TDA4VM: Can I use mixed precision on QAT(8-bit and 16-bit)?

Hi

I have been performing PTQ (mixed-precision mode) on the trained model using TiDL framework.

I feel calibration time and effort is too long to meet desired accuracy.

 So, I check the QAT on the edgeai-torchivsion.

I want to use mixed precsion for minimizing the losees of information.

For example, I want almost layers are quantized by 8-bit and

a few last layers or specific layers are quantized by 16-bit(depth estimation or achorfree detector or segmentation)

I find the bitwidth setting in QuantTrainModule.

Can I set mixed precsion(8-bit or 16-bit) on QAT as below?

I handle my custom multi-task model.

backbone, neck : 8-bit quantization(weight, activation)

last convolution layers in dectection header : 16-bit quantization(weight, activation)

last convolution layers in segmentation/depth header : 16-bit quantization(weight, activation)

if 'training' in args.phase:
    model = xnn.quantize.QuantTrainModule(model, per_channel_q=args.per_channel_q,
                histogram_range=args.histogram_range, bitwidth_weights=args.bitwidth_weights,
                bitwidth_activations=args.bitwidth_activations, constrain_bias=args.constrain_bias,
                dummy_input=dummy_input)

Thanks.