loss spike

Loss spikes during training indicate instability that can derail optimization, typically caused by learning rate issues, bad data batches, gradient explosions, or numerical precision problems, requiring immediate investigation and intervention. Symptoms: loss suddenly increases by orders of magnitude; may recover or may diverge completely. Common causes: learning rate too high (gradients overshoot), corrupted/mislabeled data in batch, gradient explosion (especially in RNNs), and NaN/Inf from numerical issues. Immediate fixes: reduce learning rate, add gradient clipping (clip by norm or value), and check for NaN in gradients. Data investigation: identify which batch caused spike; check for outliers, encoding issues, or corrupted examples. Gradient clipping: cap gradient magnitude before update (torch.nn.utils.clip_grad_norm_); prevents single large gradient from destroying weights. Learning rate schedule: warmup helps avoid early spikes; cosine or step decay prevents late instability. Mixed precision: loss scaling in FP16 training prevents underflow; check AMP scaler if using mixed precision. Checkpoint recovery: if training destabilizes, rollback to earlier checkpoint; may need different hyperparameters to proceed. Batch size: very small batches have high variance; may cause sporadic spikes. Detection: monitor loss in real-time; alert on anomalous increases. Prevention: proper initialization, normalization layers, and conservative learning rates. Loss spikes require immediate diagnosis before continuing training.