pyTorch Boilerplate

Generic code template for pytorch project.

Install

Required python >= 3.8 and pip

pip install -r requirements.txt
echo "run_name: tmp" >> settings.yaml

Template quick start

1. Remove the .idea/ directory if you don't use pyCharm IDE.
2. In utils/logger.py at the last line, set the current project name in: logger_name='<project name>'.
3. Create your own dataset in datasets/ or remove the folder if you use a dataset already implemented in pytorch.
4. Create your own neural network in networks/.
5. Replace the dataset and the network in main.py.
6. Remove / add settings in utils/settings.py
7. Run run.py
8. Update the title / description of this README file, and remove this section.

Run

Run a single training + testing based on the current settings:

python3 run.py

Run a list training + testing based on the current settings and the planner configuration:

python3 runs_planner.py

Settings

An extendable list of settings with their default values can be found in the file utils/settings.py.

Their values can be overridden by:

• the local setting file (./settings.yaml)
• the environment variables
• the arguments of the command line (--help to see the list)

File structure

• datasets/: Directory that contains the list of available datasets classes that extend the pytorch Dataset class.
• networks/: Directory that contains the list of available neural network that extend the pytorch Module class.
• out/: Directory where named run output will be saved. A subdirectory will be generated for each new run.
• plots/: Directory that contains all plotting scripts.
• utils/: Directory that contains utility scripts, using them is optional, but they simplify some tasks.
  • logger.py: Class wrapper to handle console and file logging.
  • metrics.py: Script to compute and save metrics about the network and the run.
  • output.py: Script to manage the run output (log file, images, results, ...).
  • planner.py: Class to create an iterable list run settings.
  • progress_bar.py: Class to handle visual progress bar.
  • settings.py: Dataclass to store and load settings values. Create a singleton at the first import.
  • timer.py: Class wrapper to handle, print and save timers.
• main.py: Simple script to set up dataset and neural network before to start a run.
• run.py: Full script that contains the run steps logic.
• runs_analyse.py: Script to aggregate and plot result from several runs.
• runs_planner.py: Script to start a list of runs with different settings one by one.
• settings.yaml: Local settings file to override default values. This file shouldn't be committed.
• test.py: Script that contains the network testing logic.
• train.py: Script that contains the network training logic.