The following is a list of suggestions to optimize pruning results. Following these guidelines has been found to help developers achieve higher pruning ratios and reduce accuracy loss.
- Use as much data as possible to perform model analysis. Ideally, you should use all the data in the validation dataset, but this can be time consuming. You can also use partial validation set data, but you need to make sure at least half of the dataset is used.
- During the fine-tuning stage, experiment with a few hyperparameters, including the initial learning rate and the learning rate decay policy. Use the best result as the input for the next iteration.
- The data used in fine-tuning should be a subset of the original dataset used to train the baseline model.
- If the accuracy does not improve sufficiently after conducting several fine-tuning experiments, try reducing the pruning rate parameter and then re-run pruning and fine-tuning.