How to customize the position of the quantizer?

[hcqylymzc]

hi
Can I manually declare a quantizer through aimet_torch.quantization, and then use this quantizer to customize where my network should be quantized?The parameters are the same. If so, can you give me an example?
For example, I use three Q.affine.QuantizeDequantize to define the quantization nodes of input, output, and convolution. I can insert quantization nodes in the entire model through a series of custom quantizers. But how do I cooperate with QuantizationSimModel after insertion?

[quic-mtuttle]

Hi @hcqylymzc, could you clarify a bit about what exactly your use-case is? If you mean you just want to replace the quantizers placed during QuantizationSimModel instantiation with some new Q.affine.QuantizeDequantize objects, then some workflow like this should work:

# 1) Instantiate Quantsim with whatever args you want
sim = QuantizationSimModel(model, dummy_input, ...)

# 2) Replace whatever quantizers you want
sim.model.conv1.param_quantizers["weight"] = Q.affine.QuantizeDequantize(shape, bitwidth, symmetric=True)
sim.model.conv1.output_quantizers[0] = Q.affine.QuantizeDequantize((), bitwidth, symmetric=False)

# 3) Continue with regular workflow (e.g., compute_encodings and export) as normal
sim.compute_encodings(forward_pass_callback, forward_pass_callback_args)
sim.export(path, filename_prefix, dummy_input)