Experiments approximating the posterior with diagonal Gaussians from SGLD samples

We start by building the model and showing the basic inference procedure and calculation of the performance on the MNIST classification and the outlier detection task. Then perform multiple runs of the model with different number of samples in the ensemble to calculate performance statistics.



In [1]:

    
# Let's first setup the libraries, session and experimental data
import experiment
import inferences
import edward as ed
import tensorflow as tf
import numpy as np
import os

s = experiment.setup()
mnist, notmnist = experiment.get_data()









    



Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz
Extracting notMNIST_data/train-images-idx3-ubyte.gz
Extracting notMNIST_data/train-labels-idx1-ubyte.gz
Extracting notMNIST_data/t10k-images-idx3-ubyte.gz
Extracting notMNIST_data/t10k-labels-idx1-ubyte.gz



In [2]:

    
# Builds the model and approximation variables used for the model
y_, model_variables = experiment.get_model_3layer()
approx_variables = experiment.get_gauss_approximation_variables_3layer()



In [3]:

    
# Performs inference with our custom inference class
inference_dict = {model_variables[key]: val for key, val in approx_variables.iteritems()}
inference = inferences.VariationalGaussSGLD(inference_dict, data={y_: model_variables['y']})
n_iter=1000

inference.initialize(n_iter=n_iter, step_size=0.005, burn_in=0)

tf.global_variables_initializer().run()
for i in range(n_iter):
    batch = mnist.train.next_batch(100)
    info_dict = inference.update({model_variables['x']: batch[0],
                                  model_variables['y']: batch[1]})
    inference.print_progress(info_dict)

inference.finalize()









    



1000/1000 [100%] ██████████████████████████████ Elapsed: 7s | Acceptance Rate: 1.000



In [4]:

    
# Computes the accuracy of our model
accuracy, disagreement = experiment.get_metrics(model_variables, approx_variables, num_samples=10)
print(accuracy.eval({model_variables['x']: mnist.test.images, model_variables['y']: mnist.test.labels}))
print(disagreement.eval({model_variables['x']: mnist.test.images, model_variables['y']: mnist.test.labels}))









    



0.7335
[ 0.00124708  0.00350972  0.00111239 ...,  0.003096    0.00367678
  0.00545061]



In [5]:

    
# Computes some statistics for the proposed outlier detection
outlier_stats = experiment.get_outlier_stats(model_variables, disagreement, mnist, notmnist)
print(outlier_stats)
print('TP/(FN+TP): {}'.format(float(outlier_stats['TP']) / (outlier_stats['TP'] + outlier_stats['FN'])))
print('FP/(FP+TN): {}'.format(float(outlier_stats['FP']) / (outlier_stats['FP'] + outlier_stats['TN'])))









    



{'FP': 6, 'TN': 9994, 'FN': 3959, 'TP': 6041}
TP/(FN+TP): 0.6041
FP/(FP+TN): 0.0006

The following cell performs multiple runs of this model with different number of samples within the ensemble to capture performance statistics. Results are saved in SampledGauss_SGLD_LR.csv.



In [6]:

    
import pandas as pd

results = pd.DataFrame(columns=('run', 'samples', 'acc', 'TP', 'FN', 'TN', 'FP'))

for run in range(5):
    inference_dict = {model_variables[key]: val for key, val in approx_variables.iteritems()}
    inference = inferences.VariationalGaussSGLD(inference_dict, data={y_: model_variables['y']})
    n_iter=1000

    inference.initialize(n_iter=n_iter, step_size=0.005, burn_in=0)

    tf.global_variables_initializer().run()
    for i in range(n_iter):
        batch = mnist.train.next_batch(100)
        info_dict = inference.update({model_variables['x']: batch[0],
                                      model_variables['y']: batch[1]})
        inference.print_progress(info_dict)

    inference.finalize()
    
    for num_samples in range(15):
        accuracy, disagreement = experiment.get_metrics(model_variables, approx_variables,
                                                        num_samples=num_samples + 1)
        acc = accuracy.eval({model_variables['x']: mnist.test.images, model_variables['y']: mnist.test.labels})
        outlier_stats = experiment.get_outlier_stats(model_variables, disagreement, mnist, notmnist)
        results.loc[len(results)] = [run, num_samples + 1, acc,
                                     outlier_stats['TP'], outlier_stats['FN'],
                                     outlier_stats['TN'], outlier_stats['FP']]
results.to_csv('SampledGauss_SGLD_LR.csv', index=False)









    



1000/1000 [100%] ██████████████████████████████ Elapsed: 8s | Acceptance Rate: 1.000
1000/1000 [100%] ██████████████████████████████ Elapsed: 8s | Acceptance Rate: 1.000
1000/1000 [100%] ██████████████████████████████ Elapsed: 9s | Acceptance Rate: 1.000
1000/1000 [100%] ██████████████████████████████ Elapsed: 11s | Acceptance Rate: 1.000
1000/1000 [100%] ██████████████████████████████ Elapsed: 11s | Acceptance Rate: 1.000