Fixed point input to model

[rf-ml]

When I run part7 of the tutorial I can see floats going into the model as input. In my situation I've already got some upstream processing that has created Q(1.31) / fixed32_31 data (so 1 integer sign bit and 31 other bits). I would like to avoid having to convert the existing input data I've got to floating point. In future as well I may be dealing with 8bits as well.

What I'm not sure is:

1. How to train a Qkeras model with fixed point input as I don't see a specific quantised aware "input" layer https://github.com/google/qkeras. I'm guessing I just put the data converted to fixed point straight in during training rather than putting float in.
2. During training if I try putting fixed point in on input layer and straight into a QConv2DBatchnorm with kernel_quantizer=f"quantized_bits({10},{2},alpha=1)" then I get a loss of nan. No issue though with floating point.
3. For the HLS config I'm guessing I can just set something like hls_config_q['LayerName'][**input_layer**]['Precision'] = 'ap_fixed<32,1>' (but so far not managed to successfully train a model to get to this stage with a fixed point input).
4. If I assume I'm meant to just train the QKeras model with floating point input, then in the hsl_config set the input layer to be ap_fixed<32,1> I get a different error when I try to test the model with hls_model_q.predict(np.ascontiguousarray(X_fixed))
Exception: Invalid type (int32) of numpy array. Supported types are: single, float32, double, float64, float_.. So I'm not sure how I should test the model before deploying if this approach even works.

I'm new to FPGA and fixed point so just knowing that this is possible would be a great help. Also if people do normally change the input to fixed point or this is a strange thing to do.

Thanks