File size: 10,393 Bytes
01f95f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Loss functions for ViL Tracker training.

Includes:
- FocalLoss: for center heatmap prediction (handles class imbalance)
- GIoULoss: for bounding box regression
- UncertaintyNLLLoss: uncertainty-aware NLL loss
- MemoryContrastiveLoss: contrastive loss for mLSTM memory states
- AFKDDistillationLoss: attention-free knowledge distillation
- ADWLoss: adaptive dynamic weighting for multi-task loss
- CombinedTrackingLoss: combines all losses with learned weighting
"""

import torch
import torch.nn as nn
import torch.nn.functional as F


class FocalLoss(nn.Module):
    """Focal loss for heatmap prediction (CornerNet-style).
    
    Handles extreme foreground/background imbalance in center heatmaps
    where only ~1/256 positions are positive.
    """
    def __init__(self, alpha: float = 2.0, beta: float = 4.0):
        super().__init__()
        self.alpha = alpha
        self.beta = beta
    
    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
        """
        Args:
            pred: (B, 1, H, W) predicted heatmap (logits)
            target: (B, 1, H, W) ground truth Gaussian heatmap
        """
        pred_sig = torch.sigmoid(pred)
        pred_sig = pred_sig.clamp(1e-6, 1 - 1e-6)
        
        pos_mask = target.eq(1).float()
        neg_mask = target.lt(1).float()
        
        # Positive loss
        pos_loss = -((1 - pred_sig) ** self.alpha) * torch.log(pred_sig) * pos_mask
        
        # Negative loss (weighted by distance from GT peak)
        neg_weight = (1 - target) ** self.beta
        neg_loss = -(pred_sig ** self.alpha) * torch.log(1 - pred_sig) * neg_weight * neg_mask
        
        num_pos = pos_mask.sum().clamp(min=1)
        loss = (pos_loss.sum() + neg_loss.sum()) / num_pos
        return loss


class GIoULoss(nn.Module):
    """Generalized IoU loss for bounding box regression.
    
    Better gradient signal than L1 for box prediction, especially
    for non-overlapping boxes.
    """
    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
        """
        Args:
            pred: (B, 4) predicted [cx, cy, w, h]
            target: (B, 4) ground truth [cx, cy, w, h]
        """
        # Convert to [x1, y1, x2, y2]
        pred_x1 = pred[:, 0] - pred[:, 2] / 2
        pred_y1 = pred[:, 1] - pred[:, 3] / 2
        pred_x2 = pred[:, 0] + pred[:, 2] / 2
        pred_y2 = pred[:, 1] + pred[:, 3] / 2
        
        gt_x1 = target[:, 0] - target[:, 2] / 2
        gt_y1 = target[:, 1] - target[:, 3] / 2
        gt_x2 = target[:, 0] + target[:, 2] / 2
        gt_y2 = target[:, 1] + target[:, 3] / 2
        
        # Intersection
        inter_x1 = torch.max(pred_x1, gt_x1)
        inter_y1 = torch.max(pred_y1, gt_y1)
        inter_x2 = torch.min(pred_x2, gt_x2)
        inter_y2 = torch.min(pred_y2, gt_y2)
        inter_area = (inter_x2 - inter_x1).clamp(min=0) * (inter_y2 - inter_y1).clamp(min=0)
        
        # Union
        pred_area = (pred_x2 - pred_x1).clamp(min=0) * (pred_y2 - pred_y1).clamp(min=0)
        gt_area = (gt_x2 - gt_x1).clamp(min=0) * (gt_y2 - gt_y1).clamp(min=0)
        union_area = pred_area + gt_area - inter_area
        
        iou = inter_area / union_area.clamp(min=1e-6)
        
        # Enclosing box
        enc_x1 = torch.min(pred_x1, gt_x1)
        enc_y1 = torch.min(pred_y1, gt_y1)
        enc_x2 = torch.max(pred_x2, gt_x2)
        enc_y2 = torch.max(pred_y2, gt_y2)
        enc_area = (enc_x2 - enc_x1).clamp(min=0) * (enc_y2 - enc_y1).clamp(min=0)
        
        giou = iou - (enc_area - union_area) / enc_area.clamp(min=1e-6)
        return (1 - giou).mean()


class UncertaintyNLLLoss(nn.Module):
    """Uncertainty-aware negative log-likelihood loss.
    
    Weighs the regression loss by predicted uncertainty:
    L = 0.5 * exp(-s) * |pred - target|^2 + 0.5 * s
    where s = log(variance).
    """
    def forward(self, pred: torch.Tensor, target: torch.Tensor, log_var: torch.Tensor) -> torch.Tensor:
        """
        Args:
            pred: (B, ...) predictions
            target: (B, ...) targets
            log_var: (B, ...) predicted log variance
        """
        precision = torch.exp(-log_var)
        sq_error = (pred - target) ** 2
        loss = 0.5 * (precision * sq_error + log_var)
        return loss.mean()


class MemoryContrastiveLoss(nn.Module):
    """Contrastive loss for mLSTM memory states.
    
    Encourages similar memory states for the same target across frames
    and dissimilar states for different targets.
    """
    def __init__(self, temperature: float = 0.1):
        super().__init__()
        self.temperature = temperature
    
    def forward(self, feat_a: torch.Tensor, feat_b: torch.Tensor) -> torch.Tensor:
        """
        Args:
            feat_a: (B, D) features from frame A
            feat_b: (B, D) features from frame B (same target)
        """
        # L2 normalize
        feat_a = F.normalize(feat_a, dim=-1)
        feat_b = F.normalize(feat_b, dim=-1)
        
        B = feat_a.shape[0]
        
        # Similarity matrix
        sim = torch.mm(feat_a, feat_b.t()) / self.temperature  # (B, B)
        
        # Positive pairs along diagonal
        labels = torch.arange(B, device=feat_a.device)
        loss = F.cross_entropy(sim, labels)
        return loss


class AFKDDistillationLoss(nn.Module):
    """Attention-Free Knowledge Distillation loss.
    
    For distilling from MCITrack-B256 teacher to ViL-S student.
    Uses feature matching + response-based distillation.
    """
    def __init__(self, student_dim: int = 384, teacher_dim: int = 768, temperature: float = 4.0):
        super().__init__()
        self.temperature = temperature
        # Projector to match dimensions
        self.projector = nn.Sequential(
            nn.Linear(student_dim, teacher_dim),
            nn.GELU(),
            nn.Linear(teacher_dim, teacher_dim),
        )
    
    def forward(
        self,
        student_feat: torch.Tensor,
        teacher_feat: torch.Tensor,
        student_logits: torch.Tensor = None,
        teacher_logits: torch.Tensor = None,
    ) -> torch.Tensor:
        """
        Args:
            student_feat: (B, S, D_s) student features
            teacher_feat: (B, S, D_t) teacher features
            student_logits: optional (B, ...) student predictions
            teacher_logits: optional (B, ...) teacher predictions
        """
        # Feature distillation
        student_proj = self.projector(student_feat)
        feat_loss = F.mse_loss(student_proj, teacher_feat.detach())
        
        # Response distillation (if logits provided)
        if student_logits is not None and teacher_logits is not None:
            T = self.temperature
            s_soft = F.log_softmax(student_logits.view(student_logits.shape[0], -1) / T, dim=-1)
            t_soft = F.softmax(teacher_logits.view(teacher_logits.shape[0], -1) / T, dim=-1)
            resp_loss = F.kl_div(s_soft, t_soft.detach(), reduction='batchmean') * (T ** 2)
            return feat_loss + resp_loss
        
        return feat_loss


class ADWLoss(nn.Module):
    """Adaptive Dynamic Weighting for multi-task loss.
    
    Learns task weights based on loss magnitudes using homoscedastic uncertainty.
    w_k = 1/(2*sigma_k^2), regularizer = log(sigma_k)
    """
    def __init__(self, num_tasks: int = 4):
        super().__init__()
        # Log variance parameters (initialized to 0 = equal weighting)
        self.log_vars = nn.Parameter(torch.zeros(num_tasks))
    
    def forward(self, losses: list) -> torch.Tensor:
        """
        Args:
            losses: list of scalar loss tensors (one per task)
        Returns:
            weighted sum of losses
        """
        total = 0
        for i, loss in enumerate(losses):
            precision = torch.exp(-self.log_vars[i])
            total = total + precision * loss + self.log_vars[i]
        return total


class CombinedTrackingLoss(nn.Module):
    """Combined loss for tracker training.
    
    Combines:
    - Focal loss on center heatmap
    - GIoU loss on predicted boxes
    - L1 loss on size regression
    - Optional: uncertainty NLL, contrastive, distillation
    """
    def __init__(self, use_uncertainty: bool = True, use_adw: bool = True):
        super().__init__()
        self.focal = FocalLoss()
        self.giou = GIoULoss()
        self.l1 = nn.L1Loss()
        self.use_uncertainty = use_uncertainty
        
        if use_uncertainty:
            self.uncertainty_loss = UncertaintyNLLLoss()
        
        num_tasks = 4 if use_uncertainty else 3
        self.adw = ADWLoss(num_tasks=num_tasks) if use_adw else None
    
    def forward(
        self,
        pred: dict,
        gt_heatmap: torch.Tensor,
        gt_size: torch.Tensor,
        gt_boxes: torch.Tensor,
    ) -> dict:
        """
        Args:
            pred: model output dict with 'heatmap', 'size', 'boxes', optionally 'log_variance'
            gt_heatmap: (B, 1, H, W) ground truth heatmap
            gt_size: (B, 2) ground truth normalized size [w, h]
            gt_boxes: (B, 4) ground truth boxes [cx, cy, w, h] in pixels
        """
        # Heatmap loss
        heatmap_loss = self.focal(pred['heatmap'], gt_heatmap)
        
        # Size loss (at peak location)
        B = gt_size.shape[0]
        pred_size = pred['size'].view(B, 2, -1).mean(dim=-1)  # average pool
        size_loss = self.l1(pred_size, gt_size)
        
        # GIoU box loss
        giou_loss = self.giou(pred['boxes'], gt_boxes)
        
        losses = [heatmap_loss, size_loss, giou_loss]
        
        # Uncertainty loss
        if self.use_uncertainty and 'log_variance' in pred:
            log_var = pred['log_variance'].mean(dim=[1, 2, 3])  # (B,)
            unc_loss = (0.5 * torch.exp(-log_var) * giou_loss + 0.5 * log_var).mean()
            losses.append(unc_loss)
        
        # Combine with ADW or simple sum
        if self.adw is not None:
            total_loss = self.adw(losses)
        else:
            weights = [1.0, 1.0, 2.0, 0.5] if len(losses) == 4 else [1.0, 1.0, 2.0]
            total_loss = sum(w * l for w, l in zip(weights, losses))
        
        return {
            'total': total_loss,
            'heatmap': heatmap_loss.detach(),
            'size': size_loss.detach(),
            'giou': giou_loss.detach(),
        }