Train the Neural Network

Usage

After creating the tensors that will be used to train and test the neural network, it is possible to start training the neural network. Different models and optimizers can be selected by command-line argument parsing. Additionally, the class_weight implementation from Keras can be activated. This was implemented to take into account the uneven distribution of samples across instrument labels. Given that this did not result in improvements with the IRMAS data set, the use of this feature is optional in argument parsing. Since the size of the neural network and the amount of data, it is strongly recommended to use a GPU to reduce training time.

Available model names are:

  • model_baseline - Baseline model by Han et al. using ReLU activation functions (refer to section 4.1).
  • model_leaky - Baseline model by Han et al. using Leaky ReLU activation functions (refer to section 4.1).
  • model_two_branch - Model that allows for two simultaneous inputs and late fusion (refer to section 5.6.2).
  • model_multi_res - Model that allows for three dimensional input (refer to section 5.6.3).

Available optimizer names are:

  • adam - Adam optimizer with parameters beta_1=0.9, beta_2=0.999, epsilon=1e-08.
  • sgd - Stochastic Gradient Descent optimizer with parameters momentum=0.9, decay=learning_rate/nb_epochs, Nesterov=True.

To train the network, use the following command:

python3 train.py -m <model_name> -o <optimizer_name> -c <y/n>

Finally, the resulting files are stored in the MODEL_PATH defined in settings.py:

  • <model_name>.<optimizer_name><iteration>.history.npy - Training history to evaluate training and validation loss and accuracy.
  • <model_name>.<optimizer_name><iteration>.best.hdf5 - Stored weights by using ModelCheckpoint saving the model with best validation accuracy.
  • <model_name>.<optimizer_name><iteration>.json - Structure of the neural network.

Documentation

class train.Trainer(model, optimizer, data_dist)

Trainer trains a convolutional neural network with a given model. The training data set must have already been extracted with generate_dataset.py. The balanced train-validation split is already performed during the data set generation, so the trainer only performs a randoms shuffling on every epoch. All of the parameters are inherited from settings.py.

Parameters:
  • model (object) – model to be trained with the selected data set.
  • optimizer (str) – optimizer to be used in backpropagation.
  • data_dist (pandas dataframe) – dataframe with samples distribution in dataset (if class_weight is implemented).
load_data()

This method loads the training and validation tensors (melspectrogram arrays) and corresponding labels to memory (for the case of single input). It also implements one hot encoding for the labels (both training and validation data sets).

load_multi_data()

This method loads the training and validation tensors (melspectrogram arrays) and corresponding labels to memory (for the case of multiple input). It also implements one hot encoding for the labels (both training and validation data sets).

train(epochs, batch, iter_num)

This method trains the corresponding model and implements two callback functions: Early Stopping (patience = 5 epochs) and Model Checkpointing. The number of epochs and mini-batch size are inherited from settings.py. Additionally, the training history, weights, and structure are stored in the MODEL_PATH.

Parameters:
  • epoch (int) – maximum number of epochs to train model.
  • batch (int) – mini-batch size.
  • iter_num (int) – number of training iterations to compensate for random initialization.