Comparing Iterative Pruning and One-Step Pruning - 3.5 English

Vitis AI User Guide (UG1414)

Document ID
Release Date
3.5 English

Two primary approaches to pruning neural networks are Iterative Pruning and One-Step Pruning, each offering unique strategies for achieving model sparsity while preserving accuracy. Iterative pruning progressively trims model parameters while retaining accuracy, employing multiple iterations to achieve the desired sparsity level. In contrast, One-Step Pruning rapidly identifies and fine-tunes the most promising subnetwork, making it an efficient choice for achieving model sparsity with high potential accuracy in a single step.

A comparison of these two methods is shown in the following table.

Table 1. Iterative vs. One-step Pruning
Criteria Iterative Pruning One-step Pruning
Prerequisites - BatchNormalization in network
Time taken More than one-step pruning Less than iterative pruning
Retraining requirement Required Required
Code organization Evaluation function Evaluation function

Calibration function