AutoSlim is an automated approach to finding optimal channel configurations for slimmable networks — instead of using uniform width multipliers (0.25×, 0.5×, etc.), AutoSlim searches for the best per-layer channel allocation under a given computation budget.

How AutoSlim Works

• Non-Uniform: Different layers may have different optimal widths — AutoSlim finds per-layer widths.
• Greedy Slimming: Start from the full network and greedily prune channels layer-by-layer, removing the least important ones.
• Evaluation: After each pruning step, evaluate accuracy to guide which channels to remove next.
• Pareto Frontier: Produces a set of architectures along the accuracy-FLOPs Pareto frontier.

Why It Matters

• Better Than Uniform: Non-uniform width allocation outperforms uniform scaling at the same FLOP budget.
• Automated: No manual architecture design — the search finds optimal per-layer widths.
• Efficient Search: Greedy slimming is much faster than full NAS — can complete in one training run.

AutoSlim is smart channel allocation — automatically finding the best per-layer width configuration for optimal accuracy within any computation budget.