Feature Request: Dequantization - convertion of int model to float model

[Shreeshrii]

@stweil You had mentioned at one point that it should be possible to finetune fast models. It will be useful to have this feature as many fast models use a smaller network size compared to best models, and hence the finetuned models would also be faster.

Is this something that can be included in 5.0.0?

[stweil]

I am afraid that would delay 5.0.0 further as I don't have the time to implement that until the end of November.

[stweil]

Technically the code would read the fast model with integer parameters and convert those parameters to float (or double) to get a best model which is then used for training.

[amitdo]

+1 for this feature request.

After 5.0.0...

[amitdo]

The term for converting from int model to float model is Dequantization / Dequantize.

https://www.tensorflow.org/api_docs/python/tf/quantization/dequantize

https://pytorch.org/docs/stable/generated/torch.dequantize.html