Operator Action Prediction

From what errors happened and in which sites predict operator actions:

• which actions to choose
• requested memory
• job splitting
• enabling/disabling xrootd

For more info about this problem look at slides

For information about recreating results look at testing directory

All scripts for recreating results are in test_models.ipynb

Examples

Training

For each target (action,memory,splitting,xrootd) you will have to train a model separately

from SimpleModel import SimpleModel as Model
from CNNModel import CNNModel as Model
from XGBoostModel import XGBoostModel as Model

m = Model(num_classes=3, num_error=54, num_sites=142)
m.train(X, Y, verbose=1)

• EmbeddingModel has pretraining procedure, so you initialize it a little bit differently:
  m = EmbeddingModel(X=X, num_classes=3, num_error=54, num_sites=142)

Here:

• X - Numpy array X.shape = (num_examples, num_error, num_sites)
• Y - Numpy array Y.shape = (num_examples,), where each number represents class index (like in sklearn)

Predicting

predictions = m.predict(Y)

Here:

• predictions - Numpy array predictions.shape = (num_examples,) with argmax over predictions, where each number represents class index (like in sklearn)

Hyperparameter Search (using skopt library)

best_params = m.find_optimal_parameters(X, Y, num_calls=12)

After finding optimal parameter, it automatically sets as default parameters.
So, by running m.train(X,Y) you will train on found parameters.

If you want to set parameters manually and train with them:

# set found parameters
m.model_params = best_params

# or set custom parameters
custom_params_for_SimpleModel = {
  'dense_layers':3,
  'dense_units':50,
  'regulizer_value':0.0015,
  'dropout_value':0.015,
  'learning_rate':1e-3
}
m.model_params = custom_params_for_SimpleModel

m.train(X,Y)

Using Model Attention

• Doesn't work with XGBoost Model

m = Model(num_classes=3, num_error=54, num_sites=142, use_attention=True)
predictions, site_attention, error_attention = m.predict(Y)

Here:

• predictions - Numpy array predictions.shape = (num_examples,) with argmax over predictions, where each number represents class index (like in sklearn)
• sites_attention - Numpy array sites_attention.shape = (num_sites,), where each number is [0,1]
• error_attention - Numpy array error_attention.shape = (num_error,), where each number is [0,1]
• Each [0,1] number represents how much model was focused on that site/error. Outputs are normalized so that all numbers in sites_attention sum to 1 (same for error_attention).

Using Resampling (using imblearn library)

from imblearn.over_sampling import SMOTE

m = Model(num_classes=3, num_error=54, num_sites=142)
m.train( X, Y, verbose=1, use_imblearn=True,
         imblearn_class=SMOTE(random_state=42, ratio=1.0))

Saving and Loading Model

path_to_directory = '~/my_model/'

m.save_model(dirpath=path_to_directory)
m.load_model(dirpath=path_to_directory)