Budget-aware pruning focuses on compressing the network subject to the prescribed explicit resource constraints, including targets in following table.
| Budgets | Explanation |
|---|---|
Sparsity |
Percentage of zero weights including layer, block, global network wise |
FLOPs |
Floating-Point Operations per Second |
Latency |
Execution time on specific hardware |
Energy |
Power consumption for running model |
Array |
Structured collection, like memory cells or processors |
Compiler |
Translates code for computer execution |
For example optimizing the budgets of execution latency, energy, array or compiler will involve a co-optimization between model compression including quantization and pruning, compilation/mapping and hardware, which offers new opportunities, as shown in following figure.
