Going Beyond Classification Accuracy Metrics in Model Compression