The main goal of this baseline is to have a skeleton code base structure for pytorch lightning with some nice configurations. Some of the libraries and frameworks I will be usign are listed below.
pytorch
pytorch-lightning
clearml
rich
tqdm
split-folders
simple-parsing
I downloaded and use Animal-10 function directly. dataset to make this demo. Then I useds utils/reduce.py to reduce the datatset size to approx. 50 examples per category for quick testing.
I wrote custom dataloader to process and load this data located at utils/dataloader.py. It can be modified depending on the use case. We can also just use ImageFolder dataloader directly.
I defined some custom configuration inside the config folder.
The first one is args.py which contains the hyperparameters. In my baseline, I am using simple-parsing to parse the arguments due to its clean interface. For that purpose, I define a simple dataclass in the args.py file which we later use to aprse our arguments.
The config.py mainly consists of some paths configurations and logging behavior. I am also writing logs into .log files but you can configure it according to your requriements. In case you wan tto write logs, create a logs folder first in the main directory.
This is the main file of the project. FIrst of all, I am using albimentations to preprocess the data. Then I have some functions to create dataloaders.
After that we initialize clearML task. I am using clearML to track my experiments. Results will be shown at app.clear.ml.
Then we define our checkpoint and progress bar behavior. Initialize the Trainer and then fit the model.
I am not including any test script here. That behavior depends on the pplication. Also I am planning to create some web apps using this baseline, so I will be wrtigin test scripts there. If I have time, I will include it here later.
Update: I ended up writign the test script test.py. It will load the model and return the prediction results both in the CLI as well as in plot form.
Feel free to use these baselines in your projects.