Quantization tool supporting Conv and MatMul nodes #1892
Conversation
|
|
| # Load the onnx model | ||
| model = onnx.load('path/to/the/model.onnx') | ||
| # Quantize | ||
| quantized_model = quantize(model, per_channel=False, quantization_mode=QuantizationMode.IntegerOps_Dynamic) |
There was a problem hiding this comment.
Static mode involves more inputs right? like the quantization params… can you include an example for static mode too...
| max_range = max(abs(rmin), abs(rmax)) | ||
| scale = (float(max_range)*2) / quantize_range | ||
| zero_point = 0 | ||
| quantized_data = (np.asarray(data) / scale).round().astype('b') #signed byte type |
There was a problem hiding this comment.
Does this round half to even?
| S: scale | ||
| z: zero point | ||
| ''' | ||
| rmin = min(data) |
There was a problem hiding this comment.
what happens when range does not include 0? For example is the range is 2-10 then in this case we dont have a unique representation for 0. Can you do rmin = min(min(data), 0) and similar for max...
| scale_name = weight.name + '_scale' | ||
| zero_point_name = weight.name + '_zero_point' | ||
|
|
||
| # Remove existing weight initializer |
There was a problem hiding this comment.
What happens when this input is also being used by another node? This condition should be checked.
Quantization tool for converting an onnx model into quantized onnx model. Currently only conversion of Conv and MatMul nodes is supported.