search

balezz / LacmusTflite

0 stars 1 forks source link

Quantized tflite model #4

Open balezz opened 2 years ago

balezz commented 2 years ago

Now we have tflite model without any optimizations. Please, add some optimizations and corresponding benchmarks for it.

dolmatov2000 commented 2 years ago

Added model quantization to float16, calculated benchmarks, but not sure, that received some significant and positive impact on speed

balezz commented 2 years ago

@dolmatov2000 Could you quantize model to int8 please? Also remove auc score from table and add inference time. Seems like it normal slower tflite on x86 platform https://github.com/tensorflow/tensorflow/issues/40183#issuecomment-641393739

dolmatov2000 commented 2 years ago

@balezz Added int8 quantization and it sets all the predictions to 0 class. Edited result table