# conf.py will parse the yaml and extract parameters based on what is specified

# note, the YAML parser will NOT evaluate expressions in the value fields.

# e.g. "validation_frac: 1.0/3.0" will result in str value "1.0/3.0"

# will read and write (normalization, etc.) shot data

# in fs_path / [username] / signal_data | shot_lists | processed shots, etc.

# (username is automatically added as first subdir if user_subdir==True)

# will output csvlog, trained model checkpoints, etc.

# in fs_path_output / [username] / results | csv_logs | model_checkpoints | Graph, etc.

fs_path: '/Users/'

user_subdir: True

fs_path_output: '/Users/'

user_subdir_output: True

target: 'hinge' # 'maxhinge' # 'maxhinge' # 'binary' # 'hinge'

num_gpus: 1 # per node

paths:

signal_prepath: '/signal_data/' # /signal_data/jet/

shot_list_dir: '/shot_lists/'

tensorboard_save_path: '/Graph/'

data: d3d_0D

# if specific_signals: [] left empty, it will use all valid signals defined on a machine. Only use if need a custom set

specific_signals: [] # ['q95','li','ip','betan','energy','lm','pradcore','pradedge','pradtot','pin','torquein','tmamp1','tmamp2','tmfreq1','tmfreq2','pechin','energydt','ipdirect','etemp_profile','edens_profile']

executable: "mpi_learn.py"

shallow_executable: "learn.py"

data:

bleed_in: 0 # how many shots from the test set to use in training?

bleed_in_repeat_fac: 1 # how many times to repeat shots in training and validation?

bleed_in_remove_from_test: True

bleed_in_equalize_sets: False

signal_to_augment: None # 'plasma current'

augmentation_mode: None

augment_during_training: False

cut_shot_ends: True

recompute: False

recompute_normalization: False

# specifies which of the signals in the signals_dirs order contains the plasma current info

current_index: 0

plotting: False

# maximum number of shots to use

use_shots: 200000 # 1000

positive_example_penalty: 1.0 # by what factor to upweight positive examples?

# normalization timescale

dt: 0.001

T_min_warn: 30

# maximum TTD considered

T_max: 1000.0

# warning time in seconds

# The shortest works best so far: less overfitting. log(TTD) prediction also works well. 0.5s better than 0.2s

T_warning: 1.024 # 0.512 # 0.25 # 1.0

current_thresh: 750000

current_end_thresh: 10000

# the characteristic decay length of the decaying moving average window

window_decay: 2

# the width of the actual window

window_size: 10

# TODO(KGF): optimize the normalizer parameters

normalizer: 'var'

norm_stat_range: 100.0

equalize_classes: False

# the fraction of samples with which to train the shallow model

# shallow_sample_prob: 0.01

floatx: 'float32'

model:

loss_scale_factor: 1.0

use_batch_norm: false

torch: False

shallow: False

shallow_model:

type: "xgboost" # "random_forest"

# the number of samples to use for training

num_samples: 1000000 # 1000000

n_estimators: 100 # used in random forest

max_depth: 3 # used in random forest and xgboost (def = 3)

C: 1.0 # used in svm

kernel: "rbf" # rbf, sigmoid, linear, poly, for svm

learning_rate: 0.1 # used in xgboost

scale_pos_weight: 10.0 # used in xgboost

# final layer has this many neurons, every layer before has twice as many

final_hidden_layer_size: 10

num_hidden_layers: 3

learning_rate_mlp: 0.0001

mlp_regularization: 0.0001

# should a finished model be loaded if available?

skip_train: False

# length of LSTM memory

pred_length: 200

pred_batch_size: 128

# TODO(KGF): optimize length of LSTM memory

length: 128

skip: 1

# hidden layer size

# TODO(KGF): optimize size of RNN layers

# size 100 slight overfitting, size 20 no overfitting. 200 is not better than 100.

# Prediction is much better with size 100, size 20 cannot capture the data.

rnn_size: 200

rnn_type: 'LSTM'

# TODO(KGF): optimize number of RNN layers

rnn_layers: 2

num_conv_filters: 128

size_conv_filters: 3

num_conv_layers: 3

pool_size: 2

dense_size: 128

extra_dense_input: False

# have not found a difference yet

optimizer: 'adam'

clipnorm: 10.0

regularization: 0.001

dense_regularization: 0.001

# lr=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

# lr=1e-4 also works well if we decay a lot (i.e ~0.7 or more)

lr: 0.00002 # 0.00001 # 0.0005 # for adam plots 0.0000001

lr_decay: 0.97 # 0.98 # 0.9

stateful: True

return_sequences: True

dropout_prob: 0.1

# only relevant if we want to do MPI training. The number of steps with a single replica

warmup_steps: 0

# how many initial timesteps to ignore during evaluation (to let the internal state settle)

ignore_timesteps: 100

backend: 'tensorflow'

training:

as_array_of_shots: True

shuffle_training: True

# used iff 1) test & 2) (train U validate) are both sampled from the same distribution/source lists of shots:

train_frac: 0.75

validation_frac: 0.3333333333333333

batch_size: 128 # 256

# THE MAX_PATCH_LENGTH WAS THE CULPRIT FOR NO TRAINING! Lower than 1000 performs very poorly

max_patch_length: 100000

# How many shots are we loading at once?

num_shots_at_once: 200

# large number = maximum number of epochs.

# Early stopping will occur if loss does not decrease, after some patience # of epochs

num_epochs: 1000

use_mock_data: False

data_parallel: False

hyperparam_tuning: False

batch_generator_warmup_steps: 0

use_process_generator: False

num_batches_minimum: 20 # minimum number of batches per epoch

ranking_difficulty_fac: 1.0 # how much to upweight incorrectly classified shots during training

timeline_prof: False

step_limit: 50

no_validation: True

callbacks:

list: ['earlystop']

metrics: ['val_loss','val_roc','train_loss']

mode: 'max'

monitor: 'val_roc'

patience: 5

write_grads: False

monitor_test: True

monitor_times: [30,70,200,500,1000]

env:

name: 'frnn'

type: 'anaconda'