search

mit-han-lab / tiny-training

On-Device Training Under 256KB Memory [NeurIPS'22]
http://tinytraining.mit.edu
MIT License
438 stars 60 forks source link

Doubt in Algorithm Part : how you have prepared three models in int8 format: mbv2-w0.35, mcunet-5fps, proxyless-w0.3 #5

Open chinmaykhattar opened 1 year ago

chinmaykhattar commented 1 year ago

Hi,

In Algorithm part , you have used the pkl file for building the model (For eg mcunet-5fps.pkl in assets/mcu_models). Could you please let me know how you have create the model and how you have convert the model into pickle quantized (int 8 bit) format.

Thanks, Chinmay

silvagal commented 3 months ago

I would like to know too. I was looking the structure of mbv2-w0.35.pkl and what I found was that it is a dictionary, that contains the following keys: first_conv, blocks, feature_mix and classifier.

first_conv seems to be a dictionary with 10 elements, probably containing information about the first convolutional layer of the model. blocks is a list with 17 elements, likely representing the intermediate blocks or layers of the neural network. feature_mix is set to None. And classifier is a dictionary with 6 elements, which likely correspond to the final classification layer of the model.

I didn't find any function library that saves the model like that, I tried to make my own function, but I got some problems. Did you discover how the authors have created the model and how they convert it into pickle quantized format?