v2.toml

##################################################################################################
# Defining the model
##################################################################################################


# hidden layers in order  or explicitly number them:  which?
# [layer]
#     type            = "conv"
#     image           = { x = 24, y = 24 }    # don't think we need this since prior layer conv settings define it
#     filter          = "something"
#     stride          = "something"
#     pad             = "something"
#     activation      = "relu"

# [layer]
#     activation      = "tanh"   # without specifying type, a fully connected linear layer is assumed by default
#     units           = 256
#     dropout         = false
#     droplim         = 0.8           # or for multiple layers [1.0, 0.8, 1.0], if for all except output

# [layer]
#     activation      = "relu"
#     units           = 128

# [layer]
#     activation      = "tanh"
#     units           = 64

# [layer]                           # do we need to make this explicit or leave it implicit???
#     type            = "linear"    # or "affine"  USE THIS TO FORCE FULLY CONNECTED OR DENSE AFTER CONV,RNN

[layer.1]
    type            = "conv"
    image           = { x = 24, y = 24, z = 3 }    # don't think we need this since prior layer conv settings define it
    filter          = "something"
    stride          = "something"
    pad             = "something"
    activation      = "relu"
    trainable       = true

[layer.2]
    activation      = "tanh"   # without specifying type, a fully connected linear layer is assumed by default
    units           = 256
    dropout         = false
    droplim         = 0.8           # or for multiple layers [1.0, 0.8, 1.0], if for all except output
    trainable       = true          # this is the default.  OK to leave this out.  use primarily for transfer
                                    # learning with trainable = false
[layer.3]
    activation      = "relu"
    units           = 128

[layer.4]
    activation      = "tanh"
    units           = 64

[layer.5]                           # do we need to make this explicit or leave it implicit???
    type            = "linear"

[output] # should this be an element in the layer array?--we can always stick it there
# required
    classify        = "softmax"
    target          = "something" # we need a variable


##################################################################################################
# Defining the data
##################################################################################################

[data.train] # should this be an element in the layer array?--we can always stick it there  NO: use layer to mean hidden layer
# required
    norm_mode       = "none"   # maybe put this in training?
    input           = "something" # we need a variable
    image           = { x = 28, y = 28, z = 3 }  # need to know if it is stacked or flat
    target          = "something"
    # should we put the input variables in here?

[data.test]
    input           = "something"
    target          = "something"

##################################################################################################
# Controlling the training process
##################################################################################################

[training]                  # required
    epochs              = 24  # required
    alpha               = 0.74  # learning rate or nu in some notation; required
    learn_decay         = [0.5, 4.0] # or {factor = 0.5, steps = 4.0}
    quiet               = true
    initializer         = "xavier"
    scale_init          = 2.0
    bias_initializer    = 0.3
    input_norm          = "none"   # put in training or input layer of model


[minibatch]                             # global--applies to all layers; if excluded no minibatches
    do_batch        = true
    do_batch_norm   = true  # this can be by layer, but minibatch applies to the entire training run
    mb_size_in      = 50
    norm_after      = "linear"    # or "activation"--not implemented


[optimization]                  # optional: if absent, no optimzation
    opt             = "adam"
    opt_params      = [0.9, 0.999]

[regularization]                # global--applies to all hidden layers--this could be done per layer--not implemented
# optional--if absent, no regularization
    reg             = "Maxnorm"  # or "L2", "L1"
    maxnorm_lim     = 4.0
    lambda          = 0.000191   # only applies to L2 or L1--ignored otherwise

[plotting]          # optional: if absent, no plotting or plot data
    stats           = ["Train", "Learning", "Test", "Cost", "epoch"]    # or "batch"
    plot_now        = true

[results]
    # all = true  # this is the default--and what you get if you don't include the results table
    train_inputs = true
    train_targets = true
    train_preds = true
    test_inputs = true
    test_targets = true
    test_preds = true
    hyper_params = true
    wgts = true
    batchnorm_params = true