Mathieu Poumeyrol
Mathieu Poumeyrol
Hey, thanks for your interest. You should give a try to the NNEF serialization. It's significantly faster to load and optimise than an ONNX model.
The NNEF serialization is a step towards this, as you'll save the "decluttered" model. And decluttering account for the most expensive part of the the loading/declutter/optimize workflow (more than the...
Yeah, the "IR" is just tract-core with TypedModel with some optimized operators. Most operators will retain their "decluttered" form, because there is not much to gain in optimizing them, but...
Hello, thanks for your interest in tract. I believe this issue is gone with recent versions of tract. It did not reproduce with my top of tree, at least (tried...
It is a missing op problem. I'm not sure what the SplitV is meant to be, but it can maybe be remapped to one or more tract existing operators. Can...
There is no pro-active plan to augment tract support for tflite. PR are always welcome, and I don't mind binding operators myself when a missing tflite op can be re-expressed...
Thanks for the report, I'll try to have a look
I have merged a fix on main. Please be give it a shot. Also, be aware that tract Resize operator is weak and incomplete. It's a whole project of its...
"Quantized models" is a very overloaded terminology :) tract has some support for the QMatMul and QConv operators in ONNX, but ONNX lagged behind the SOTA for a lot of...
Hey, thanks for your interest! Could you have a look at #1358 ? I think by replacing the data_provider on the Onnx structure by your own implementation, you would be...