1{
2  "end_line": 16,
3  "filename": "bench/poison/custom_robust_loss.py",
4  "start_line": 12,
5  "op": "replace",
6  "after_line": 1,
7  "content": "    def __init__(self):\n        self.k = 1.0 # Decay rate for high-loss samples\n\n    def compute_loss(self, logits, labels, epoch):\n        losses = F.cross_entropy(logits, labels, reduction='none')\n        \n        with torch.no_grad():\n            mu = losses.mean()\n            sigma = losses.std()\n            \n            # Z-score normalization of losses\n            z = (losses - mu) / (sigma + 1e-8)\n            \n            # Suppress samples with higher than average loss\n            w = torch.exp(-self.k * F.relu(z))\n            \n            # Normalize to preserve overall gradient magnitude\n            w = w / (w.mean() + 1e-8)\n            \n        return (w * losses).mean()\n"
8}

1{
2  "after_line": 1,
3  "start_line": 12,
4  "op": "replace",
5  "content": "    def __init__(self):\n        self.k = 1.0 # Decay rate for high-loss samples\n\n    def compute_loss(self, logits, labels, epoch):\n        losses = F.cross_entropy(logits, labels, reduction='none')\n        \n        with torch.no_grad():\n            mu = losses.mean()\n            sigma = losses.std()\n            \n            # Z-score normalization of losses\n            z = (losses - mu) / (sigma + 1e-8)\n            \n            # Suppress samples with higher than average loss\n            w = torch.exp(-self.k * F.relu(z))\n            \n            # Normalize to preserve overall gradient magnitude\n            w = w / (w.mean() + 1e-8)\n            \n        return (w * losses).mean()",
6  "filename": "pytorch-vision/bench/poison/custom_robust_loss.py",
7  "end_line": 16
8}

1{
2  "end_line": 31,
3  "filename": "pytorch-vision/bench/poison/custom_robust_loss.py",
4  "start_line": 12,
5  "op": "replace",
6  "content": "    def __init__(self):\n        self.alpha = 1.0\n        self.beta = 1.0\n\n    def compute_loss(self, logits, labels, epoch):\n        probs = F.softmax(logits, dim=1).clamp_min(1e-8)\n        log_probs = torch.log(probs)\n        \n        # NCE (Normalized Cross Entropy)\n        ce = F.cross_entropy(logits, labels, reduction='none')\n        denom = -log_probs.sum(dim=1)\n        nce = (ce / (denom + 1e-8)).mean()\n        \n        # MAE / RCE (Mean Absolute Error)\n        p_y = probs.gather(1, labels[:, None]).squeeze()\n        mae = (1.0 - p_y).mean()\n        \n        return self.alpha * nce + self.beta * mae",
7  "after_line": 1
8}

1{
2  "start_line": 10,
3  "end_line": 16,
4  "filename": "bench/poison/custom_robust_loss.py",
5  "op": "replace",
6  "content": "    \"\"\"Generalized Jensen-Shannon Divergence Loss.\"\"\"\n\n    def __init__(self):\n        self.pi = 0.5\n\n    def compute_loss(self, logits, labels, epoch):\n        p = torch.softmax(logits, dim=1)\n        y = F.one_hot(labels, num_classes=logits.shape[1]).float()\n        \n        m = self.pi * y + (1 - self.pi) * p\n        m = m.clamp_min(1e-8)\n        p_safe = p.clamp_min(1e-8)\n        \n        # KL(Y || M) = - \\sum Y \\log(M) + \\sum Y \\log Y = - \\log(M_y)\n        kl_y_m = -torch.log(m.gather(1, labels[:, None])).squeeze(1)\n        \n        # KL(P || M) = \\sum P \\log(P / M)\n        kl_p_m = (p * (torch.log(p_safe) - torch.log(m))).sum(dim=1)\n        \n        loss = self.pi * kl_y_m + (1.0 - self.pi) * kl_p_m\n        return loss.mean()"
7}

1{
2  "filename": "pytorch-vision/bench/poison/custom_robust_loss.py",
3  "content": "    \"\"\"Generalized Jensen-Shannon Divergence Loss.\"\"\"\n\n    def __init__(self):\n        self.pi = 0.5\n\n    def compute_loss(self, logits, labels, epoch):\n        p = torch.softmax(logits, dim=1)\n        y = F.one_hot(labels, num_classes=logits.shape[1]).float()\n        \n        m = self.pi * y + (1.0 - self.pi) * p\n        m = m.clamp_min(1e-8)\n        p_safe = p.clamp_min(1e-8)\n        \n        # KL(Y || M) = - \\sum Y \\log(M)\n        kl_y_m = -torch.log(m.gather(1, labels[:, None])).squeeze(1)\n        \n        # KL(P || M) = \\sum P \\log(P / M)\n        kl_p_m = (p * (torch.log(p_safe) - torch.log(m))).sum(dim=1)\n        \n        loss = self.pi * kl_y_m + (1.0 - self.pi) * kl_p_m\n        return loss.mean()",
4  "start_line": 10,
5  "end_line": 16,
6  "op": "replace"
7}

1{
2  "end_line": 18,
3  "op": "replace",
4  "start_line": 9,
5  "content": "class RobustLoss:\n    \"\"\"Dynamic Bootstrapped Generalized Cross-Entropy (DBGCE).\n    \n    Combines the robust gradient-bounding of Generalized Cross-Entropy (GCE)\n    with dynamic self-distillation (soft bootstrapping) to prevent memorization\n    of noisy labels and correct targets dynamically over the training schedule.\n    \"\"\"\n\n    def __init__(self):\n        self.q = 0.7\n\n    def compute_loss(self, logits, labels, epoch):\n        probs = torch.softmax(logits, dim=1)\n        \n        # Hard targeted labels\n        y_hard = F.one_hot(labels, num_classes=logits.shape[1]).float()\n        # Soft targeted pseudo-labels from the model's current predictions\n        y_soft = probs.detach()\n        \n        # Dynamic mixing coefficient (beta):\n        # - Epochs 0-15 (Warmup): beta=1.0. The model learns clean, easy samples \n        #   using GCE to bound the influence of noisy labels.\n        # - Epochs 15-100: beta decays linearly from 1.0 to 0.4. The model increasingly \n        #   relies on its own clean representations (self-distillation) to correct \n        #   the poisoned targets.\n        warmup = 15\n        if epoch < warmup:\n            beta = 1.0\n        else:\n            beta = max(0.4, 1.0 - 0.6 * (epoch - warmup) / (100.0 - warmup))\n            \n        target = beta * y_hard + (1.0 - beta) * y_soft\n        \n        # Element-wise Generalized Cross Entropy for soft targets\n        p_q = probs.clamp(min=1e-7).pow(self.q)\n        loss = target * (1.0 - p_q) / self.q\n        \n        return loss.sum(dim=1).mean()",
6  "filename": "bench/poison/custom_robust_loss.py"
7}

1{
2  "start_line": 9,
3  "op": "replace",
4  "content": "class RobustLoss:\n    \"\"\"Dynamic Bootstrapped Generalized Cross-Entropy (DBGCE).\n    \n    Combines the robust gradient-bounding of Generalized Cross-Entropy (GCE)\n    with dynamic self-distillation (soft bootstrapping) to prevent memorization\n    of noisy labels and correct targets dynamically over the training schedule.\n    \"\"\"\n\n    def __init__(self):\n        self.q = 0.7\n\n    def compute_loss(self, logits, labels, epoch):\n        probs = torch.softmax(logits, dim=1)\n        \n        # Hard targeted labels\n        y_hard = F.one_hot(labels, num_classes=logits.shape[1]).float()\n        # Soft targeted pseudo-labels from the model's current predictions\n        y_soft = probs.detach()\n        \n        # Dynamic mixing coefficient (beta):\n        # - Epochs 0-15 (Warmup): beta=1.0. The model learns clean, easy samples \n        #   using GCE to bound the influence of noisy labels.\n        # - Epochs 15-100: beta decays linearly from 1.0 to 0.4. The model increasingly \n        #   relies on its own clean representations (self-distillation) to correct \n        #   the poisoned targets.\n        warmup = 15\n        if epoch < warmup:\n            beta = 1.0\n        else:\n            beta = max(0.4, 1.0 - 0.6 * (epoch - warmup) / (100.0 - warmup))\n            \n        target = beta * y_hard + (1.0 - beta) * y_soft\n        \n        # Element-wise Generalized Cross Entropy for soft targets\n        p_q = probs.clamp(min=1e-7).pow(self.q)\n        loss = target * (1.0 - p_q) / self.q\n        \n        return loss.sum(dim=1).mean()",
5  "filename": "pytorch-vision/bench/poison/custom_robust_loss.py",
6  "end_line": 18
7}