File size: 10,841 Bytes

"""
Define loss functions needed for training the model — padding safe (-1 sentinel)
"""

import torch
import torch.nn.functional as F
from torchmetrics.functional.classification import (
    auroc, average_precision, roc, precision_recall_curve
)
import rootutils
from dpacman.utils import pylogger

root = rootutils.setup_root(__file__, indicator=".project-root", pythonpath=True)
logger = pylogger.RankedLogger(__name__, rank_zero_only=True)


def _expand_like(mask: torch.Tensor, like: torch.Tensor):
    # Make mask broadcastable to logits/targets (handles (B,L) vs (B,L,1))
    while mask.dim() < like.dim():
        mask = mask.unsqueeze(-1)
    return mask.expand_as(like)

def bce_loss_masked(logits, targets, mask, pos_weight=None, eps=1e-8):
    """
    Compute masked BCE with logits over non-peak positions only.
    Expects nonpeak_mask already broadcastable to logits.
    """
    # Clamp targets into [0,1] to be safe, even if pads slip through earlier
    t = targets.clamp(0.0, 1.0)
    loss = F.binary_cross_entropy_with_logits(
        logits, t, reduction="none", pos_weight=pos_weight
    )
    m = _expand_like(mask, loss).to(loss.dtype)
    denom = m.sum().clamp_min(eps)
    return (loss * m).sum() / denom

def mse_peaks_only(logits, targets, peak_mask, eps=1e-8):
    """
    Calculate MSE on peaks only (on probabilities), masking everything else.
    """
    probs = torch.sigmoid(logits)
    per_elem = F.mse_loss(probs, targets, reduction="none")
    m = _expand_like(peak_mask, per_elem).to(per_elem.dtype)
    denom = m.sum().clamp_min(eps)
    return (per_elem * m).sum() / denom

def calculate_loss(
    logits,
    targets,
    binder_kpm,
    glm_kpm,
    eps: float = 1e-8,
    alpha: float = 1.0,
    gamma: float = 1.0,
    pos_weight=None,
    pad_value: float = -1.0,
    loss_type="mixed"
):
    """
    Combine masked-BCE (non-peak) + masked-MSE on probs (peak), ignoring padding.
    Assumes targets == -1 are pads; non-peak = 0; peak > 0.
    
    binder_kpm is 1 at PAD positions, 0 elsewhere
    glm_kpm is 1 at PAD positions, 0 elsewhere
    
    if loss_type is mixed, we're doing binary cross entropy off the peaks and MSE on the peaks. 
    if loss_type is binary, we're doing binary cross entropy everywhere because the labels are binary. 
    """
    # calculate validity in two ways; these should be the same. 
    # targets are padded to -1 where there is not really a DNA sequence there
    valid = (targets != pad_value)
    if glm_kpm is not None:
        nvalid = torch.sum(valid).item()
        nvalid_2 = torch.sum(~glm_kpm).item()
        assert nvalid==nvalid_2

    # Peak / non-peak masks that exclude pads
    nonpeak_mask = valid & (targets == 0)
    peak_mask    = valid & (targets > 0)

    # For safety, zero-out targets at pad positions so they never feed into BCE/MSE
    targets_safe = torch.where(valid, targets, torch.zeros_like(targets))

    if loss_type=="mixed":
        bce_nonpeak = bce_loss_masked(logits, targets_safe, nonpeak_mask, pos_weight=pos_weight, eps=eps)
        mse_peak    = mse_peaks_only(logits, targets_safe, peak_mask, eps=eps)
        return alpha * bce_nonpeak + gamma * mse_peak
    else:
        # we're expecting all binary labels. make sure. 
        all_binary = ((targets_safe==1) | (targets_safe==0)).all().item()
        if not(all_binary):
            logger.info(f"WARNING: expecting all binary labels for loss_type={loss_type}. Did not get all binary labels.")
        # bce over all valid positions
        bce_all = bce_loss_masked(logits, targets_safe, valid, pos_weight=pos_weight, eps=eps)
        return alpha*bce_all

@torch.no_grad()
def auroc_zeros_vs_ones_from_logits(
    logits: torch.Tensor,            # (B, L)
    labels: torch.Tensor,            # (B, L)
    glm_kpm: torch.Tensor | None = None,  # (B, L) True=PAD
    pos_thresh: float = 0.99,
):
    """
    Returns:
      auc:            scalar tensor (AUROC)
      n_pos, n_neg:   ints
      tpr, fpr:       tensors of shape (T,)
      thresholds:     tensor of shape (T,)
      tp, fp:         integer counts per threshold (shape (T,))
    """
    device = logits.device
    # glm_kpm is 1 where there's a pad, so ~glm_kpm is valid positions
    valid = ~glm_kpm if glm_kpm is not None else torch.ones_like(labels, dtype=torch.bool, device=device)
    keep = valid & ((labels > pos_thresh) | (labels == 0.0))
    if keep.sum() == 0:
        return (torch.tensor(float('nan'), device=device), 0, 0,
                torch.empty(0, device=device), torch.empty(0, device=device),
                torch.empty(0, device=device), torch.empty(0, device=device), torch.empty(0, device=device))

    y = (labels[keep] > pos_thresh).to(torch.int)
    s = logits[keep]

    n_pos = int(y.sum().item())
    n_neg = y.numel() - n_pos
    if n_pos == 0 or n_neg == 0:
        return (torch.tensor(float('nan'), device=device), n_pos, n_neg,
                torch.empty(0, device=device), torch.empty(0, device=device),
                torch.empty(0, device=device), torch.empty(0, device=device), torch.empty(0, device=device))

    # Full ROC curve
    fpr, tpr, thresholds = roc(s, y, task="binary")
    # AUROC (TM handles logits)
    auc = auroc(s, y, task="binary")

    # Convert rates to counts (round to nearest to avoid float off-by-one)
    tp = (tpr * n_pos).round().to(torch.long)
    fp = (fpr * n_neg).round().to(torch.long)

    return auc.to(device), n_pos, n_neg, tpr.to(device), fpr.to(device), thresholds.to(device), tp.to(device), fp.to(device)


@torch.no_grad()
def auprc_zeros_vs_ones_from_logits(
    logits: torch.Tensor,            # (B, L)
    labels: torch.Tensor,            # (B, L)
    glm_kpm: torch.Tensor | None = None,  # (B, L) True=PAD
    pos_thresh: float = 0.99,
):
    """
    Returns:
      ap:             scalar tensor (Average Precision / AUPRC)
      n_pos, n_neg:   ints
      precision:      (T,)
      recall:         (T,)
      thresholds:     (T,)
    """
    device = logits.device
    # glm_kpm is 1 where there's a pad, so ~glm_kpm is valid
    valid = ~glm_kpm if glm_kpm is not None else torch.ones_like(labels, dtype=torch.bool, device=device)
    keep = valid & ((labels > pos_thresh) | (labels == 0.0))
    if keep.sum() == 0:
        return (torch.tensor(float('nan'), device=device), 0, 0,
                torch.empty(0, device=device), torch.empty(0, device=device), torch.empty(0, device=device))

    y = (labels[keep] > pos_thresh).to(torch.int)
    s = logits[keep]

    n_pos = int(y.sum().item())
    n_neg = y.numel() - n_pos
    if n_pos == 0:
        # By convention, AP=0 when there are no positives
        return (torch.tensor(0.0, device=device), 0, n_neg,
                torch.empty(0, device=device), torch.empty(0, device=device), torch.empty(0, device=device))

    # Full PR curve
    precision, recall, thresholds = precision_recall_curve(s, y, task="binary")
    # Average Precision / AUPRC
    ap = average_precision(s, y, task="binary")

    return ap.to(device), n_pos, n_neg, precision.to(device), recall.to(device), thresholds.to(device)

def accuracy_percentage(
    logits,
    targets,
    peak_thresh: float = 0.5,
    eps: float = 1e-8,
    pad_value: float = -1.0,
):
    """
    Compute accuracy for predicting high-confidence peaks (prob >= 0.5), ignoring padding.
    """
    valid = (targets != pad_value)
    probs = torch.sigmoid(logits)
    preds_bin = (probs >= 0.5)
    labels    = (targets >= peak_thresh)

    v = _expand_like(valid, preds_bin)
    correct = ((preds_bin == labels) & v).to(torch.float32).sum()
    total   = v.to(torch.float32).sum().clamp_min(eps)
    return (correct / total).item() * 100.0

if __name__ == "__main__":
    import torch

    torch.manual_seed(0)
    PAD = -1.0

    def make_targets_BL(B=2, L=8, pad_positions=(6, 7)):
        """Create (B,L) targets: 0=non-peak, >0=peak, -1=pad."""
        t = torch.zeros(B, L)
        # sprinkle a few peaks (values in [0.6, 1.0])
        t[:, 1] = torch.rand(B) * 0.4 + 0.6
        t[:, 3] = torch.rand(B) * 0.4 + 0.6
        # pads
        for p in pad_positions:
            t[:, p] = PAD
        return t

    def make_targets_BLC(B=2, L=8, C=3, pad_positions=(6, 7)):
        """
        Create (B,L,C) targets by broadcasting a (B,L) base across channels
        (so masking needs to expand correctly).
        """
        base = make_targets_BL(B, L, pad_positions)  # (B,L)
        t = base.unsqueeze(-1).expand(-1, -1, C).clone()
        # Make channel 1 slightly different to show per-channel variety
        t[..., 1] = torch.where(t[..., 1] > 0, (t[..., 1] * 0.85).clamp(0, 1), t[..., 1])
        return t

    def mask_stats(name, logits, targets, pad_value=PAD):
        valid = (targets != pad_value)
        nonpeak_mask = valid & (targets == 0)
        peak_mask    = valid & (targets > 0)

        m_nonpeak = _expand_like(nonpeak_mask, logits)
        m_peak    = _expand_like(peak_mask, logits)

        print(f"\n[{name}]")
        print(f"  logits.shape  = {tuple(logits.shape)}")
        print(f"  targets.shape = {tuple(targets.shape)}")
        # Previews (first batch)
        if targets.dim() == 2:  # (B,L)
            print(f"  targets[0,:] preview: {targets[0]}")
        else:  # (B,L,C)
            print(f"  targets[0,:,0] ch0 preview: {targets[0,:,0]}")
            print(f"  targets[0,:,1] ch1 preview: {targets[0,:,1]}")
        # Mask counts after EXPANSION (these define denominators)
        print(f"  #non-peak elems used = {m_nonpeak.sum().item():.0f}")
        print(f"  #peak elems used     = {m_peak.sum().item():.0f}")

    # =========================
    # Case A: (B, L)
    # =========================
    B, L = 2, 8
    logits_BL  = torch.randn(B, L)                 # raw scores
    targets_BL = make_targets_BL(B, L)             # 0, >0, and -1 pads

    mask_stats("BL", logits_BL, targets_BL, pad_value=PAD)

    loss_BL = calculate_loss(
        logits_BL, targets_BL, pad_value=PAD, alpha=1.0, gamma=1.0
    )
    acc_BL = accuracy_percentage(
        logits_BL, targets_BL, pad_value=PAD, peak_thresh=0.5
    )

    print(f"  loss_BL = {loss_BL.item():.6f}")
    print(f"  acc_BL  = {acc_BL:.2f}%")

    # =========================
    # Case B: (B, L, C)
    # =========================
    B, L, C = 2, 8, 3
    logits_BLC  = torch.randn(B, L, C)             # raw scores with channels
    targets_BLC = make_targets_BLC(B, L, C)        # broadcasted targets + tweaks

    mask_stats("BLC", logits_BLC, targets_BLC, pad_value=PAD)

    loss_BLC = calculate_loss(
        logits_BLC, targets_BLC, pad_value=PAD, alpha=1.0, gamma=1.0
    )
    acc_BLC = accuracy_percentage(
        logits_BLC, targets_BLC, pad_value=PAD, peak_thresh=0.5
    )

    print(f"  loss_BLC = {loss_BLC.item():.6f}")
    print(f"  acc_BLC  = {acc_BLC:.2f}%")