base_ml/optim_factory.py

# UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio, Video, Point Cloud, Time-Series and Image Recognition
# Github source: https://github.com/AILab-CVC/UniRepLKNet
# Licensed under The Apache License 2.0 License [see LICENSE for details]
# Based on RepLKNet, ConvNeXt, timm, DINO and DeiT code bases
# https://github.com/DingXiaoH/RepLKNet-pytorch
# https://github.com/facebookresearch/ConvNeXt
# https://github.com/rwightman/pytorch-image-models/tree/master/timm
# https://github.com/facebookresearch/deit/
# https://github.com/facebookresearch/dino
# --------------------------------------------------------'
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.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_convnext(var_name):
    """
    Divide [3, 3, 27, 3] layers into 12 groups; each group is three 
    consecutive blocks, including possible neighboring downsample layers;
    adapted from https://github.com/microsoft/unilm/blob/master/beit/optim_factory.py
    """
    num_max_layer = 12
    if var_name.startswith("downsample_layers"):
        stage_id = int(var_name.split('.')[1])
        if stage_id == 0:
            layer_id = 0
        elif stage_id == 1 or stage_id == 2:
            layer_id = stage_id + 1
        elif stage_id == 3:
            layer_id = 12
        return layer_id

    elif var_name.startswith("stages"):
        stage_id = int(var_name.split('.')[1])
        block_id = int(var_name.split('.')[2])
        if stage_id == 0 or stage_id == 1:
            layer_id = stage_id + 1
        elif stage_id == 2:
            layer_id = 3 + block_id // 3 
        elif stage_id == 3:
            layer_id = 12
        return layer_id
    else:
        return num_max_layer + 1

class LayerDecayValueAssigner(object):
    def __init__(self, values):
        self.values = values

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

    def get_layer_id(self, var_name):
        return get_num_layer_for_convnext(var_name)


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.
        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:
                scale = 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(model, weight_decay, lr, opt, get_num_layer=None, opt_eps=None, opt_betas=None, get_layer_scale=None, filter_bias_and_bn=True, skip_list=None, momentum = 0.9):
    opt_lower = opt.lower()
    weight_decay = weight_decay
    # if weight_decay and filter_bias_and_bn:
    if 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=lr, weight_decay=weight_decay)
    if opt_eps is not None:
        opt_args['eps'] = opt_eps
    if opt_betas is not None:
        opt_args['betas'] = opt_betas

    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=momentum, nesterov=True, **opt_args)
    elif opt_lower == 'momentum':
        opt_args.pop('eps', None)
        optimizer = optim.SGD(parameters, momentum=momentum, nesterov=False, **opt_args)
    elif opt_lower == 'adam':
        optimizer = optim.Adam(parameters, **opt_args)
    if 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=momentum, nesterov=True, **opt_args)
    elif opt_lower == 'adadelta':
        optimizer = optim.Adadelta(parameters, **opt_args)

    elif opt_lower == 'adahessian':
        optimizer = Adahessian(parameters, **opt_args)
    elif opt_lower == 'rmsprop':
        optimizer = optim.RMSprop(parameters, alpha=0.9, momentum=momentum, **opt_args)
    elif opt_lower == 'rmsproptf':
        optimizer = RMSpropTF(parameters, alpha=0.9, momentum=momentum, **opt_args)
    elif opt_lower == 'fusedsgd':
        opt_args.pop('eps', None)
        optimizer = FusedSGD(parameters, momentum=momentum, nesterov=True, **opt_args)
    elif opt_lower == 'fusedmomentum':
        opt_args.pop('eps', None)
        optimizer = FusedSGD(parameters, momentum=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"

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

    return optimizer