INT8 Quantization

[sriram487]

I'm working on quantizing the model, and while the accuracy remains stable when using FP16 precision, the model's performance degrades significantly when quantizing to INT8 or FP8. The issue could be related to the calibration data. For INT8 extrinsic quantization, it's typically recommended to use at least 500 images for CNNs. However, I'm not sure about the minimum number of samples required and the appropriate batch size for the calibration data, especially the batch size changes during registration and tracking, which could affect the calibration parameters.

What data should be used for calibration? Since the model is trained on synthetic data, is it ok to use a portion of that synthetic training data for calibration?

[wenbowen123]

I haven't tried quantizing. But I'm also curious. You are welcome to report back your findings : )

[A7eNg]

+1 waiting for the quantization result