You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
specific_signals: [] #['q95','li','ip','betan','energy','lm','pradcore','pradedge','pradtot','pin','torquein','tmamp1','tmamp2','tmfreq1','tmfreq2','pechin','energydt','ipdirect','etemp_profile','edens_profile'] #if left empty will use all valid signals defined on a machine. Only use if need a custom set
max_depth: 3#for random forest and xgboost (def = 3)
66
-
C: 1.0#for svm
67
-
kernel: "rbf"#rbf, sigmoid, linear, poly, for svm
68
-
learning_rate: 0.1#xgboost
69
-
scale_pos_weight: 10.0#xgboost
70
-
final_hidden_layer_size: 10#final layers has this many neurons, every layer before twice as many
63
+
num_samples: 1000000#1000000 #the number of samples to use for training
64
+
type: "xgboost"#"xgboost" #"random_forest"
65
+
n_estimators: 100#for random forest
66
+
max_depth: 3#for random forest and xgboost (def = 3)
67
+
C: 1.0#for svm
68
+
kernel: "rbf"#rbf, sigmoid, linear, poly, for svm
69
+
learning_rate: 0.1# used in xgboost
70
+
scale_pos_weight: 10.0# used in xgboost
71
+
final_hidden_layer_size: 10#final layers has this many neurons, every layer before twice as many
71
72
num_hidden_layers: 3
72
73
learning_rate_mlp: 0.0001
73
74
mlp_regularization: 0.0001
74
-
skip_train: False #should a finished model be loaded if available
75
-
#length of LSTM memory
75
+
skip_train: False #should a finished model be loaded if available
76
+
#length of LSTM memory
76
77
pred_length: 200
77
78
pred_batch_size: 128
78
-
#TODO optimize
79
+
#TODO(KGF): optimize length of LSTM memory
79
80
length: 128
80
81
skip: 1
81
-
#hidden layer size
82
-
#TODO optimize
82
+
#hidden layer size
83
+
#TODO(KGF): optimize size of RNN layers
83
84
rnn_size: 200
84
-
#size 100 slight overfitting, size 20 no overfitting. 200 is not better than 100. Prediction much better with size 100, size 20 cannot capture the data.
85
+
#size 100 slight overfitting, size 20 no overfitting. 200 is not better than 100. Prediction much better with size 100, size 20 cannot capture the data.
85
86
rnn_type: 'LSTM'
86
-
#TODO optimize
87
+
#TODO(KGF): optimize number of RNN layers
87
88
rnn_layers: 2
88
89
num_conv_filters: 128
89
90
size_conv_filters: 3
90
91
num_conv_layers: 3
91
92
pool_size: 2
92
93
dense_size: 128
93
94
extra_dense_input: False
94
-
#have not found a difference yet
95
+
#have not found a difference yet
95
96
optimizer: 'adam'
96
97
clipnorm: 10.0
97
98
regularization: 0.001
98
99
dense_regularization: 0.001
99
-
#1e-4 is too high, 5e-7 is too low. 5e-5 seems best at 256 batch size, full dataset and ~10 epochs, and lr decay of 0.90. 1e-4 also works well if we decay a lot (i.e ~0.7 or more)
100
-
lr: 0.00002#0.00001 #0.0005 #for adam plots 0.0000001 #0.00005 #0.00005 #0.00005
101
-
lr_decay: 0.97#0.98 #0.9
100
+
# lr=1e-4 is too high, 5e-7 is too low. 5e-5 seems best at 256 batch size, full dataset
101
+
# and ~10 epochs, and lr decay of 0.90
102
+
# lr=1e-4 also works well if we decay a lot (i.e ~0.7 or more)
103
+
lr: 0.00002# 0.00001 # 0.0005 # for adam plots 0.0000001 # 0.00005 # 0.00005 # 0.00005
104
+
lr_decay: 0.97# 0.98 # 0.9
102
105
stateful: True
103
106
return_sequences: True
104
107
dropout_prob: 0.1
105
-
#only relevant if we want to do mpi training. The number of steps with a single replica
108
+
#only relevant if we want to do mpi training. The number of steps with a single replica
106
109
warmup_steps: 0
107
-
ignore_timesteps: 100#how many initial timesteps to ignore during evaluation (to let the internal state settle)
110
+
ignore_timesteps: 100#how many initial timesteps to ignore during evaluation (to let the internal state settle)
108
111
backend: 'tensorflow'
109
112
training:
110
113
as_array_of_shots: True
111
114
shuffle_training: True
112
115
train_frac: 0.75
113
116
validation_frac: 0.33
114
-
batch_size: 128#256
115
-
#THIS WAS THE CULPRIT FOR NO TRAINING! Lower than 1000 performs very poorly
117
+
batch_size: 128#256
118
+
# THE MAX_PATCH_LENGTH WAS THE CULPRIT FOR NO TRAINING! Lower than 1000 performs very poorly
116
119
max_patch_length: 100000
117
-
#How many shots are we loading at once?
120
+
#How many shots are we loading at once?
118
121
num_shots_at_once: 200
119
-
num_epochs: 1000
122
+
num_epochs: 1000# large number = maximum number of epochs. Early stopping will occur if loss does not decrease
120
123
use_mock_data: False
121
124
data_parallel: False
122
125
hyperparam_tuning: False
123
126
batch_generator_warmup_steps: 0
124
127
use_process_generator: False
125
-
num_batches_minimum: 20#minimum number of batches per epoch
126
-
ranking_difficulty_fac: 1.0#how much to upweight incorrectly classified shots during training
128
+
num_batches_minimum: 20#minimum number of batches per epoch
129
+
ranking_difficulty_fac: 1.0#how much to upweight incorrectly classified shots during training
0 commit comments