Predicting Training Batch Time for ResNet Image Classification: A Lightweight Model for Media Workloads

This paper presents a data-driven approach for predicting the execution time of hyperparameter optimization (HPO) trials in heterogeneous computing environments. Differences in node performance often lead to imbalanced workloads and reduced efficiency in distributed HPO. To address this, we propose a Multi-Layer Perceptron (MLP) regression model that estimates per-batch execution time using trial hyperparameters, model size, and hardware characteristics such as CPU/GPU specifications and frequencies. The proposed model achieves high accuracy ( \(r = 0.993\) ) and low inference latency ( 0.66 ms), making it suitable for real-time scheduling applications. Our workflow includes data collection, feature engineering, and sample balancing to ensure robustness across diverse nodes. Experimental results demonstrate the model’s effectiveness and potential integration into HPO schedulers for cost-aware, adaptive resource allocation in heterogeneous clusters. We evaluate our predictor on ResNet-based image-classification trials (e.g., ResNet-18, ResNet-50), so the results reflect realistic media workloads.