Exemplo n.º 1
0
def get_config():
    c = Config()
    # cli flags using trixi, overwrite using e.g. --learning_rate=0.001
    c.txt_file = 'assets/001ssb.txt'  # Path to a .txt file to train on
    c.seq_length = 30        # Length of an input sequence
    c.gen_length = 250       # Length of the generated sequence
    c.lstm_num_hidden = 128  # Number of hidden units in the LSTM
    c.lstm_num_layers = 2    # Number of LSTM layers in the model

    # Training params
    c.batch_size = 64        # Number of examples to process in a batch
    c.learning_rate = 2e-3   # Learning rate

    # It is not necessary to implement the following three params, but it may help training.
    c.learning_rate_decay = 0.96  # Learning rate decay fraction
    c.learning_rate_step = 5000  # Learning rate step
    c.dropout_keep_prob = 1.0  # Dropout keep probability

    c.train_steps = 1e6      # Number of training steps
    c.max_norm = 5.0

    # Misc params
    c.summary_path = './summaries/'  # Output path for summaries
    c.print_every = 5        # How often to print training progress
    c.sample_every = 100     # How often to sample from the model
    c.device = 'cuda:0'      # Training device 'cpu' or 'cuda:0'
    c.temperature = 0.5      # balances the sampling strategy between fully-greedy (near 0) and fully-random (higher). e.g. 0.5, 1.0, 2.0.

    return c
Exemplo n.º 2
0
def get_config():
    c = Config()
    # cli flags using trixi, overwrite using e.g. --learning_rate=0.001
    c.model_type = 'RNN'    # Model type, should be 'RNN' or 'LSTM'
    c.input_length = 10     # Length of an input sequence
    c.input_dim = 1         # Dimensionality of input sequence
    c.num_classes = 10      # Dimensionality of output sequence
    c.num_hidden = 128      # Number of hidden units in the model
    c.batch_size = 128      # Number of examples to process in a batch
    c.learning_rate = 0.001 # Learning rate
    c.train_steps = 10000   # Number of training steps
    c.max_norm = 10.0
    c.device = 'cuda:0'     # Training device 'cpu' or 'cuda:0'
    return c