How does FINN deal with the float-point layer by default

[maruAYA]

Code abstract:

net = nn.Sequential(
nn.Linear(4, 4)
qnn.QuantLinear(4, 4, bias=True, input_quant=Int8ActPerTensorFloat)
)
ckpt_dir = './ckpt'
os.makedirs(ckpt_dir, exist_ok=True)
qonnx_fp = 'linear.qonnx.onnx'
qonnx_fp = os.path.join(ckpt_dir, qonnx_fp)
_ = export_qonnx(
    module=net,
    args=torch.randn(1, 4),
    export_path=qonnx_fp,
    disable_warnings=True,
    input_names=['x'],
    output_names=['y'],
    # dynamic_axes={
        # 'x':{0: 'batch'},
        # 'y':{0: 'batch'}
    # }
)
qonnx_cleanup(qonnx_fp, out_file=qonnx_fp)
model = ModelWrapper(qonnx_fp)
model.set_tensor_datatype(model.graph.input[0].name, DataType["FLOAT32"])
model = model.transform(ConvertQONNXtoFINN())
finn_fp = qonnx_fp.replace('qonnx', 'finn')
model.save(finn_fp)

model_for_sim = model
finnonnx_in_tensor_name = model_for_sim.graph.input[0].name
finnonnx_out_tensor_name = model_for_sim.graph.output[0].name
finnonnx_model_in_shape = model_for_sim.get_tensor_shape(finnonnx_in_tensor_name)
finnonnx_model_out_shape = model_for_sim.get_tensor_shape(finnonnx_out_tensor_name)
finnonnx_model_in_dt = model_for_sim.get_tensor_datatype(finnonnx_in_tensor_name)
finnonnx_model_out_dt = model_for_sim.get_tensor_datatype(finnonnx_out_tensor_name)

verif_fp = finn_fp.replace('finn', 'verified')
model_for_sim.save(verif_fp)

cfg_estimates = build.DataflowBuildConfig(
    output_dir          = estimates_output_dir,
    mvau_wwidth_max     = 4,
    target_fps          = 10000,
    synth_clk_period_ns = 10.0,
    # fpga_part           = "xc7z020clg400-1",
    fpga_part           = "xczu9eg-ffvb1156-2-e",
    steps               = build_cfg.estimate_only_dataflow_steps,
    generate_outputs=[
        build_cfg.DataflowOutputType.ESTIMATE_REPORTS,
    ]
)
build.build_dataflow_cfg(finn_fp, cfg_estimates)

We apply several float-point layers with FINN and the generated IP by default set-up is included in here. The input width and the weight width are automatically truncated to 8-bit (fixed-point I guess). We are wondering what the data type of the 8-bit input of the MVU is and whether FINN should perform such transfomation by default.