tflite_micro_compiler icon indicating copy to clipboard operation
tflite_micro_compiler copied to clipboard
verified publisher icon cpetig

Future development ideas

Open cpetig opened this issue 5 years ago • 2 comments

This issue contains discussion about potential long term improvements of the compiler

cpetig avatar Jul 23 '20 16:07 cpetig

  • Find a way to make reshape a no-op (by implementing in-place operators inside tflite?) see https://github.com/cpetig/tflite_micro_compiler/issues/53
  • Find a way to support runtime calculated quantization (really nice feature, but sadly not yet supported by tf lite, see https://arxiv.org/pdf/1906.00532.pdf, especially figure 1 (this uses the combination of Quantize,QuantizedConv2D,RequantizationRange, Requantize...,Dequantize, as supported by graph transform), see also https://github.com/tensorflow/tensorflow/issues/38285 ). https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with-tensorflow/ also has good explanations of the methods involved.

cpetig avatar Jul 23 '20 17:07 cpetig

  • Deploy an LSTM example as a more complex test case (I am currently working on this) Update: I was finally able to get an LSTM defined, trained and converted. I will write about the process and how to navigate the many pitfalls in the coming days. Update: It works fine https://github.com/cpetig/tflite_micro_compiler/commit/2c21a364e9763e51706cf6f6b447ed908314e117

cpetig avatar Jul 23 '20 17:07 cpetig