Manual Tuning

If the accuracy after sparsity does not meet the requirements, refer to this section to perform tuning.

If the accuracy of the retrained sparsified model is not as expected, you can modify the simplified configuration file (see Simplified QAT Configuration File for details) and perform sparsity and training again. The common methods are as follows:

• Modify the update interval (for calculating which elements are reserved) by setting the update_freq parameter in the simplified configuration file. You can change the update interval to a positive integer and perform sparsity for debugging again. Generally, the smaller the update interval, the higher the accuracy gain.
• Skip certain layers in sparsity by setting the regular_prune_skip_layers parameter in the simplified configuration file.