network_final_submission.inc 1.18 KB
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
# Definition of network architecture and other model parameters
# The network has a total of 373169 weights.

# number of models for bagging
model_count=5

# length of spectrogram slice (must correspond to architecture below)
net_width=1000

# network layers
cnvmult=16
nl=leaky_rectify
# augmentation
featpart="augment:Rotation|augment:Shift(low=-1,high=1,axis=3)"
# normalization
normpart="normalize:BatchNorm(axes=0x1x2,alpha=0.1,beta=None,gamma=None)" # normalize over all except frequency axis
# convolutional layers
convpart="Conv2D(${cnvmult}x3x3)|${nl}|Pool2D(3x3)|Conv2D(${cnvmult}x3x3)|${nl}|Pool2D(3x3)|Conv2D(${cnvmult}x3x1)|${nl}|Pool2D(3x1)|Conv2D(${cnvmult}x3x1)|${nl}|Pool2D(3x1)"
# dense layers
densepart="Dropout(p=0.5)|Dense(256)|${nl}|Dropout(p=0.5)|Dense(32)|${nl}|Dropout(p=0.5)|Dense(1)|output:sigmoid"
# all together
net_layers="${featpart}|${normpart}|${convpart}|${densepart}"

# training options
net_options="--var updates=adam --var batchsize=64 --var epochsize=1500 --var learning_rate_mode=adaptive --var learning_rate=0.001 --var learning_rate_stop=0.00001"


# number of folds for pseudo-labeling
pseudo_folds=2
# confidence threshold for pseudo-labeling
pseudo_threshold=0.1