optim_factory.py

import torch
from torch import optim as optim

from timm.optim.adafactor import Adafactor
from timm.optim.adahessian import Adahessian
from timm.optim.adamp import AdamP
from timm.optim.lookahead import Lookahead
from timm.optim.nadam import Nadam
from timm.optim.novograd import NovoGrad
from timm.optim.nvnovograd import NvNovoGrad
from timm.optim.radam import RAdam
from timm.optim.rmsprop_tf import RMSpropTF
from timm.optim.sgdp import SGDP

import json

try:
    from apex.optimizers import FusedNovoGrad, FusedAdam, FusedLAMB, FusedSGD
    has_apex = True
except ImportError:
    has_apex = False


def get_num_layer_for_vit(var_name, num_max_layer):
    if var_name in ("pos_embed", "temporal_embed"): ###!!!
        return 0
    elif var_name.startswith("visual.conv1"):
        return 0
    elif var_name.startswith("visual.ln_pre"):
        return 0
    elif var_name.startswith("visual.transformer.resblocks"):
        layer_id = int(var_name.split('.')[3]) #// 2
        return layer_id + 1
    else:
        return num_max_layer - 1

def get_num_layer_for_eva_vit(var_name, num_max_layer):
    if var_name in ("pos_embed", "temporal_embed"): ###!!!
        return 0
    elif var_name.startswith("visual.patch_embed"):
        return 0
    #elif var_name.startswith("visual.ln_pre"):
    #    return 0
    elif var_name.startswith("visual.blocks"):
        layer_id = int(var_name.split('.')[2]) #// 2
        return layer_id + 1
    else:
        return num_max_layer - 1
    
class LayerDecayValueAssigner(object):
    def __init__(self, values, vit_arch='clip'):
        self.values = values
        self.vit_arch = vit_arch

    def get_scale(self, layer_id):
        return self.values[layer_id]

    def get_layer_id(self, var_name):
        if self.vit_arch == 'clip':
            return get_num_layer_for_vit(var_name, len(self.values))
        elif self.vit_arch == 'eva_clip':
            return get_num_layer_for_eva_vit(var_name, len(self.values))

def get_parameter_groups(model, weight_decay=1e-5, skip_list=(), get_num_layer=None, get_layer_scale=None):
    parameter_group_names = {}
    parameter_group_vars = {}

    for name, param in model.named_parameters():
        if not param.requires_grad:
            continue  # frozen weights
        

        if len(param.shape) == 1 or name.endswith(".bias") or name in skip_list: ###!!!
            group_name = "no_decay"
            this_weight_decay = 0.
        elif 'control_point' in name:
            group_name = "control_point"
            this_weight_decay = weight_decay
            group_name = 'control_point'
        else:
            group_name = "decay"
            this_weight_decay = weight_decay
            
        if get_num_layer is not None:
            layer_id = get_num_layer(name)
            group_name = "layer_%d_%s" % (layer_id, group_name)
        else:
            layer_id = None

        if group_name not in parameter_group_names:
            if get_layer_scale is not None:
                if 'control_point' not in group_name and 'visual' in name: #for clip part
                    scale = get_layer_scale(layer_id) #0.001
                else:
                    scale = 1. #get_layer_scale(layer_id)
            else:
                scale = 1.

            parameter_group_names[group_name] = {
                "weight_decay": this_weight_decay,
                "params": [],
                "lr_scale": scale
            }
            parameter_group_vars[group_name] = {
                "weight_decay": this_weight_decay,
                "params": [],
                "lr_scale": scale
            }

        parameter_group_vars[group_name]["params"].append(param)
        parameter_group_names[group_name]["params"].append(name)
    print("Param groups = %s" % json.dumps(parameter_group_names, indent=2))
    return list(parameter_group_vars.values())


def create_optimizer(args, model, get_num_layer=None, get_layer_scale=None, filter_bias_and_bn=True, skip_list=None):
    opt_lower = args.opt.lower()
    weight_decay = args.weight_decay
    if weight_decay and filter_bias_and_bn:
        skip = {}
        if skip_list is not None:
            skip = skip_list
        elif hasattr(model, 'no_weight_decay'):
            skip = model.no_weight_decay()
        parameters = get_parameter_groups(model, weight_decay, skip, get_num_layer, get_layer_scale)
        weight_decay = 0.
    else:
        parameters = model.parameters()

    if 'fused' in opt_lower:
        assert has_apex and torch.cuda.is_available(), 'APEX and CUDA required for fused optimizers'

    opt_args = dict(lr=args.lr, weight_decay=weight_decay)
    if hasattr(args, 'opt_eps') and args.opt_eps is not None:
        opt_args['eps'] = args.opt_eps
    if hasattr(args, 'opt_betas') and args.opt_betas is not None:
        opt_args['betas'] = args.opt_betas

    print("optimizer settings:", opt_args)

    opt_split = opt_lower.split('_')
    opt_lower = opt_split[-1]
    if opt_lower == 'sgd' or opt_lower == 'nesterov':
        opt_args.pop('eps', None)
        optimizer = optim.SGD(parameters, momentum=args.momentum, nesterov=True, **opt_args)
    elif opt_lower == 'momentum':
        opt_args.pop('eps', None)
        optimizer = optim.SGD(parameters, momentum=args.momentum, nesterov=False, **opt_args)
    elif opt_lower == 'adam':
        optimizer = optim.Adam(parameters, **opt_args)
    elif opt_lower == 'adamw':
        optimizer = optim.AdamW(parameters, **opt_args)
    elif opt_lower == 'nadam':
        optimizer = Nadam(parameters, **opt_args)
    elif opt_lower == 'radam':
        optimizer = RAdam(parameters, **opt_args)
    elif opt_lower == 'adamp':
        optimizer = AdamP(parameters, wd_ratio=0.01, nesterov=True, **opt_args)
    elif opt_lower == 'sgdp':
        optimizer = SGDP(parameters, momentum=args.momentum, nesterov=True, **opt_args)
    elif opt_lower == 'adadelta':
        optimizer = optim.Adadelta(parameters, **opt_args)
    elif opt_lower == 'adafactor':
        if not args.lr:
            opt_args['lr'] = None
        optimizer = Adafactor(parameters, **opt_args)
    elif opt_lower == 'adahessian':
        optimizer = Adahessian(parameters, **opt_args)
    elif opt_lower == 'rmsprop':
        optimizer = optim.RMSprop(parameters, alpha=0.9, momentum=args.momentum, **opt_args)
    elif opt_lower == 'rmsproptf':
        optimizer = RMSpropTF(parameters, alpha=0.9, momentum=args.momentum, **opt_args)
    elif opt_lower == 'novograd':
        optimizer = NovoGrad(parameters, **opt_args)
    elif opt_lower == 'nvnovograd':
        optimizer = NvNovoGrad(parameters, **opt_args)
    elif opt_lower == 'fusedsgd':
        opt_args.pop('eps', None)
        optimizer = FusedSGD(parameters, momentum=args.momentum, nesterov=True, **opt_args)
    elif opt_lower == 'fusedmomentum':
        opt_args.pop('eps', None)
        optimizer = FusedSGD(parameters, momentum=args.momentum, nesterov=False, **opt_args)
    elif opt_lower == 'fusedadam':
        optimizer = FusedAdam(parameters, adam_w_mode=False, **opt_args)
    elif opt_lower == 'fusedadamw':
        optimizer = FusedAdam(parameters, adam_w_mode=True, **opt_args)
    elif opt_lower == 'fusedlamb':
        optimizer = FusedLAMB(parameters, **opt_args)
    elif opt_lower == 'fusednovograd':
        opt_args.setdefault('betas', (0.95, 0.98))
        optimizer = FusedNovoGrad(parameters, **opt_args)
    else:
        assert False and "Invalid optimizer"
        raise ValueError

    if len(opt_split) > 1:
        if opt_split[0] == 'lookahead':
            optimizer = Lookahead(optimizer)

    return optimizer