TroubleShooting Documentation (MCT XQuant Extension Tool)¶

Overview¶

The Model Compression Toolkit (MCT) offers numerous functionalities to compress neural networks with minimal accuracy lost. However, in some cases, the compressed model may experience a significant decrease in accuracy. Fear not, as this lost accuracy can often be reclaimed by adjusting the quantization configuration or setup.

Outlined below are a series of steps aimed at recovering lost accuracy resulting from compression with MCT. Some steps may be applicable to your model, while others may not.

Quantization Troubleshooting for MCT[1]¶

1. Judgeable Troubleshoots

The following items are automatically identified by the XQuant Extension Tool.

Please read the following items indicated by XQuant Extension Tool, especially the Solution section.

• Outlier Removal

• Shift Negative Activation

• Unbalanced “concatenation”

• Mixed Precision with model output loss objective

2. General Troubleshoots

The following items are general troubleshoots for quantization accuracy improvement.

If quantization accuracy of your model does not improve after reading Judgeable Troubleshoots, please read the following items.

• Representative Dataset size and diversity

• Representative and Validation Dataset Mismatch

• Bias Correction

• Using more samples in Mixed Precision quantization

• Threshold selection error method

• Enabling Hessian-based Mixed Precision

• GPTQ - Gradient-Based Post Training Quantization

References¶

[1] Quantization Troubleshooting for MCT

[2] PyTorch documentation (v2.5)