File size: 13,763 Bytes
c29babb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
import math

import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as T
from PIL import Image
from transformers import CLIPModel


class EffortModel(nn.Module):
    def __init__(self, config=None):
        super().__init__()
        self.config = config
        self.backbone = self.build_backbone(config)
        self.head = nn.Linear(1024, 2)
        self.loss_func = nn.CrossEntropyLoss()
        self.prob, self.label = [], []
        self.correct, self.total = 0, 0

    def build_backbone(self, config):
        # ⚠⚠⚠ Download CLIP model using the below link
        # https://drive.google.com/drive/folders/1fm3Jd8lFMiSP1qgdmsxfqlJZGpr_bXsx?usp=drive_link

        # mean: [0.48145466, 0.4578275, 0.40821073]
        # std: [0.26862954, 0.26130258, 0.27577711]

        # ViT-L/14 224*224
        # the path of this folder in your disk (download from the above link)
        clip_model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")

        # Apply SVD to self_attn layers only
        # ViT-L/14 224*224: 1024-1
        clip_model.vision_model = apply_svd_residual_to_self_attn(clip_model.vision_model, r=1024 - 1)

        # for name, param in clip_model.vision_model.named_parameters():
        #    print('{}: {}'.format(name, param.requires_grad))
        # num_param = sum(p.numel() for p in clip_model.vision_model.parameters() if p.requires_grad)
        # num_total_param = sum(p.numel() for p in clip_model.vision_model.parameters())
        # print('Number of total parameters: {}, tunable parameters: {}'.format(num_total_param, num_param))

        return clip_model.vision_model

    def features(self, inputs: torch.Tensor) -> torch.tensor:
        return self.backbone(inputs).pooler_output

    def classifier(self, features: torch.tensor) -> torch.tensor:
        return self.head(features)

    # def get_losses(self, data_dict: dict, pred_dict: dict) -> dict:
    #    label = data_dict['label']
    #    pred = pred_dict['cls']
    #    loss = self.loss_func(pred, label)
    #
    #    if self.training:
    #        # Regularization term
    #        lambda_reg = 1.0
    #        reg_term = 0.0
    #        num_reg = 0
    #        for module in self.backbone.modules():
    #            if isinstance(module, SVDResidualLinear):
    #                reg_term += module.compute_orthogonal_loss()
    #                reg_term += module.compute_keepsv_loss()
    #                num_reg += 1
    #
    #        loss += lambda_reg * reg_term / num_reg
    #
    #    loss_dict = {'overall': loss}
    #    return loss_dict

    def compute_weight_loss(self):
        weight_sum_dict = {}
        num_weight_dict = {}
        for name, module in self.backbone.named_modules():
            if isinstance(module, SVDResidualLinear):
                weight_curr = module.compute_current_weight()
                if str(weight_curr.size()) not in weight_sum_dict.keys():
                    weight_sum_dict[str(weight_curr.size())] = weight_curr
                    num_weight_dict[str(weight_curr.size())] = 1
                else:
                    weight_sum_dict[str(weight_curr.size())] += weight_curr
                    num_weight_dict[str(weight_curr.size())] += 1

        loss2 = 0.0
        for k in weight_sum_dict.keys():
            _, S_sum, _ = torch.linalg.svd(weight_sum_dict[k], full_matrices=False)
            loss2 += -torch.mean(S_sum)
        loss2 /= len(weight_sum_dict.keys())
        return loss2

    def get_losses(self, data_dict: dict, pred_dict: dict) -> dict:
        label = data_dict["label"]  # Tensor of shape [batch_size]
        pred = pred_dict["logits"]  # Tensor of shape [batch_size, num_classes]

        # Compute overall loss using all samples
        loss = self.loss_func(pred, label)

        # Create masks for real and fake classes
        mask_real = label == 0  # Boolean tensor
        mask_fake = label == 1  # Boolean tensor

        # Compute loss for real class
        if mask_real.sum() > 0:
            pred_real = pred[mask_real]
            label_real = label[mask_real]
            loss_real = self.loss_func(pred_real, label_real)
        else:
            # No real samples in batch
            loss_real = torch.tensor(0.0, device=pred.device)

        # Compute loss for fake class
        if mask_fake.sum() > 0:
            pred_fake = pred[mask_fake]
            label_fake = label[mask_fake]
            loss_fake = self.loss_func(pred_fake, label_fake)
        else:
            # No fake samples in batch
            loss_fake = torch.tensor(0.0, device=pred.device)

        # loss2 = self.compute_weight_loss()
        # overall_loss = loss + loss2

        # Return a dictionary with all losses
        loss_dict = {
            "overall": loss,
            "real_loss": loss_real,
            "fake_loss": loss_fake,
            # 'erank_loss': loss2
        }
        return loss_dict

    # def get_train_metrics(self, data_dict: dict, pred_dict: dict) -> dict:
    #     label = data_dict["label"]
    #     pred = pred_dict["logits"]
    #     # compute metrics for batch data
    #     auc, eer, acc, ap = calculate_metrics_for_train(label.detach(), pred.detach())
    #     metric_batch_dict = {"acc": acc, "auc": auc, "eer": eer, "ap": ap}
    #     return metric_batch_dict

    def forward(self, inputs: torch.Tensor):
        # Get features from the backbone
        features = self.features(inputs)

        # Get logits from the classifier
        logits = self.classifier(features)

        normalized_features = F.normalize(features, p=2, dim=1)

        return logits, normalized_features


# Custom module to represent the residual using SVD components
class SVDResidualLinear(nn.Module):
    def __init__(self, in_features, out_features, r, bias=True, init_weight=None):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.r = r  # Number of top singular values to exclude

        # Original weights (fixed)
        self.weight_main = nn.Parameter(torch.Tensor(out_features, in_features), requires_grad=False)
        if init_weight is not None:
            self.weight_main.data.copy_(init_weight)
        else:
            nn.init.kaiming_uniform_(self.weight_main, a=math.sqrt(5))

        # Bias
        if bias:
            self.bias = nn.Parameter(torch.Tensor(out_features))
            nn.init.zeros_(self.bias)
        else:
            self.register_parameter("bias", None)

    def compute_current_weight(self):
        if self.S_residual is not None:
            return self.weight_main + self.U_residual @ torch.diag(self.S_residual) @ self.V_residual
        else:
            return self.weight_main

    def forward(self, x):
        if hasattr(self, "U_residual") and hasattr(self, "V_residual") and self.S_residual is not None:
            # Reconstruct the residual weight
            residual_weight = self.U_residual @ torch.diag(self.S_residual) @ self.V_residual
            # Total weight is the fixed main weight plus the residual
            weight = self.weight_main + residual_weight
        else:
            # If residual components are not set, use only the main weight
            weight = self.weight_main

        return F.linear(x, weight, self.bias)

    def compute_orthogonal_loss(self):
        if self.S_residual is not None:
            # According to the properties of orthogonal matrices: A^TA = I
            UUT = torch.cat((self.U_r, self.U_residual), dim=1) @ torch.cat((self.U_r, self.U_residual), dim=1).t()
            VVT = torch.cat((self.V_r, self.V_residual), dim=0) @ torch.cat((self.V_r, self.V_residual), dim=0).t()
            # print(self.U_r.size(), self.U_residual.size())  # torch.Size([1024, 1023]) torch.Size([1024, 1])
            # print(self.V_r.size(), self.V_residual.size())  # torch.Size([1023, 1024]) torch.Size([1, 1024])
            # UUT = self.U_residual @ self.U_residual.t()
            # VVT = self.V_residual @ self.V_residual.t()

            # Construct an identity matrix
            UUT_identity = torch.eye(UUT.size(0), device=UUT.device)
            VVT_identity = torch.eye(VVT.size(0), device=VVT.device)

            # Using frobenius norm to compute loss
            loss = 0.5 * torch.norm(UUT - UUT_identity, p="fro") + 0.5 * torch.norm(VVT - VVT_identity, p="fro")
        else:
            loss = 0.0

        return loss

    def compute_keepsv_loss(self):
        if (self.S_residual is not None) and (self.weight_original_fnorm is not None):
            # Total current weight is the fixed main weight plus the residual
            weight_current = self.weight_main + self.U_residual @ torch.diag(self.S_residual) @ self.V_residual
            # Frobenius norm of current weight
            weight_current_fnorm = torch.norm(weight_current, p="fro")

            loss = torch.abs(weight_current_fnorm**2 - self.weight_original_fnorm**2)
            # loss = torch.abs(weight_current_fnorm ** 2 + 0.01 * self.weight_main_fnorm ** 2 - 1.01 * self.weight_original_fnorm ** 2)
        else:
            loss = 0.0

        return loss

    def compute_fn_loss(self):
        if self.S_residual is not None:
            weight_current = self.weight_main + self.U_residual @ torch.diag(self.S_residual) @ self.V_residual
            weight_current_fnorm = torch.norm(weight_current, p="fro")

            loss = weight_current_fnorm**2
        else:
            loss = 0.0

        return loss


# Function to replace nn.Linear modules within self_attn modules with SVDResidualLinear
def apply_svd_residual_to_self_attn(model, r):
    for name, module in model.named_children():
        if "self_attn" in name:
            # Replace nn.Linear layers in this module
            for sub_name, sub_module in module.named_modules():
                if isinstance(sub_module, nn.Linear):
                    # Get parent module within self_attn
                    parent_module = module
                    sub_module_names = sub_name.split(".")
                    for module_name in sub_module_names[:-1]:
                        parent_module = getattr(parent_module, module_name)
                    # Replace the nn.Linear layer with SVDResidualLinear
                    setattr(parent_module, sub_module_names[-1], replace_with_svd_residual(sub_module, r))
        else:
            # Recursively apply to child modules
            apply_svd_residual_to_self_attn(module, r)
    # After replacing, set requires_grad for residual components
    for param_name, param in model.named_parameters():
        if any(x in param_name for x in ["S_residual", "U_residual", "V_residual"]):
            param.requires_grad = True
        else:
            param.requires_grad = False
    return model


# Function to replace a module with SVDResidualLinear
def replace_with_svd_residual(module, r):
    if isinstance(module, nn.Linear):
        in_features = module.in_features
        out_features = module.out_features
        bias = module.bias is not None

        # Create SVDResidualLinear module
        new_module = SVDResidualLinear(in_features, out_features, r, bias=bias, init_weight=module.weight.data.clone())

        if bias and module.bias is not None:
            new_module.bias.data.copy_(module.bias.data)

        new_module.weight_original_fnorm = torch.norm(module.weight.data, p="fro")

        # Perform SVD on the original weight
        U, S, Vh = torch.linalg.svd(module.weight.data, full_matrices=False)

        # Determine r based on the rank of the weight matrix
        r = min(r, len(S))  # Ensure r does not exceed the number of singular values

        # Keep top r singular components (main weight)
        U_r = U[:, :r]  # Shape: (out_features, r)
        S_r = S[:r]  # Shape: (r,)
        Vh_r = Vh[:r, :]  # Shape: (r, in_features)

        # Reconstruct the main weight (fixed)
        weight_main = U_r @ torch.diag(S_r) @ Vh_r

        # Calculate the frobenius norm of main weight
        new_module.weight_main_fnorm = torch.norm(weight_main.data, p="fro")

        # Set the main weight
        new_module.weight_main.data.copy_(weight_main)

        # Residual components (trainable)
        U_residual = U[:, r:]  # Shape: (out_features, n - r)
        S_residual = S[r:]  # Shape: (n - r,)
        Vh_residual = Vh[r:, :]  # Shape: (n - r, in_features)

        if len(S_residual) > 0:
            new_module.S_residual = nn.Parameter(S_residual.clone())
            new_module.U_residual = nn.Parameter(U_residual.clone())
            new_module.V_residual = nn.Parameter(Vh_residual.clone())

            new_module.S_r = nn.Parameter(S_r.clone(), requires_grad=False)
            new_module.U_r = nn.Parameter(U_r.clone(), requires_grad=False)
            new_module.V_r = nn.Parameter(Vh_r.clone(), requires_grad=False)
        else:
            new_module.S_residual = None
            new_module.U_residual = None
            new_module.V_residual = None

            new_module.S_r = None
            new_module.U_r = None
            new_module.V_r = None

        return new_module
    else:
        return module


# This is the original preprocessing used in Effort paper
# Gives almost the same results as `preprocessing`
_preprocessing_original = T.Compose(
    [
        T.ToTensor(),
        T.Normalize(
            [0.48145466, 0.4578275, 0.40821073],
            [0.26862954, 0.26130258, 0.27577711],
        ),
    ]
)


def preprocessing_original(image: Image) -> torch.Tensor:
    image = np.array(image)
    image = cv2.resize(image, (224, 224), interpolation=cv2.INTER_LINEAR)
    return _preprocessing_original(image)