forked from amazon-archives/amazon-dsstne
-
Notifications
You must be signed in to change notification settings - Fork 1
/
LDL.txt
152 lines (147 loc) · 11.5 KB
/
LDL.txt
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
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
Neural network Layer Definition Language (Rev 0.8):
{
"Version" : <String> # Indicates version (by default uses latest)
"Name" : <String> # Optional name for network
"WeightData" : <String> # Optional NetCDF dataset file containing weights
"Kind" : <String> # Either AutoEncoder or FeedForward (default)
"ShuffleIndices" : <Boolean> # Shuffle data ordering during training (Default true)
"LocalResponseNormalization" :
{
"k" : <float>, # Local Response Normalization offset (default 2)
"n" : <Integer>, # Local Response Normalization spread (default 5)
"alpha" : <float>, # Local Response Normalization scaling (default 0.0001)
"beta" : <float> # Local Response Normalization exponent (default 0.75)
},
"SparsenessPenalty" :
{
"p" : <float>, # Target sparseness probability for sparse hidden units (default 0.0)
"beta" : <float> # Sparse penalty weight on sparse hidden units (default 0.0)
},
"Denoising" :
{
"p" : <float>, # Probability of denoising inputs (for sparse input layers, only denoise on > 0)
},
"DeltaBoost" : # Scales gradient when using a binary crossentropy loss function
{
"one" : <float> # Scaling factor for non-zero target values (Default 1.0)
"zero" : <float> # Scaling factor for zero target values (Default 1.0)
},
"ScaledMarginalCrossEntropy" : { # Parameters for binary scaled marginal cross entropy loss function
"oneTarget" : <float> # Relaxed target for non-zero target values (Default 0.9)
"zeroTarget" : <float> # Relaxed target for zero target values (Default 0.1)
"oneScale" : <float> # Scaling factor for non-zero target values (Default 1.0)
"zeroScale" : <float> # Scaling factor for zero target values (Default 1.0)
},
"Layers" : <Array> # List of input, hidden, and output units. There must
# be at least one input and one output layer for this
# for this to be a valid neural network
[
# Input layer definition
{
"Kind" : "Input" # Indicates this layer is an input layer
"Name" : <String>, # Provides ASCII identifier for complex topologies (optional)
"DataSet" : <String>, # Name of dataset that provides input values
"N" : <Integer> or <Array> or "auto", # Input unit dimensions, use "auto" to base them on DataSet dimensions
"Sparse" : <Boolean>, # Indicates whether layer is sparse (default false)
"pDropout" : <float> # Droput probability (default 0.0)
},
.
.
.
# Hidden layer definition
{
"Kind" : "Hidden" # Indicates this is a hidden layer
"Name" : <String> # Provides an ASCII identifier for a layer (optional)
"Source" : <String> or <Array> # Specifies source layer for hidden/output or source data set for input layers
"Type" : <String>, # FullyConnected or Convolution layer (default FullyConnected)
"N" : <Integer>, # Number of output units (Mandatory)
"WeightInit" : <Object> # Incoming weight initialization scheme, default Xavier
{
"Scheme" : <String> # Xavier, CaffeXavier, Gaussian, Uniform, Constant, UnitBall (default Xavier)
"Scale" : <float> # Scaling factor for all initializations (default 1)
"Bias" : <float> # Bias initialization value (default 0)
}
"Activation" : <String>, # Sigmoid, Relu, Tanh, Linear, PRelu, LinearMax, MaxOut (default sigmoid)
"Kernel" : <Integer> or <Array>, # kernel dimensions X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"KernelStride" : <Integer> or <Array>, # kernel stride X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"pDropout" : <float> # Droput probability (default 0.0)
"WeightNorm": <float> # Indicates maximum allowable length of individual incoming weight vectors (default 0, unconstrained)
"DeltaNorm" : <float> # Indicates maximum allowable length of delta vectors for backpropagation (default 0, unconstrained)
"BatchNormalization" : <Boolean> # Indicates whether output should be rescaled to N(0,1) (default false)
"Sparse" : <Boolean>, # Indicates whether layer is sparse (default false)
"SparsenessPenalty" : <Boolean> # Indicates whether sparseness penalty should be applied (default false)
"L1Penalty" : <float> # L1 penalty factor to apply to activations (default 0.0, off)
"Recurrent" : <String> # Makes layer recurrent: "LSTM", "SRN", "SCRN", "FSMN", or "None" (default "None")
"SharedWeights" : <Array> or <Object> # Indicates whether to re-use weights from a previous layer (optional)
[
{ "InputLayer" : <String>, # Input to this layer from which to attach shared weights
"SourceInputLayer" : <String>, # SourceInputLayer and SourceOutputLayer collectively
"SourceOutputLayer" : <String>, # specify which weight matrix will be shared here
"Transposed" : <Boolean> # Use transpose of matrix? (default false)
},
.
.
],
},
# Pooling Layer Definition
{
"Kind" : "Pooling", # Indicates this is a pooling layer
"Name" : <String>, # Provides an ASCII identifier for the layer (optional)
"Source" : <String>, # Pooling layers process a single hidden or input layer
"Function" : <String>, # Max, Stochastic, LocalContrastNormalization, LocalResponseNormalization, GlobalTemporal
"Kernel" : <Integer> or <Array>, # kernel dimensions X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"KernelStride" : <String>, # kernel stride X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"pDropout" : <float> # Droput probability (default 0.0)
},
# Target Layer Definition
{
"Kind" : "Target", # Indicates this is a target layer
"Name" : <String>, # Provides an ASCII identifier for the layer (optional)
"Target" : <String>, # Indicates layer that serves as source data and provides dimensions
"WeightInit" : <Object> # Incoming weight initialization scheme, default Xavier
{
"Scheme" : <String> # Xavier, CaffeXavier, Gaussian, Uniform, Constant, UnitBall (default Xavier)
"Scale" : <float> # Scaling factor for all but Xavier Initializations (default 1)
"Bias" : <String> or <float> # Bias initialization either drawn from N(0,1) or a constant (default 0)
}
"ErrorFunction" : <String>, # Either SMCE, L1, L2, or CrossEntropy (default)
},
.
.
.
# Output layer definition
{
"Kind" : "Output", # Indicates this is an output layer
"Name" : <String>, # Provides an ASCII identifier for a layer (optional)
"DataSet" : <String>, # Name of dataset that provides target values (optional)
"Source" : <String> or <Array>, # Specifies input(s) to this layer by name (default hidden layer or first input)
"Type" : <String>, # FullyConnected or Convolution layer (default FullyConnected)
"N" : <Integer> or "auto", # Number of output units, use "auto" to base them on DataSet dimensions
"Activation" : <String>, # Sigmoid, Relu, Tanh, Linear, SoftMax, ReluMax, LinearMax (default sigmoid)
"Kernel" : <Integer> or <Array>, # kernel dimensions X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"KernelStride" : <String>, # kernel stride X, [X], [X, Y], or [X, Y, Z] (default 1 for each)
"WeightInit" : <Object> # Incoming weight initialization scheme, default Xavier
{
"Scheme" : <String> # Xavier, CaffeXavier, Gaussian, Uniform, Constant, UnitBall (default Xavier)
"Scale" : <float> # Scaling factor for all but Xavier Initializations (default 1)
"Bias" : <String> or <float> # Bias initialization either drawn from N(0,1) or a constant (default 0)
}
"WeightNorm" : <float>, # Indicates maximum allowable length of individual incoming weight vectors (default 0, unconstrained)
"DeltaNorm" : <float>, # Indicates maximum allowable length of delta vectors for backpropagation (default 0, unconstrained)
"pDropout" : <float> # Droput probability (default 0.0)
"Sparse" : <Boolean>, # Indicates whether layer is sparse (default false)
"SparsenessPenalty" : <Boolean> # Indicates whether sparseness penalty should be applied (default false)
"SharedWeights" : <Array> or <Object> # Indicates whether to re-use weights from a previous layer (optional)
[
{ "InputLayer" : <String>, # Input to this layer from which to attach shared weights
"SourceInputLayer" : <String>, # SourceInputLayer and SourceOutputLayer collectively
"SourceOutputLayer" : <String>, # specify which weight matrix will be shared here
"Transpose" : <Boolean> # Use transpose of matrix? (default false)
},
.
.
],
},
]
"ErrorFunction" : <String> # Either SMCE, L1, L2, or CrossEntropy (default)
}