NVlabs / timeloop

Timeloop performs modeling, mapping and code-generation for tensor algebra workloads on various accelerator architectures.
https://timeloop.csail.mit.edu/
BSD 3-Clause "New" or "Revised" License
335 stars 101 forks source link

Channel pruning #186

Closed nivi1501 closed 1 year ago

nivi1501 commented 1 year ago

Hi, Is it possible to implement and analyze the effects of channel-wise, filter-wise and shape-wise pruning in Spareloop? Paper- An Efficient Hardware Accelerator for Structured Sparse Convolutional Neural Networks on FPGAs (Refer to Figure 1) Link- https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9130762&tag=1

nellie-wu commented 1 year ago

Yes, Sparseloop can model such sparsity patterns with the appropriate mappings, but such coarse-grained structured sparsity, e.g., channel-wise and filter-wise, can be more easily modeled as a change of tensor shape. For example, if channel-wise pruning is applied, the workload can be modeled still as a dense workload with decreased number of channels.

nivi1501 commented 1 year ago

Thanks a lot !