PyMC3 simulation of MNIST56

Author: dizcza

JAGS/mcmc_jags.R renamed to MCMC/mnist56_jags.R

@@ -0,0 +1,57 @@
+import pymc3 as pm
+import theano.tensor as tt
+import pandas as pd
+import numpy as np
+
+mnist56_url = {
+    "train": "https://www.dropbox.com/s/l7uppxi1wvfj45z/MNIST56_train.csv?dl=1",
+    "test": "https://www.dropbox.com/s/399gkdk9bhqvz86/MNIST56_test.csv?dl=1"
+}
+
+
+def prepare_data(fold_name="train"):
+    """
+    Processes MNIST56 dataset.
+    :param fold_name: either 'train' or 'test'
+    :return: binary pixels matrix of size (n, 25) and corresponding labels (vector of size n)
+    """
+    dataframe = pd.read_csv(mnist56_url[fold_name])
+    x = dataframe.iloc[:, :25]  # MNIST 5x5 flatten
+    x = (x > 0).astype(int)  # binarize pixels
+    y = dataframe.iloc[:, 25]  # labels 0 or 1
+    return x, y
+
+
+def predict(x_data, w):
+    """
+    :param x_data: matrix of size (n, 25)
+    :param w: matrix of size (25, 2)
+    :return: vector of size n of predicted labels
+    """
+    logit_vec = np.dot(x_data, w)
+    y_pred = logit_vec[:, 1] > logit_vec[:, 0]
+    y_pred = y_pred.astype(int)
+    return y_pred
+
+
+def main():
+    np.random.seed(113)
+    x_train, y_train = prepare_data(fold_name="train")
+    model = pm.Model()
+
+    with model:
+        w = pm.Bernoulli('w', p=0.5, shape=(25, 2))
+        logit_vec = tt.dot(x_train, w)
+        logit_p = logit_vec[:, 1] - logit_vec[:, 0]  # logit of p(y=1)
+        y_obs = pm.Bernoulli('y_obs', logit_p=logit_p, observed=y_train)
+        trace = pm.sample(draws=10, njobs=1, chains=1, n_init=1000, tune=0)
+    w_mean = trace.get_values('w').mean(axis=0)
+    w_binary = (w_mean > 0.5).astype(int)
+    x_test, y_test = prepare_data(fold_name="test")
+    y_pred = predict(x_data=x_test, w=w_binary)
+    accuracy = (y_pred == y_test).sum() / len(y_test)
+    print("Test accuracy: {}".format(accuracy))
+
+
+if __name__ == '__main__':
+    main()

MCMC/mnist56_pymc.py

@@ -54,4 +54,5 @@ trainer.train(n_epoch=100)
 
 * Train plots. Navigate to [http://ec2-34-227-113-244.compute-1.amazonaws.com:8099](http://ec2-34-227-113-244.compute-1.amazonaws.com:8099) and choose the Environment you want (`main` env is empty).
 * For your local results, go to [http://localhost:8097](http://localhost:8097)
-* JAGS simulation in _R_: [paper](JAGS/paper.pdf), [code](JAGS/mcmc_jags.R)
+* JAGS simulation in _R_: [paper](MCMC/paper.pdf), [source](MCMC/mnist56_jags.R)
+* PyMC3 simulation in Python: [source](MCMC/mnist56_pymc.py)