Fine-tuned CRAFT model works much slower on CPU than default one.

[romanvelichkin]

I fine-tuned CRAFT model according to this guide: https://github.com/JaidedAI/EasyOCR/tree/master/trainer/craft

But this model works 5 times slower than default model 'craft_mlt_25k' on some server CPUs (on some CPUs speeds are same). What can it be? Is 'craft_mlt_25k' quantized in some way?

[romanvelichkin]

I solved this speed problem by quantizing my CRAFT model into float16.

However 'craft_mlt_25k' is in float32.