English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Zendy Plus

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Zendy Tools

Zendy Open

The exponentially large discrete search space in mixed-precision quantization(MPQ) makes it hard to determine the optimal bit-width for each layer. Previousworks usually resort to iterative search methods on the training set, whichconsume hundreds or even thousands of GPU-hours. In this study, we reveal thatsome unique learnable parameters in quantization, namely the scale factors inthe quantizer, can serve as importance indicators of a layer, reflecting thecontribution of that layer to the final accuracy at certain bit-widths. Theseimportance indicators naturally perceive the numerical transformation duringquantization-aware training, which can precisely provide quantizationsensitivity metrics of layers. However, a deep network always contains hundredsof such indicators, and training them one by one would lead to an excessivetime cost. To overcome this issue, we propose a joint training scheme that canobtain all indicators at once. It considerably speeds up the indicatorstraining process by parallelizing the original sequential training processes.With these learned importance indicators, we formulate the MPQ search problemas a one-time integer linear programming (ILP) problem. That avoids theiterative search and significantly reduces search time without limiting thebit-width search space. For example, MPQ search on ResNet18 with our indicatorstakes only 0.06 s, which improves time efficiency exponentially compared toiterative search methods. Also, extensive experiments show our approach canachieve SOTA accuracy on ImageNet for far-ranging models with variousconstraints (e.g., BitOps, compress rate). Code is available onhttps://github.com/1hunters/LIMPQ.

Mixed-Precision Neural Network Quantization via Learned Layer-wise Importance