1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_loss.py",
4  "start_line": 246,
5  "end_line": 266,
6  "content": "def compute_loss(logits, targets, config):\n    \"\"\"Adaptive Curriculum Margin Loss (ACML).\n\n    Combines:\n    1. Curriculum margin that increases with epoch progression.\n    2. Focal weighting for hard examples.\n    3. Label smoothing for calibration.\n    4. Adaptive temperature scaling.\n\n    Hyperparameters:\n        base_margin: maximum margin at final epoch\n        gamma: controls margin progression rate\n        beta: focal weighting strength\n        tau: temperature scaling magnitude\n        smoothing_eps: label smoothing factor\n    \"\"\"\n    # Hyperparameters (hardcoded)\n    base_margin = 0.5\n    gamma = 2.0\n    beta = 2.0\n    tau = 1.0\n    smoothing_eps = 0.1\n\n    num_classes = config['num_classes']\n    epoch = config['epoch']\n    total_epochs = config['total_epochs']\n\n    B = logits.size(0)\n\n    # Curriculum progression factor (0 to 1)\n    progress = epoch / max(total_epochs - 1, 1)  # avoid division by zero\n\n    # Dynamic margin that increases with progress\n    margin = base_margin * (progress ** gamma)\n\n    # Adaptive temperature: starts warm (higher temp), ends cool (lower temp)\n    temperature = 1.0 + tau * (1.0 - progress)  # temperature decreases as progress increases\n\n    # Apply margin to logits: subtract margin from true class logit\n    logits_adjusted = logits.clone()\n    logits_adjusted[range(B), targets] -= margin\n\n    # Apply temperature scaling\n    logits_adjusted = logits_adjusted / temperature\n\n    # Compute softmax probabilities\n    probs = F.softmax(logits_adjusted, dim=-1)\n    pt = probs[range(B), targets]  # probability of true class\n\n    # Focal weighting: down-weight easy examples\n    focal_weight = (1.0 - pt) ** beta\n\n    # Label smoothing target distribution\n    one_hot = F.one_hot(targets, num_classes).float()\n    smoothed_targets = (1.0 - smoothing_eps) * one_hot + smoothing_eps / num_classes\n\n    # Compute cross-entropy between smoothed targets and adjusted logits\n    log_probs = F.log_softmax(logits_adjusted, dim=-1)\n    ce = -(smoothed_targets * log_probs).sum(dim=-1)  # [B]\n\n    # Weighted loss\n    loss = (focal_weight * ce).mean()\n\n    return loss"
7}

1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_loss.py",
4  "start_line": 246,
5  "end_line": 309,
6  "content": "def compute_loss(logits, targets, config):\n    \"\"\"Adaptive Margin Focal Loss (AMFL).\n\n    Combines:\n    1. Sample-dependent margin: harder examples get larger margin.\n    2. Curriculum progression: margin magnitude increases over epochs.\n    3. Focal weighting for hard examples.\n    4. Label smoothing for calibration.\n    5. Adaptive temperature scaling.\n\n    Hyperparameters:\n        base_margin: maximum margin at final epoch for easy samples\n        gamma: controls margin progression rate over epochs\n        alpha: controls margin scaling with sample hardness\n        beta: focal weighting strength\n        tau: temperature scaling magnitude\n        smoothing_eps: label smoothing factor\n    \"\"\"\n    # Hyperparameters (hardcoded)\n    base_margin = 0.3\n    gamma = 1.0\n    alpha = 1.0\n    beta = 2.0\n    tau = 0.5\n    smoothing_eps = 0.1\n\n    num_classes = config['num_classes']\n    epoch = config['epoch']\n    total_epochs = config['total_epochs']\n\n    B = logits.size(0)\n\n    # Curriculum progression factor (0 to 1)\n    progress = epoch / max(total_epochs - 1, 1)  # avoid division by zero\n\n    # Adaptive temperature: starts warm (higher temp), ends cool (lower temp)\n    temperature = 1.0 + tau * (1.0 - progress)  # temperature decreases as progress increases\n\n    # Apply temperature scaling to raw logits\n    logits_scaled = logits / temperature\n\n    # Compute raw probabilities (before margin)\n    probs_raw = F.softmax(logits_scaled, dim=-1)\n    pt_raw = probs_raw[range(B), targets]  # probability of true class before margin\n\n    # Sample-dependent margin: larger margin for harder examples\n    hardness = 1.0 - pt_raw  # hardness ∈ [0,1]\n    margin = base_margin * (progress ** gamma) * (hardness ** alpha)  # [B]\n\n    # Apply margin to logits: subtract margin from true class logit\n    logits_adjusted = logits_scaled.clone()\n    logits_adjusted[range(B), targets] -= margin\n\n    # Compute probabilities after margin\n    probs = F.softmax(logits_adjusted, dim=-1)\n    pt = probs[range(B), targets]  # probability of true class after margin\n\n    # Focal weighting: down-weight easy examples\n    focal_weight = (1.0 - pt) ** beta\n\n    # Label smoothing target distribution\n    one_hot = F.one_hot(targets, num_classes).float()\n    smoothed_targets = (1.0 - smoothing_eps) * one_hot + smoothing_eps / num_classes\n\n    # Compute cross-entropy between smoothed targets and adjusted logits\n    log_probs = F.log_softmax(logits_adjusted, dim=-1)\n    ce = -(smoothed_targets * log_probs).sum(dim=-1)  # [B]\n\n    # Weighted loss\n    loss = (focal_weight * ce).mean()\n\n    return loss"
7}

1{
2  "op": "replace",
3  "filename": "pytorch-vision/custom_loss.py",
4  "start_line": 246,
5  "end_line": 317,
6  "content": "def compute_loss(logits, targets, config):\n    \"\"\"Curriculum PolyLoss with Adaptive Smoothing.\n\n    Combines:\n    1. Label smoothing (fixed epsilon).\n    2. Poly loss term with curriculum: epsilon decreases over epochs.\n    3. Adaptive temperature scaling.\n\n    Loss = CE_smoothed + epsilon * (1 - pt)\n    where epsilon = base_epsilon * (1 - progress)^gamma\n    and temperature = 1 + tau * (1 - progress)\n\n    Hyperparameters:\n        smoothing_eps: label smoothing factor\n        base_epsilon: maximum poly weight at start\n        gamma: controls decay rate of poly weight\n        tau: temperature scaling magnitude\n    \"\"\"\n    # Hyperparameters (hardcoded)\n    smoothing_eps = 0.1\n    base_epsilon = 2.0\n    gamma = 0.5\n    tau = 0.5\n\n    num_classes = config['num_classes']\n    epoch = config['epoch']\n    total_epochs = config['total_epochs']\n\n    # Curriculum progression factor (0 to 1)\n    progress = epoch / max(total_epochs - 1, 1)  # avoid division by zero\n\n    # Adaptive temperature: starts warm (higher temp), ends cool (lower temp)\n    temperature = 1.0 + tau * (1.0 - progress)  # temperature decreases as progress increases\n\n    # Apply temperature scaling\n    logits_scaled = logits / temperature\n\n    # Compute probabilities\n    probs = F.softmax(logits_scaled, dim=-1)\n    pt = probs[range(len(targets)), targets]  # probability of true class\n\n    # Label smoothing target distribution\n    one_hot = F.one_hot(targets, num_classes).float()\n    smoothed_targets = (1.0 - smoothing_eps) * one_hot + smoothing_eps / num_classes\n\n    # Cross-entropy with smoothed targets\n    log_probs = F.log_softmax(logits_scaled, dim=-1)\n    ce = -(smoothed_targets * log_probs).sum(dim=-1)  # [B]\n\n    # Curriculum poly weight\n    epsilon = base_epsilon * ((1.0 - progress) ** gamma)\n    poly_term = epsilon * (1.0 - pt)\n\n    # Total loss\n    loss = (ce + poly_term).mean()\n\n    return loss"
7}

1{
2  "n": 1
3}