about "ckpt_path" and "data_path" for running the code

[zhongyi-zhou]

Hi! I am currently working in a similar field and I'm really excited to find this project.

I find I cannot run the code probably due to these two variables: https://github.com/ArantxaCasanova/ralis/blob/83325f318621b2b856a91cf1b26d1d16e4558657/scripts/launch_train_ralis.sh#L3-L4

I have tried replacing the following code into the path of the model you provided in google drive, it seems not working.

ckpt_path='/home/casanova/scratch/ckpt_seg'

Also, could you briefly introduce how data_path works in the project? It seems that the python file does not use this argument.

[ArantxaCasanova]

Hi, you have to replace data_path by the path that you downloaded your datasets into. The datasets need to be saved in a structure such as this one: <data_path>/datasets/camvid <data_path>/datasets/gta5 <data_path>/datasets/cityscapes (Note that you only need to download the datasets you are going to use).

With respect to ckpt_path, it is the path where you have saved the downloaded models. Following the structure of the pretrained models I provided, let's put an example: user_name/ralis/cs/<exp_name> In this case, ckpt_path='user_name/ralis/cs'.

Additionally, you will use the experiment name (--exp-name <exp_name>) in the scripts. exp_name will be the name of the folder where your experiment will be saved. You can run this with launch_test_ralis.sh. In this folder, you will save the weights of the segmentation model trained on data that RALIS feeds it with its active learning policy. For the other variables, exp-name-toload is name of the pretrained network for the segmentation network on GTA5 (also provided in the pretrained models), and exp-name-toload-rl is the name of the already trained RALIS model, in case you want to test it and train a segmentation network with it.

The script does use the data_path argument. More specifically, you can find it here.