config.py

from training_settings import configs

class Config():
    """Hyper Parameter configurations"""

    # DRAGAN experiments
    e_lr = 1e-3
    d_lr = 1e-4
    g_lr = 1e-4
    decay_steps=5000
    decay_rate =1.0

    #######################
    # Architecture params #
    #######################
    batch_size = 32
    scale = 10.
    g_bn  = True
    g_ksize = 5
    g_dim   = 64
    e_ksize = 5
    e_dim   = 128
    d_ksize = 5
    d_dim   = 256
    z_dim   = 512
    noise   = 'default' # 'multi_view' or 'default'
    use_normal_noise = True
    use_tanh = False
    use_auto_d_update=False
    encoder_use_mlp=True
    encoder_use_bn =True
    encoder_weight_decay = 1e-3
    encoder_hidden_layers = [1024, 1024]
    encoder_use_avg_pooling = False
    d_use_bn = False
    d_use_ln = True
    gan_loss_type='DCGAN' # LSGAN
    d_input_norm = False


    ########################
    # Training/Loss params #
    ########################
    content_loss_weight=10.
    clf_loss_weight=0.
    adv_loss_weight=0.
    pose_inv_loss_weight=1.
    enc_content_loss_weight = 0.
    enc_clf_loss_weight = 0.
    gp = 'dragan'  # None, wgan, dragan
    img_loss_types = ['l1', 'l2']
    vox_inv_loss_types = {
        'l1' : 1.,
        'l2' : 1.
    }
    vox_inv_loss_weight = 1.
    optimizer_type = 'adam'

    max_d_acc=1. # 0.75 for PrGAN | 1. for default

    g_iters = (lambda _,t : 1) # G's update iterations in GAN pass
    d_iters = (lambda _,t : 1) # D's update iterations in GAN pass
    num_autoencoder_iters = 1
    num_gan_iters = 1
    update_g = True  # whether update G during the training
    update_d = True  # whether update D during the training
    update_e = True  # whether update E during the training

    ####################
    # Projector params #
    ####################
    projection_typ = 'perspective' # perspective | orthographic
    resolution = 32
    vox_size   = 32
    distance   = 100
    projection_temperature = 1.
    pooling_typ='exp' # exp | max
    binarize_data=True
    # "X", "Y", "Z" for discrete case, and "XYZ" for cont. case,
    # "PTN" for X,Z cont, Y diescrete
    # rotation_axis = 'PTN'
    rotation_axis = 'XYZ'
    viewpoints = 5
    rx_range = [-20., 40.]
    ry_range = [0.,   360.]
    rz_range = [180., 180.]

    #####################
    # Validation params #
    #####################
    max_validation_batches = None
    validation_interval = 200

    ##################
    # Dataset params #
    ##################
    category_settings = configs