Skip to content

Latest commit

 

History

History
66 lines (49 loc) · 3.66 KB

TRAINING.md

File metadata and controls

66 lines (49 loc) · 3.66 KB
Raw

Training

Description:
  1. Trains a model specified in config file
  2. Saves classifier and best checkpoints
  3. Saves logs for tensorboard

Step 1: Dataset

OPTION 1 - using preprocessed files(from PREPROCESSING STEP)

1. Create train config in configs like train_.yml

Example: example_config_files/train_dense_net_example.yml In the train config you can choose model, dataset, trainer, metrics, dataloader configs, features see the example

Requirement: Add these two parameters under special_inputs, more details in example config

special_inputs: input_dim: 53 n_outputs: 2

2. python run_train --config-yml configs/train_<something>.yml

a. The results will be pickle file of classifier saved in classifiers
b. If training torch model the epoch loss and other metrics specified in the model itself will be logged for tensorboard in train_results

3. Repeat steps 1 and 2 but with other examples

a. Logistic Regression config example_config_files/train_logistic_regression_example.yml

Note changes in Logistic Regression:

  1. Trainer changed to sklearn_trainer
  2. model.name, model.tag changed
  3. special inputs are accords with inputs of sklearn.linear_model.LogisticRegression class
  4. dataset.Dataloader was removed as it's not needed
b. Logistic Regression config example_config_files/train_xgboost_example.yml

4. The result classifiers are saved in classifiers folder

If trained torch model the epoch wise and additional metrics are saved in train_results in tensorboard format

OPTION 2 - if you have your own torch.Dataset

Example:

1. Prepare data

  1. You must be in G42 network to access this dataset
  2. Download dataset of x y coordinates of players in NBA games
  3. Unzip it and put in data folder
  4. Original dataset and code reference

2. Insert the appropriate torch.Dataset class in path_to_project/modules/datasets/

The torch.Dataset for this specific example is already inserted For reference see path_to_project/modules/datasets/basket_dataset.py

3. Create and insert training config in path_to_project/configs folder

Use example example_config_files/train_dagnet_example.yml

4. python run_train --config-yml configs/train_<something>.yml

Step 2: Model

1. Take a look at one of the models already created in modules/models folder. You already can use them or any of sklearn models. In addition, you can create your own by inheriting BaseModel or DefaultModel. Just override the abstract methods.

2. Don't forget to register it line other models(put "@registry.register_model('my_custom_model')" above your model class declaration). In this case my_custom_model will be the name of your model that you specify in your train and prediction config files.

3. You have to specify tag of your model in config file(see examples). This way models of the same type(for example: 2 xgboosts) will be distinguished. They will be named with model class name and tag name in underscore style.

4. Once the model is written make sure it's in modules/models folder and run

python run_train --config-yml configs/train_<something>.yml

Tips:

  1. For binary classification torch models use n_outputs: 1
  2. For all classification torch models beware you need to specify self.output_function (last layer activation) and self.prediction_function (used during the prediction step, usually nn.Softmax) yourself in your model class.