Marry Dijan team's repository for Lumen Data science 2022.

Implementation of architecture module, train.py and inference.py scripts and various models.

The easiest and most straightforward way to set things up is using Docker.

To install Docker, follow these instructions: https://docs.docker.com/get-docker/

Now build a Docker image from Dockerfile:

docker build -t marry_dijan -f LUMEN_DS_2022/model3/Docker/Dockerfile_cpu model3/Docker/

Now run new container and mount your workspace directory to container's workspace directory:

docker run -it --rm -v [FULL PATH TO YOUR WORKSPACE]:/workspaces marry_dijan /bin/bash

Now a new bash instance should be running inside the created container.

Install architecture module in development mode:

cd LUMEN_DS_2022/model3/src
pip install -e .

Now you are ready to run either train.py or inference.py from model3/src. Usage instructions are explained below.

Trainer refers to a class which has .train() method implemented which dictates the training process. TrainerDefault is the default training process.

For creating new Trainer class follow these steps. Let's suppose that we want to create a training process for a neural network we have just implemented inside model_resnet34_fc11.py file. We are not sure what hyperparameters to use, so we will create two training processes that differ only in hyperparameters. Each training process gets it's own .py file, and those training processes that differ only in hyperparameters should be inside the same parent directory.

1. Create parent directory src/architecture/train/trainers_resnet34_fc11/ and inside this directory create files trainer_resnet34_fc11_v1.py and trainer_resnet34_fc11_v2.py.

2. * Inside each of these files create create a new class with name signature Trainer[details], in our case Trainer_ResNet34_FC11_v1 and Trainer_ResNet34_FC11_v2 respectively.

3. Let the created classes inherit from default trainer:

   class Trainer_ResNet34_FC11_v1(architecture.train.TrainerDefault)

4. Update all the hyperparameters and modify the training process.

5. * Declare the newly created class by adding the following line inside architecture.train.__init__.py:

   from .trainers_resnet34_fc11.trainer_resnet34_fc11_v1 import Trainer_ResNet34_FC11_v1.

   This step is very important in order that main.py would be able to find the correct trainer class.

6. * Choose a new trainer by specifying --trainer_name Trainer_ResNet34_FC11_v1 when invoking main.py script.

Steps marked with * probably cannot be skipped.

The term 'inference' means generating predictions of the trained model on some input images. For running inference, run inference.py script.

In the following example we will generate predictions of one model.

• Parameters is located at /workspaces/parameters_id-25.prms.

• Dataset is located at:

  /workspaces/Dataset/
      |- data/
      |- 2022-mar-31_data_val.csv
  

• (Empty) output directory is located at /workspaces/id_25_out.

Now run the following:

python3 src/inference.py
   --trainer_name=Trainer_ResNet50_FC3_v1 \
   --parameters_path=/workspaces/parameters_id-25.prms \
   --dataset_root=/workspaces/Dataset \
   --target_csv=2022-mar-31_data_val.csv \
   --output_dir=/workspaces/id_25_out \
   --device=cpu \
   --batch_size=1 \
   --num_workers=4

Result is /workspaces/id_25_out/output.csv where all the predictions are generated in columns mo_latitude, mo_longitude.

To run a process on a remote machine with ssh: ssh name@host "python3 /path/to/script.py --param1 val1 > /path/to/stdout 2>&1 &"

Add this configuration to .vscode/launch.json for a debug training.

{
   "name": "Training with train.py",
   "type": "python",
   "request": "launch",
   "program": "src/train.py",
   "console": "integratedTerminal",
   "justMyCode": true,
   "args": [
         "--trainer_name=Trainer_Debug_V1",
         "--output_dir=outputs/debug-output-3",
         "--train_csv=2022-mar-31_data-debug_train.csv",
         "--val_csv=2022-mar-31_data-debug_val.csv",
         "--dataset_root=/workspaces/Dataset",
         "--device=cpu",
         "--run_mode=d"
   ]
}