code about mobilenet_v2

[H-cool]

Hello, I'm encountering some issues while trying to reproduce the MobileNet-v2 results from the paper. Specifically, I'm having trouble with quantizing very small models, such as those with 0.9M or 1.1M parameters. What is the quantization bit width setting for these models? I see in the code that the search range is 2 to 4 bits. If 1 bit is not included in the search range, the model size cannot be reduced below 1M, and it cannot be used with BRECQ for data calibration. Could you please explain how the search space for MobileNet-v2 is selected for small models, and how it is used in conjunction with BRECQ for quantization?

[BobxmuMa]

Hello, I'm encountering some issues while trying to reproduce the MobileNet-v2 results from the paper. Specifically, I'm having trouble with quantizing very small models, such as those with 0.9M or 1.1M parameters. What is the quantization bit width setting for these models? I see in the code that the search range is 2 to 4 bits. If 1 bit is not included in the search range, the model size cannot be reduced below 1M, and it cannot be used with BRECQ for data calibration. Could you please explain how the search space for MobileNet-v2 is selected for small models, and how it is used in conjunction with BRECQ for quantization?

Thank you for your interest in our work and for raising this question.

For extremely small models (0.9M-1.1M parameters), you can experiment with different bit-width configurations to find the optimal setting for your specific use case. The 2-4 bit search range in our code represents a balance between model size and accuracy for general scenarios.

Regarding 1-bit quantization: compressing to 1-bit typically causes significant accuracy degradation for CNN models. Standard quantization methods like BRECQ are not designed to handle such extreme compression without substantial accuracy loss. If you need to achieve sub-1M model sizes with 1-bit quantization, you'll likely need specialized techniques specifically designed for binary neural networks, such as:

• Binary Neural Networks (BNN) with specialized training procedures
• Knowledge distillation for binary models
• Architecture modifications tailored for 1-bit weights

I recommend exploring recent research on 1-bit quantization methods for CNNs if you need to push below the 1M parameter threshold while maintaining acceptable accuracy.

Feel free to experiment with extending the search space to include 1-bit in your own experiments, but be prepared for the accuracy trade-offs.

Thanks again for your attention to our work!