File size: 6,582 Bytes
aea73e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# UniRepLKNet
# https://github.com/AILab-CVC/UniRepLKNet
# Licensed under The Apache 2.0 License [see LICENSE for details]
# --------------------------------------------------------
import json
from mmcv.runner import OPTIMIZER_BUILDERS, DefaultOptimizerConstructor
from mmcv.runner import get_dist_info
from mmdet.utils import get_root_logger

def get_layer_id(var_name, max_layer_id,):
    """Get the layer id to set the different learning rates in ``layer_wise``
    decay_type.

    Args:
        var_name (str): The key of the model.
        max_layer_id (int): Maximum layer id.

    Returns:
        int: The id number corresponding to different learning rate in
        ``LearningRateDecayOptimizerConstructor``.
    """

    if var_name in ('backbone.cls_token', 'backbone.mask_token',
                    'backbone.pos_embed'):
        return 0

    elif var_name.startswith('backbone.downsample_layers'):
        stage_id = int(var_name.split('.')[2])
        if stage_id == 0:
            layer_id = 0
        elif stage_id == 1:
            layer_id = 2
        elif stage_id == 2:
            layer_id = 3
        elif stage_id == 3:
            layer_id = max_layer_id
        return layer_id

    elif var_name.startswith('backbone.stages'):
        stage_id = int(var_name.split('.')[2])
        block_id = int(var_name.split('.')[3])
        if stage_id == 0:
            layer_id = 1
        elif stage_id == 1:
            layer_id = 2
        elif stage_id == 2:
            layer_id = 3 + block_id // 3
        elif stage_id == 3:
            layer_id = max_layer_id
        return layer_id

    else:
        return max_layer_id + 1



def get_stage_id(var_name, max_stage_id):
    """Get the stage id to set the different learning rates in ``stage_wise``
    decay_type.

    Args:
        var_name (str): The key of the model.
        max_stage_id (int): Maximum stage id.

    Returns:
        int: The id number corresponding to different learning rate in
        ``LearningRateDecayOptimizerConstructor``.
    """

    if var_name in ('backbone.cls_token', 'backbone.mask_token',
                    'backbone.pos_embed'):
        return 0
    elif var_name.startswith('backbone.downsample_layers'):
        return 0
    elif var_name.startswith('backbone.stages'):
        stage_id = int(var_name.split('.')[2])
        return stage_id + 1
    else:
        return max_stage_id - 1


@OPTIMIZER_BUILDERS.register_module()
class UniRepLKNetLearningRateDecayOptimizerConstructor(DefaultOptimizerConstructor):
    # Different learning rates are set for different layers of backbone.
    # The design is inspired by and adapted from ConvNeXt.

    def add_params(self, params, module, **kwargs):
        """Add all parameters of module to the params list.

        The parameters of the given module will be added to the list of param
        groups, with specific rules defined by paramwise_cfg.

        Args:
            params (list[dict]): A list of param groups, it will be modified
                in place.
            module (nn.Module): The module to be added.
        """
        logger = get_root_logger()

        parameter_groups = {}
        logger.info(f'self.paramwise_cfg is {self.paramwise_cfg}')
        num_layers = self.paramwise_cfg.get('num_layers') + 2
        decay_rate = self.paramwise_cfg.get('decay_rate')
        decay_type = self.paramwise_cfg.get('decay_type', 'layer_wise')
        dw_scale = self.paramwise_cfg.get('dw_scale', 1)
        logger.info('Build UniRepLKNetLearningRateDecayOptimizerConstructor  '
                    f'{decay_type} {decay_rate} - {num_layers}')
        weight_decay = self.base_wd
        for name, param in module.named_parameters():
            if not param.requires_grad:
                continue  # frozen weights
            if len(param.shape) == 1 or name.endswith('.bias') or name in (
                    'pos_embed', 'cls_token'):
                group_name = 'no_decay'
                this_weight_decay = 0.
            else:
                group_name = 'decay'
                this_weight_decay = weight_decay
            if 'layer_wise' in decay_type:
                layer_id = get_layer_id(name, self.paramwise_cfg.get('num_layers'))
                logger.info(f'set param {name} as id {layer_id}')
            elif decay_type == 'stage_wise':
                layer_id = get_stage_id(name, num_layers)
                logger.info(f'set param {name} as id {layer_id}')

            if dw_scale == 1 or 'dwconv' not in name:
                group_name = f'layer_{layer_id}_{group_name}'
                if group_name not in parameter_groups:
                    scale = decay_rate ** (num_layers - layer_id - 1)
                    parameter_groups[group_name] = {
                        'weight_decay': this_weight_decay,
                        'params': [],
                        'param_names': [],
                        'lr_scale': scale,
                        'group_name': group_name,
                        'lr': scale * self.base_lr,
                    }

                parameter_groups[group_name]['params'].append(param)
                parameter_groups[group_name]['param_names'].append(name)
            else:
                group_name = f'layer_{layer_id}_{group_name}_dwconv'
                if group_name not in parameter_groups:
                    scale = decay_rate ** (num_layers - layer_id - 1) * dw_scale
                    parameter_groups[group_name] = {
                        'weight_decay': this_weight_decay,
                        'params': [],
                        'param_names': [],
                        'lr_scale': scale,
                        'group_name': group_name,
                        'lr': scale * self.base_lr,
                    }

                parameter_groups[group_name]['params'].append(param)
                parameter_groups[group_name]['param_names'].append(name)

        rank, _ = get_dist_info()
        if rank == 0:
            to_display = {}
            for key in parameter_groups:
                to_display[key] = {
                    'param_names': parameter_groups[key]['param_names'],
                    'lr_scale': parameter_groups[key]['lr_scale'],
                    'lr': parameter_groups[key]['lr'],
                    'weight_decay': parameter_groups[key]['weight_decay'],
                }
            logger.info(f'Param groups = {json.dumps(to_display, indent=2)}')
        params.extend(parameter_groups.values())