TensorFlow Liteのエコシステムはどんどん強化されている

@Vengineerの戯言 : Twitter
SystemVerilogの世界へようこそ、すべては、SystemC v0.9公開から始まった 

TLから

tensorflow model-optimizationのロードマップが更新された。

・[WIP] Post training quantization for (8b) fixed-point RNNs
・[WIP] Quality and performance improvements to post training dynamic-range quantizationhttps://t.co/nTlRozZ4lY

— nb.o (@Nextremer_nb_o) 2020年8月7日

これだけいろいろロードマップあって、改善しているのって、TensorFlow Liteだけ？

 引用します。

Quantization

• Post training quantization for dynamic-range kernels -- Launched
• Post training quantization for (8b) fixed-point kernels -- Launched
• Quantization aware training for (8b) fixed-point kernels and experimentation for <8b -- Launched
• [WIP] Post training quantization for (8b) fixed-point RNNs
• Quantization aware training for (8b) fixed-point RNNs
• [WIP] Quality and performance improvements to post training dynamic-range quantization

Pruning / Sparsity

• During-training magnitude-based weight pruning -- Launched
• Sparse model execution support in TensorFlow Lite -- WIP

こちらは、Latency完全っぽい

Weight clustering

• During-training weight clustering
• Cascading compression techniques
• [WIP] Additional support for combining different compression techniques. Today, users can only combine one during-training technique with post-training quantization. The proposal is coming soon.

Compression

• [WIP] Tensor compression API