search

drprojects / superpoint_transformer

Official PyTorch implementation of Superpoint Transformer introduced in [ICCV'23] "Efficient 3D Semantic Segmentation with Superpoint Transformer" and SuperCluster introduced in [3DV'24 Oral] "Scalable 3D Panoptic Segmentation As Superpoint Graph Clustering"
MIT License
560 stars 72 forks source link

How to find `state_dict` in a `.ckpt` file ? #38

Closed hansoogithub closed 10 months ago

hansoogithub commented 10 months ago

sorry i am having trouble accessing the trained model weights of the PointSegmentationModule and especially the encoder and the decoder, i would like to inspect them and edit them how can i do it? thank you

cfg = init_config(overrides=[
    "experiment=kitti360",
    "ckpt_path=mycheckpoint.ckpt"
])
datamodule = hydra.utils.instantiate(cfg.datamodule)
datamodule.prepare_data()
datamodule.setup()
model = hydra.utils.instantiate(cfg.model)
model = model.load_from_checkpoint(cfg.ckpt_path, net=model.net, criterion=model.criterion)
model = model.eval().cuda()
model.state_dict() # error here
drprojects commented 10 months ago

Hi @hansoogithub

If you only want to inspect the state_dict, you can simply:

ckpt = torch.load("mycheckpoint.ckpt")

ckpt is a dictionary and you will find what you want in ckpt['state_dict'] if I am not mistaken. Otherwise just inspect the keys in ckpt and you should easily find it :wink:

PS: if you are interested in this project, don't forget to give our project a :star:, it matters to us !

hansoogithub commented 10 months ago

i gave this project a star, thank you sorry after trying your suggestion, based on the kitti360 checkpoint

is it possible to identify which transformer blocks are the encoders and decoders in that state_dict? which sections are the point segmentation module? do the kitti360 config files determine these?

drprojects commented 10 months ago

is it possible to identify which transformer blocks are the encoders and decoders in that state_dict? which sections are the point segmentation module?

To answer these questions, I invite you to have a look at the code, it is fairly commented and should help you in understanding how the project works. In particular, check out:

You want to understand how these two work to be able to parse whatever you are interested in from the state_dict.

do the kitti360 config files determine these?

The configs file rely on hydra and have a nested inheriting structure. You can find the model configurations. To understand how these work, being familiar with hydra and lightning-hydra will help.