While training large models with PyTorch, sometimes the model does not update in certain epochs. All detections are incorrect, and the precision, recall, and F1-score are inconsistent. What should I do to resolve this issue during updates? This problem does not seem to occur with TensorFlow.