Link to Google Colab SPN Implementation: Google Colab Link
-
Click on the Google Colab Link to open the notebook.
-
Once the notebook is open, go through each cell one by one and execute them by clicking on the play button next to each cell.
-
Make sure to execute the cells in sequential order, as some cells may depend on variables or functions defined in previous cells.
-
As you execute each cell, observe the output and any instructions provided in comments or Markdown cells within the notebook.
-
If any packages need to be installed, Colab will prompt you to install them when you run the corresponding cell. You only need libspn-keras and TensorFlow.
-
Once all cells have been executed, you should have the SPN model set up and ready for training and evaluation.
-
You can then proceed with training the model on the provided dataset and evaluating its performance as described in the notebook.
Dataset used: mnist_*.csv is a small sample of the MNIST database, which is described at: MNIST Dataset Description
-
!pip install libspn-keras tensorflow-datasets: This line installs the required Python packages, namely libspn-keras and tensorflow-datasets, using pip. -
import libspn_keras as spnk: This imports the libspn-keras library and aliases it as spnk for easier reference. -
print(spnk.get_default_sum_op()): This line prints the default sum operation used in the library. -
spnk.set_default_sum_op(spnk.SumOpGradBackprop()): It sets the default sum operation to use gradient backpropagation for backpropagating gradients during training. -
from tensorflow import keras: This imports the Keras module from TensorFlow. -
spnk.set_default_accumulator_initializer(...): This sets the default initializer for accumulator nodes in the SPN. Accumulator nodes are used for summing up values in the SPN. -
import tensorflow_datasets as tfds: This imports the TensorFlow Datasets module, which provides convenient access to various datasets. -
batch_size = 32: Sets the batch size for training. -
normalize = spnk.layers.NormalizeStandardScore(...): Defines a normalization layer to normalize input data to have zero mean and unit variance. -
def take_first(a, b): ...: Defines a function take_first to be used for preprocessing data. -
`normalize.adapt(...): Adapts the normalization layer to the training data.
-
sum_product_network = keras.Sequential([...]): Defines the architecture of the SPN model using a Sequential API from Keras. -
sum_product_network.summary(): Prints a summary of the model architecture. -
mnist_train = tfds.load(name="mnist", split="train", as_supervised=True)...: Loads the MNIST training dataset. -
mnist_test = tfds.load(name="mnist", split="test", as_supervised=True)...: Loads the MNIST test dataset. -
optimizer = keras.optimizers.Adam(learning_rate=1e-2): Defines the Adam optimizer with a specified learning rate. -
metrics = [keras.metrics.SparseCategoricalAccuracy()]: Defines the evaluation metric for the model. -
loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True): Defines the loss function for the model. -
sum_product_network.compile(...): Compiles the model with the specified optimizer, loss function, and metrics. -
sum_product_network.fit(...): Trains the model on the MNIST training dataset. -
sum_product_network.evaluate(...): Evaluates the trained model on the MNIST test dataset.
Overall, this code sets up a SPN model for digit classification on the MNIST dataset, compiles it, trains it, and evaluates its performance.