Repairing

The clean cells are used as training examples to learn the parameters (weights) of a softmax regression model. Once those weights are defined, we use this model to perform inference on the “don’t-know” cells and insert the most likely value for each cell.

Softmax

class holoclean.learning.softmax.SoftMax(session, X_training)

build_model(featurizers, input_dim_non_dc, input_dim_dc, output_dim, tie_init=True, tie_DC=True)

Initializes the logreg part of our model

Parameters:

• input_dim_non_dc – number of init + cooccur features
• featurizers – list of featurizers
• input_dim_dc – number of dc features
• output_dim – number of classes
• tie_init – boolean to decide weight tying for init features
• tie_DC – boolean to decide weight tying for dc features

Returns:

newly created LogReg model

log_weights()

Writes weights in the logger

Returns:Null

logreg(featurizers)

Trains our model on clean cells and predicts vals for clean cells

Returns:predictions

predict(model, x_val, mask=None)

Runs our model on the test set

Parameters:

• model – trained logreg model
• x_val – test x tensor
• mask – masking tensor to restrict domain

Returns:

predicted classes with probabilities

save_prediction(Y)

Stores our predicted values in the database

Parameters:Y – tensor with probability for each class Returns:Null

setupMask(clean=1, N=1, L=1)

Initializes a masking tensor for ignoring impossible classes

Parameters:

• clean – 1 if clean cells, 0 if don’t-know
• N – number of examples
• L – number of classes

Returns:

masking tensor

setuptrainingX(sparse=0)

Initializes an X tensor of features for training

Parameters:sparse – 0 if dense tensor, 1 if sparse Returns:x tensor of features

train(model, loss, optimizer, x_val, y_val, mask=None)

Trains our model on the clean cells

Parameters:

• model – logistic regression model
• loss – loss function used for evaluating performance
• optimizer – optimizer for our neural net
• x_val – x tensor - features
• y_val – y tensor - output for comparison
• mask – masking tensor

Returns:

cost of traininng