search

mit-han-lab / tinyengine

[NeurIPS 2020] MCUNet: Tiny Deep Learning on IoT Devices; [NeurIPS 2021] MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep Learning; [NeurIPS 2022] MCUNetV3: On-Device Training Under 256KB Memory
https://mcunet.mit.edu
MIT License
757 stars 127 forks source link

Could you tell me model hyper-parameter in MCUNet paper #70

Open gorangeeeeeeeeeeeeeeeeeeeeeeeeeee opened 1 year ago

gorangeeeeeeeeeeeeeeeeeeeeeeeeeee commented 1 year ago

Hello, I found your MCUNet paper interesting and I am writing to ask for information on the hyper-parameters used in the implementation of the models used in the experiments. While I believe that you will eventually provide the training code for MCUNet or ProxylessNas, at least, I would appreciate it if you could let me know what hyper-parameters you used for the MobileNet V2 that was used in the MCUNet experiments. Specifically, I am interested in the following models: MobileNet w0.35-r64 in MCUNet V1 and MobileNet w0.35-r144 in MCUNet V2, as well as MobileNet V2, MobileNetV2-RD, and MobileNetV2 (Non-overlap) in Table 5. It would be great if you could provide the information in fp32.

image image image

I am looking forward to a prompt response.

meenchen commented 1 year ago

Hi @gorangeeeeeeeeeeeeeeeeeeeeeeeeeee,

For model-specific details, please check our MCUNet repo: https://github.com/mit-han-lab/mcunet

gorangeeeeeeeeeeeeeeeeeeeeeeeeeee commented 1 year ago

Thanks @meenchen I have one more question. For the model labeled as "non-overlap", did you train it separately from scratch or can we achieve the same results by using a pre-trained model?