File size: 6,979 Bytes
2d9a728
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
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 ("cls_token", "mask_token", "pos_embed"):
        return 0
    elif var_name.startswith("patch_embed"):
        return 0
    elif var_name.startswith("rel_pos_bias"):
        return num_max_layer - 1
    elif var_name.startswith("blocks"):
        layer_id = int(var_name.split('.')[1])
        return layer_id + 1
    elif var_name.startswith("transformer.resblocks"):
        layer_id = int(var_name.split('.')[2])
        return layer_id + 1
    elif var_name in ("class_embedding", "positional_embedding", "temporal_positional_embedding"):
        return 0
    elif var_name.startswith("conv1"):
        return 0
    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_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.
        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(
        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