search

pUmpKin-Co / offical-IndexNet

This is the official implementation of the IndexNet.
Apache License 2.0
10 stars 1 forks source link

Use with custom model #6

Open simonebonato opened 7 months ago

simonebonato commented 7 months ago

Hi all, I have recently done an implementation of BYOL to do the pretraining of ResNet50 for a semantic segmentation task, but was not too satisfied with the results, and then I bumped into your paper and repo.

What is the fastest way to try and incorporate this with a custom model and dataset? :)

Great job for the results you obtained!

pUmpKin-Co commented 7 months ago

Hi~Sorry for the late reply. For the custom backbone (maybe?), make sure the backbone output list that contains feature map (usually four stage features map). Then, configure the config file according to your feature dimension, import the custom backbone, and replace the line 188 in IndexNetModel.

For the dataset, change the config file according to your image size (config.data.size) and the spatial size of last feature map (config.network.output_size), and then write your data loading pipline similar to class Potsdam.

Hope this can help you.

pUmpKin-Co commented 7 months ago

Moreover, you can refer to our new work CMID with improved pretraining strategy and outstanding preformance.