Human Interpretable AI: Randomized StochasticTsetlin Machine with Drop Clause

Install the drop clause multi-gpu Tsetlin machine:

python setup.py install

pip install -r requirements.txt

Using the drop clause TM:

from PyTsetlinMachineCUDA.tm import MultiClassConvolutionalTsetlinMachine2D

tm = MultiClassConvolutionalTsetlinMachine2D(clauses, T, s, (patch_size, patch_size), clause_drop_p = drop_clause, number_of_gpus=n_gpus, number_of_state_bits=number_of_state_bits)
tm.fit(train_data, train_labels, epochs=1, incremental=True)
accuracy = 100*(tm.predict(test_data) == test_labels).mean()

Running the code (from examples folder):

With default parameters- python CIFARDemo2DConvTM_Interpret.py -gpus 1 -stop_train 1000

With Interpretability- python CIFARDemo2DConvTM_Interpret.py -interpret True

With changes in hyperparameters- python CIFARDemo2DConvTM_Interpret.py -n_clauses_per_class 30000 -s 10.0 -T 750 -drop_clause 0.5 -patch_size 8 -gpus 8

Similarly for NLP Interpretability- python SSTDemoWeightedClauses_Interpret.py