After properly setup the environment and dataset, you may now use several command to start the training process.
All options and their function could be found in the src/lib/opts.py file. Most of the value the the opts.py file is already given an default value, which allow you to simple type the below command to start training the CenterSAM model:
#cd src/lib
python main.py ctdet
you can easily adjust most of training option by command line, for example, if you intend to use res_18 architecture instead of hourglass, with batch size equal to 64 and l2 regression loss, the following command should be use for the training process:
python main.py ctdet --arch res_18 --batch_size 64 --reg_loss l2
There are many option including conv layer channels for output head, learning rate and so on, explore it by dive into the opts.py file if you want.
The trained model will be saved to the path: CenterSAM/exp/ctdet/TissueNet_ctdet/model_best.pth or CenterSAM/exp/ctdet/TissueNet_ctdet/model_last.pth (here we use example of ctdet task and TissueNet dataset). For convenience， we will use (Model_path) as a refer to the CenterSAM/exp/ctdet/TissueNet_ctdet/model_best.pth for the following tutorial.
let's say you have multiple images want to test that located under CenterSAM/demo_images folder, You may type the below command to visulize the effect of trained models：
python demo.py ctdet --demo CenterSAM/demo_images --load_model CenterSAM/exp/ctdet/TissueNet_ctdet/model_best.pth
The --demo argument accept a set of images or a single image.
This demo.py script allow you to visulize the result of detection stage, which shows the quality of automatically generated prompts and also produce the *json file of prompts. This json file would be the input file used for the segment stage.
In order to make it easier to use and facilitate subsequent development, we split the segment stage scripts and code and put them in CenterSAM/SAM/notebooks. You can find the scripts and code we use for produce the evaluation results and figure in that folder.
Basiclly, here we first generate a json file which storage all the prompts information from the detectoion stage, then generate final instance segment result with the help of SAM. After getting prompts json file from above, you can now run through notebook code in the CenterSAM/SAM/notebooks folders accordingly (dependding on the dataset you are testing). The evaluation and visualization of final instance segment result is all inside the ipynb files