcmul_(1 - alpha, grad, grad) # sum total_norm param_norm = grad.norm(self._ignore_norm_type) total_norm += param_norm.item() ** self._ignore_norm_type # sum momentum_norm momentum = (state['thre_square_avg'].sqrt() * self._ignore_coef).norm(self._ignore_norm_type) total_momentum_norm += momentum.item() ** self._ignore_norm_type step = min(step, state['step']) if step > self._ignore_momentum_timestep: total_norm = total_norm ** (1. / self._ignore_norm_type) total_momentum_norm = total_momentum_norm ** (1. / self._ignore_norm_type) ignore_coef = total_momentum_norm / (total_norm + 1e-6) if ignore_coef < 1: for p in group['params']: p.grad.zero_() return super().step(closure=closure)
[docs] def get_grad(self) -> float: """ Overview: calculate grad norm of the parameters whose grad norms are not None in the model. """ total_norm = 0. params = [t for group in self.param_groups for t in group['params'] if t.requires_grad and t.grad is not None] for p in params: param_norm = p.grad.data.norm(self._clip_norm_type) total_norm += param_norm.item() ** self._clip_norm_type return total_norm
[docs]class PCGrad(): """ Overview: PCGrad optimizer to support multi-task. you can view the paper in the following link https://arxiv.org/pdf/2001.06782.pdf Interfaces: ``__init__``, ``zero_grad``, ``step``, ``pc_backward`` Properties: - optimizer (:obj:`torch.optim`): the optimizer to be used """
[docs] def __init__(self, optimizer, reduction='mean'): """ Overview: Initialization of PCGrad optimizer Arguments: - optimizer (:obj:`torch.optim`): the optimizer to be used - reduction (:obj:`str`): the reduction method, support ['mean', 'sum'] """ self._optim, self._reduction = optimizer, reduction
@property def optimizer(self): """ Overview: get the optimizer """ return self._optim
[docs] def zero_grad(self): """ Overview: clear the gradient of the parameters """ return self._optim.zero_grad(set_to_none=True)
[docs] def step(self): """ Overview: update the parameters with the gradient """ return self._optim.step()
[docs] def pc_backward(self, objectives): """ Overview: calculate the gradient of the parameters Arguments: - objectives: a list of objectives """ grads, shapes, has_grads = self._pack_grad(objectives) pc_grad = self._project_conflicting(grads, has_grads) pc_grad = self._unflatten_grad(pc_grad, shapes[0]) self._set_grad(pc_grad) return
[docs] def _project_conflicting(self, grads, has_grads, shapes=None): """ Overview: project the conflicting gradient to the orthogonal space Arguments: - grads (:obj:`list`): a list of the gradient of the parameters - has_grads (:obj:`list`): a list of mask represent whether the parameter has gradient - shapes (:obj:`list`): a list of the shape of the parameters """ shared = torch.stack(has_grads).prod(0).bool() pc_grad, num_task = copy.deepcopy(grads), len(grads) for g_i in pc_grad: random.shuffle(grads) for g_j in grads: g_i_g_j = torch.dot(g_i, g_j) if g_i_g_j < 0: g_i -= (g_i_g_j) * g_j / (g_j.norm() ** 2) merged_grad = torch.zeros_like(grads[0]).to(grads[0].device) if self._reduction: merged_grad[shared] = torch.stack([g[shared] for g in pc_grad]).mean(dim=0) elif self._reduction == 'sum': merged_grad[shared] = torch.stack([g[shared] for g in pc_grad]).sum(dim=0) else: raise KeyError("invalid reduction method") merged_grad[~shared] = torch.stack([g[~shared] for g in pc_grad]).sum(dim=0) return merged_grad
[docs] def _set_grad(self, grads): """ Overview: set the modified gradients to the network Arguments: - grads (:obj:`list`): a list of the gradient of the parameters """ idx = 0 for group in self._optim.param_groups: for p in group['params']: # if p.grad is None: continue p.grad = grads[idx] idx += 1 return
[docs] def _pack_grad(self, objectives): """ Overview: pack the gradient of the parameters of the network for each objective Arguments: - objectives: a list of objectives Returns: - grad: a list of the gradient of the parameters - shape: a list of the shape of the parameters - has_grad: a list of mask represent whether the parameter has gradient """ grads, shapes, has_grads = [], [], [] for obj in objectives: self._optim.zero_grad(set_to_none=True) obj.backward(retain_graph=True) grad, shape, has_grad = self._retrieve_grad() grads.append(self._flatten_grad(grad, shape)) has_grads.append(self._flatten_grad(has_grad, shape)) shapes.append(shape) return grads, shapes, has_grads
[docs] def _unflatten_grad(self, grads, shapes): """ Overview: unflatten the gradient of the parameters of the network Arguments: - grads (:obj:`list`): a list of the gradient of the parameters - shapes (:obj:`list`): a list of the shape of the parameters """ unflatten_grad, idx = [], 0 for shape in shapes: length = np.prod(shape) unflatten_grad.append(grads[idx:idx + length].view(shape).clone()) idx += length return unflatten_grad
[docs] def _flatten_grad(self, grads, shapes): """ Overview: flatten the gradient of the parameters of the network Arguments: - grads (:obj:`list`): a list of the gradient of the parameters - shapes (:obj:`list`): a list of the shape of the parameters """ flatten_grad = torch.cat([g.flatten() for g in grads]) return flatten_grad
[docs] def _retrieve_grad(self): """ Overview: get the gradient of the parameters of the network with specific objective Returns: - grad: a list of the gradient of the parameters - shape: a list of the shape of the parameters - has_grad: a list of mask represent whether the parameter has gradient """ grad, shape, has_grad = [], [], [] for group in self._optim.param_groups: for p in group['params']: # if p.grad is None: continue # tackle the multi-head scenario if p.grad is None: shape.append(p.shape) grad.append(torch.zeros_like(p).to(p.device)) has_grad.append(torch.zeros_like(p).to(p.device)) continue shape.append(p.grad.shape) grad.append(p.grad.clone()) has_grad.append(torch.ones_like(p).to(p.device)) return grad, shape, has_grad
[docs]def configure_weight_decay(model: nn.Module, weight_decay: float) -> List: """ Overview: Separating out all parameters of the model into two buckets: those that will experience weight decay for regularization and those that won't (biases, and layer-norm or embedding weights). Arguments: - model (:obj:`nn.Module`): The given PyTorch model. - weight_decay (:obj:`float`): Weight decay value for optimizer. Returns: - optim groups (:obj:`List`): The parameter groups to be set in the latter optimizer. """ decay = set() no_decay = set() whitelist_weight_modules = (torch.nn.Linear, ) blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding) for mn, m in model.named_modules(): for pn, p in m.named_parameters(): fpn = '%s.%s' % (mn, pn) if mn else pn # full param name # Because named_modules and named_parameters are recursive # we will see the same tensors p many times. But doing it this way # allows us to know which parent module any tensor p belongs to. if pn.endswith('bias'): # all biases will not be decayed no_decay.add(fpn) elif pn.endswith('weight') and isinstance(m, whitelist_weight_modules): # weights of whitelist modules will be weight decayed decay.add(fpn) elif pn.endswith('weight') and isinstance(m, blacklist_weight_modules): # weights of blacklist modules will NOT be weight decayed no_decay.add(fpn) else: decay.add(fpn) decay = decay - no_decay # validate that we considered every parameter param_dict = {pn: p for pn, p in model.named_parameters()} union_params = decay | no_decay assert len( param_dict.keys() - union_params) == 0, "parameters %s were not separated into either decay/no_decay set!" \ % (str(param_dict.keys() - union_params),) optim_groups = [ { "params": [param_dict[pn] for pn in sorted(list(decay))], "weight_decay": weight_decay }, { "params": [param_dict[pn] for pn in sorted(list(no_decay))], "weight_decay": 0.0 }, ] return optim_groups