File size: 7,521 Bytes
a7dedf9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
from torch import nn, Tensor
import open_clip
from peft import get_peft_model, LoraConfig

from ..utils import ConvRefine, ConvUpsample, ConvAdapter
from ..utils import _get_norm_layer, _get_activation


resnet_names_and_weights = {
    "RN50": ["openai", "yfcc15m", "cc12m"],
    "RN101": ["openai", "yfcc15m", "cc12m"],
    "RN50x4": ["openai", "yfcc15m", "cc12m"],
    "RN50x16": ["openai", "yfcc15m", "cc12m"],
    "RN50x64": ["openai", "yfcc15m", "cc12m"],
}

refiner_channels = {
    "RN50": 2048,
    "RN101": 2048,
    "RN50x4": 2560,
    "RN50x16": 3072,
    "RN50x64": 4096,
}

refiner_groups = {
    "RN50": refiner_channels["RN50"] // 512,  # 4
    "RN101": refiner_channels["RN101"] // 512, # 4
    "RN50x4": refiner_channels["RN50x4"] // 512, # 5
    "RN50x16": refiner_channels["RN50x16"] // 512, # 6
    "RN50x64": refiner_channels["RN50x64"] // 512, # 8
}


class ResNet(nn.Module):
    def __init__(
        self,
        model_name: str,
        weight_name: str,
        block_size: int = 16,
        adapter: bool = False,
        adapter_reduction: int = 4,
        norm: str = "none",
        act: str = "none"
    ) -> None:
        super(ResNet, self).__init__()
        assert model_name in resnet_names_and_weights, f"Model name should be one of {list(resnet_names_and_weights.keys())}, but got {model_name}."
        assert weight_name in resnet_names_and_weights[model_name], f"Pretrained should be one of {resnet_names_and_weights[model_name]}, but got {weight_name}."
        assert block_size in [32, 16, 8], f"block_size should be one of [32, 16, 8], got {block_size}"
        self.model_name, self.weight_name = model_name, weight_name
        self.block_size = block_size

        model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual

        self.adapter = adapter
        if adapter:
            for param in model.parameters():
                param.requires_grad = False
        
        # Stem
        self.conv1 = model.conv1
        self.bn1 = model.bn1
        self.act1 = model.act1
        self.conv2 = model.conv2
        self.bn2 = model.bn2
        self.act2 = model.act2
        self.conv3 = model.conv3
        self.bn3 = model.bn3
        self.act3 = model.act3
        self.avgpool = model.avgpool
        # Stem: reduction = 4

        # Layers
        for idx in range(1, 5):
            setattr(self, f"layer{idx}", getattr(model, f"layer{idx}"))
            if adapter:
                setattr(self, f"adapter{idx}", ConvAdapter(
                    in_channels=getattr(model, f"layer{idx}")[-1].conv3.out_channels,
                    bottleneck_channels=getattr(model, f"layer{idx}")[-1].conv3.out_channels // adapter_reduction,
                ) if idx < 4 else nn.Identity())  # No adapter for the last layer

        self.in_features = model.attnpool.c_proj.weight.shape[1]
        self.out_features = model.attnpool.c_proj.weight.shape[0]

        if norm == "bn":
            norm_layer = nn.BatchNorm2d
        elif norm == "ln":
            norm_layer = nn.LayerNorm
        else:
            norm_layer = _get_norm_layer(model)

        if act == "relu":
            activation = nn.ReLU(inplace=True)
        elif act == "gelu":
            activation = nn.GELU()
        else:
            activation = _get_activation(model)

        if block_size == 32:
            self.refiner = ConvRefine(
                in_channels=self.in_features,
                out_channels=self.in_features,
                norm_layer=norm_layer,
                activation=activation,
                groups=refiner_groups[self.model_name],
            )
        elif block_size == 16:
            self.refiner = ConvUpsample(
                in_channels=self.in_features,
                out_channels=self.in_features,
                norm_layer=norm_layer,
                activation=activation,
                groups=refiner_groups[self.model_name],
            )
        else:  # block_size == 8
            self.refiner = nn.Sequential(
                ConvUpsample(
                    in_channels=self.in_features,
                    out_channels=self.in_features,
                    norm_layer=norm_layer,
                    activation=activation,
                    groups=refiner_groups[self.model_name],
                ),
                ConvUpsample(
                    in_channels=self.in_features,
                    out_channels=self.in_features,
                    norm_layer=norm_layer,
                    activation=activation,
                    groups=refiner_groups[self.model_name],
                ),
            )
    
    def train(self, mode: bool = True):
        if self.adapter and mode:
            # training:
            self.conv1.eval()
            self.bn1.eval()
            self.act1.eval()
            self.conv2.eval()
            self.bn2.eval()
            self.act2.eval()
            self.conv3.eval()
            self.bn3.eval()
            self.act3.eval()
            self.avgpool.eval()

            for idx in range(1, 5):
                getattr(self, f"layer{idx}").eval()
                getattr(self, f"adapter{idx}").train()

            self.refiner.train()

        else:
            # evaluation:
            for module in self.children():
                module.train(mode)

    def stem(self, x: Tensor) -> Tensor:
        x = self.act1(self.bn1(self.conv1(x)))
        x = self.act2(self.bn2(self.conv2(x)))
        x = self.act3(self.bn3(self.conv3(x)))
        x = self.avgpool(x)
        return x

    def forward(self, x: Tensor) -> Tensor:
        x = self.stem(x)
        
        x = self.layer1(x)
        if self.adapter:
            x = self.adapter1(x)
        
        x = self.layer2(x)
        if self.adapter:
            x = self.adapter2(x)

        x = self.layer3(x)
        if self.adapter:
            x = self.adapter3(x)
        
        x = self.layer4(x)
        if self.adapter:
            x = self.adapter4(x)
    
        x = self.refiner(x)
        return x


def _resnet(
    model_name: str,
    weight_name: str,
    block_size: int = 16,
    adapter: bool = False,
    adapter_reduction: int = 4,
    lora: bool = False,
    lora_rank: int = 16,
    lora_alpha: float = 32.0,
    lora_dropout: float = 0.1,
    norm: str = "none",
    act: str = "none"
) -> ResNet:
    assert not (lora and adapter), "Lora and adapter cannot be used together."
    model = ResNet(
        model_name=model_name,
        weight_name=weight_name,
        block_size=block_size,
        adapter=adapter,
        adapter_reduction=adapter_reduction,
        norm=norm,
        act=act
    )

    if lora:
        target_modules = []
        for name, module in model.named_modules():
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                target_modules.append(name)
        
        lora_config = LoraConfig(
            r=lora_rank,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            bias="none",
            target_modules=target_modules,
        )
        model = get_peft_model(model, lora_config)

        # Unfreeze BN layers
        for name, module in model.named_modules():
            if isinstance(module, nn.BatchNorm2d) and "refiner" not in name:
                module.requires_grad_(True)

        # Unfreeze refiner
        for name, module in model.named_modules():
            if "refiner" in name:
                module.requires_grad_(True)
    
    return model