README.md

Install

Clone the repository to somewhere, navigate into my_modules and enter:

pip install -e .

This will install the modules into your python environment.

Create a New Project

Navigate to somewhere you wanna create a new project in and enter:

create_project.py -p `your_project_name` -a `your_app_name`

For simplicity, assume your project name is project and your app name is app.

This will create a folder project in the directory where you issue the command above. The content of the directory is like:

  project
  |__ app
  |   |__ confs
  |   |   |__ __init__.py
  |   |   |__ config_1.py
  |   |
  |   |__ model
  |   |   |__ __init__.py
  |   |   |__ model.py
  |   |
  |   |__ dataset
  |   |   |__ __init__.py
  |   |   |__ dataset.py
  |   |
  |   |__ trainer
  |       |__ __init__.py
  |       |__ trainer.py
  |
  |__ scripts
  |   |__ clean.py
  |   |__ cont.py
  |
  |__ main.py

• Write Your Model

  You should create a class named Model in model.py and subclass torch.nn.Module. Refer to model.py in the templates to see what's needed in your model code.

• Write Your Dataset

  You should create a class named Dataset in dataset.py and subclass torch.utils.data.Dataset. Refer to dataset.py in the templates to see what's needed in your dataset code.

• Write Your Trainer

  The trainer is responsible for model checkpoint loading, input data parsing, loss computation, logging, checkpoint saving and so on. Most of the core parts are implemented within my_modules.modules.trainer. Your trainer should subclass this core trainer and override its interfaces. Refer to trainer.py in the templates to see what's needed in your trainer code.

• Write a Config File

  The config files are stored in the confs directory and named as config_1.py, config_2.py, ... config_[n].py, each representing one of the configuration (hyper-parameters, data paths, models to use and so on). The config is the core part throughout the whole pipeline.

  In this manner, you needn't modify the main part of your code in order to apply a new configuration. Rather, you just create a new config file and specify it when running the program. The config file is in fact a python module defining some global variables, and is dynamically loaded during runtime, passed among functions like a normal python object.

  It's encouraged to create a new config file if a new set of configuration is to be tested, while the pipeline (data reading, model processing, etc) remains almost the same. However, if the pipeline is changed to some extent but you're still doing the same task, it's encouraged to create a separate app in the same proj. To do this, simply copy the app_dir to your project dir, rename it to your new app name, and do the rest just like implementing your first app.

  A sample config file is created as you create a new project. You can modify, add or remove the variables at your need. But some of them are necessary so that you can just modify them.

key	description
PROJECT_ROOT_DIR	The project dir. By default it's the project dir you create. Leave it unchanged.
APP_DIR	The app dir. Leave it unchanged.
DATA_DIR	The directory to place your data (pretrained models, etc). Leave it unchanged.
MODEL_DIR	The directory to store trained model data. For example, model parameters, metric logs and so on. For config_n.py, the data is stored under MODEL_DIR/model_n. Leave it unchanged.
PRETRAIN_PATH	The pretrained model path.
STATE_DIR	The directory to store model parameters. By convention it's a subdirectory of MODEL_DIR. Leave it unchanged.
STATE_PREFIX	The name prefix of the saved model parameters. They're saved as {STATE_PREFIX}_{epoch_id}.pth. Leave it unchanged.
STATE_INDEX	The .pth file to be used in the next train/test phase. Set to None if you want to continue from the latest state dict.
SAVE_EPOCH_FREQ	Frequency to save model weights (in epochs). If None or 0, save every epoch.
MAX_EPOCHS	Number of epochs to train.
BATCH_SIZE	A dict specifying batch_size for each mode.
PARAM_GROUPS	See below.
GPUS	A list of gpu ids. If more than one gpus are specified, the model will be running on multiple gpus. Otherwise, only use DEFAULT_GPU.
DEFAULT_GPU	The default gpu to hold data. When running on multiple gpus, this should be one of the gpu specified in GPUS.
NUM_WORKERS	Number of workers to use in dataloader.

Note: Leave it unchanged means that you don't need to modify it since the default setting is enough. You're able to change it but you should take care of where the files are placed.

The PARAM_GROUPS specifies how you wanna train you model. It's a list of dictionaries representing different param groups. Each dictionary should has at least two keys: params and lr. The value of params is a list of strings, each representing a submodule of the model. The value of lr is the learning rate to be applied to these submodules. If the value of params only contains ['default'], it means all the remaining submodules except for those specified in other param groups. Other keys such as weight_decay are available, which depend on the optimizer's settings.

Here's an example.

PARAM_GROUPS = [{
    'params': ['fc'],
    'lr': 1e-3
  }, {
    'params': ['default'],
    'lr': 1e-4
  }]

In this example, the fc module is trained with a learning rate of 1e-3, while the rest are trained with 1e-4. Any module that is not mentioned in the PARAM_GROUPS still requires gradient computation, but won't be optimized. Any module whose learning rate is set to 0 neither requires gradient computation nor will be optimized.

Prepare Your Data

By convention you should put your data in the DATA_DIR directory, so that the program can find them. Typical data includes:

• pretrained model When loading model parameters, the trainer proceeds in the following order:
  1. Initialze the model according its initialize method.
  2. Look for saved state files ({STATE_PREFIX}_{epoch_id}.pth). If found, load it. Otherwise go to 3.
  3. Look for the pretrained model file according to PRETRAIN_PATH. If the file exists, load it. Otherwise do nothing.

Run the Program

python main.py -a {app_name} [-c {config_number} [-m {mode}]]

Start to train the model of the app {app_name} with the {config_number} configuration under mode {mode}. {mode} can be one of {all, train, test}. all is the default option and means alternatively train and test epochs until MAX_EPOCHS. train means only train 1 epoch while test means only test 1 epoch.

Continue and Clean

python scripts/clean -a {app_name} [-c {config_number}]

Clean the data generated by the {config_number} configuration of app {app_name}.

python script/cont {epoch_id} -a {app_name} [-c {config_number}]

For the {config_number} configuration, save the data until the {epoch_id} epoch, and delete those after it. Note that, for state files, only the {epoch_id} one would be reserved.