There has been a significant surge in the adoption of deep learning methods across various industrial and research domains. This increase requires efficient training strategies to improve convergence speed, robustness, and overall model optimization. The paper presents a benchmark analysis that highlights the benefits of learning rate (LR) schedulers. The benchmark has been conducted using four training methods that integrate the LR finder, the LR scheduler, and the LR optimizers like Adam. The first three training methods use the LR finder to compute the correct range and then use, respectively, the AdaptiveCycle LR scheduler, the OneCycle LR scheduler, and the static LR for training; the last training method uses only a static LR. The OneCycle LR scheduler has been used as it balances exploration and exploitation during training by adjusting the LR cyclically. Additionally, we introduce a novel scheduling method, AdaptiveCycle, that modifies the LR based on validation loss plateaus over a cyclic LR curve, further optimizing the training process. We have performed the benchmark using diverse time-series data sets from the UCR and UCE repositories. The benchmark is performed across eleven deep learning models, simple such as LSTM, FCN, and complex models such as Inception-time, XCM, and LSTM-FCN. The accuracy, F1 score, and computational time allow us to highlight the adaptability of training methods across various datasets and models. The findings emphasize the advantages of employing LR schedulers for efficient training. The AdaptiveCycle scheduler achieves optimal performance across different models and datasets. The benchmark results will encourage the adoption of the LR scheduler in research and industry.