Training Your Own Model
To train your own model invoke the train_model.py as follows:
python train_model.py --pre-processed-dir <pre-processed-dir> --model-checkpoint-path <checkpoint-dir>
Complete usage details of this script are as follows:
usage: train_model.py [-h] --pre-processed-dir PRE_PROCESSED_DIR
[--transfer-learning-model {CHAP_ALL_ADULTS, NONE}]
[--learning-rate LEARNING_RATE]
[--num-epochs NUM_EPOCHS] [--batch-size BATCH_SIZE]
[--amp-factor AMP_FACTOR]
[--cnn-window-size CNN_WINDOW_SIZE]
[--bi-lstm-window-size BI_LSTM_WINDOW_SIZE]
[--shuffle-buffer-size SHUFFLE_BUFFER_SIZE]
[--training-data-fraction TRAINING_DATA_FRACTION]
[--validation-data-fraction VALIDATION_DATA_FRACTION]
[--testing-data-fraction TESTING_DATA_FRACTION]
[--model-checkpoint-path MODEL_CHECKPOINT_PATH]
[--num-classes NUM_CLASSES]
[--class-weights CLASS_WEIGHTS]
[--down-sample-frequency DOWN_SAMPLE_FREQUENCY]
[--silent]
Argument parser for training CNN model.
required arguments:
--pre-processed-dir PRE_PROCESSED_DIR
Pre-processed data directory
optional arguments:
-h, --help show this help message and exit
--transfer-learning-model {CHAP_ALL_ADULTS, NONE}
Transfer learning model name (default:
CHAP_ALL_ADULTS)
--learning-rate LEARNING_RATE
Learning rate for training the model (default: 0.0001)
--num-epochs NUM_EPOCHS
Number of epochs to train the model (default: 15)
--batch-size BATCH_SIZE
Training batch size (default: 16)
--amp-factor AMP_FACTOR
Factor to increase the number of neurons in the CNN
layers (default: 2)
--cnn-window-size CNN_WINDOW_SIZE
CNN window size in seconds on which the predictions to
be made (default: 10)
--bi-lstm-window-size BI_LSTM_WINDOW_SIZE
BiLSTM window size in minutes on which the predictions
to be smoothed (default: 7)
--shuffle-buffer-size SHUFFLE_BUFFER_SIZE
Training data shuffle buffer size in terms of number
of records (default: 10000)
--training-data-fraction TRAINING_DATA_FRACTION
Percentage of subjects to be used for training
(default: 60)
--validation-data-fraction VALIDATION_DATA_FRACTION
Percentage of subjects to be used for validation
(default: 20)
--testing-data-fraction TESTING_DATA_FRACTION
Percentage of subjects to be used for testing
(default: 20)
--model-checkpoint-path MODEL_CHECKPOINT_PATH
Path where the trained model will be saved (default:
./model-checkpoint)
--num-classes NUM_CLASSES
Number of classes in the training dataset (default: 2)
--class-weights CLASS_WEIGHTS
Class weights for loss aggregation (default: [1.0,
1.0])
--down-sample-frequency DOWN_SAMPLE_FREQUENCY
Downsample frequency in Hz for GT3X data (default: 10)
--silent Whether to hide info messages
Model Selection: Notice that this script relies on several hyperparameters required for training the model such as learning rate, batch size, and BiLSTM window size etc. The script comes with set of default values for these parameters. However, you may need to tweak these parameters for your dataset to get the best performance.
Transfer Learning: Instead of start training a model from scratch, you can also start with an existing model and tune it for your dataset. This is also called transfer learning and can be used to train a high-quality model with limited amount of training data. Currently we support transfer learning using the CHAP_ALL_ADULTS model. In order to use transfer learning pass the --transfer-learning-model CHAP_ALL_ADULTS option when invoking the train_model.py script. Note that with transfer learning only the training hyperparameters (e.g., batch size, learning rate, num epochs etc.) can be tweaked. The architectural hyperparameters (e.g., BiLSTM window size, CNN window size) will be set to the values of the source model. When transfer learning, it is recommeded to use a low learning rate (e.g., 0.00001) to avoid overfitting.