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. This experiment uses the same base learning rate as the noisy-Adam example to produce comparable results.
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()
In [2]:
# Builds the model and approximation variables used for the model
y_, model_variables = experiment.get_model_3layer()
approx_variables = experiment.get_pointmass_approximation_variables_3layer()
In [3]:
# Performs inference with edward's MAP class and save each model state
models = []
num_models = 10
optimizer = tf.train.AdamOptimizer(0.005)
inference_dict = {model_variables[key]: val for key, val in approx_variables.iteritems()}
for _ in range(num_models):
inference = ed.MAP(inference_dict, data={y_: model_variables['y']})
n_iter=1000
inference.initialize(n_iter=n_iter, optimizer=optimizer)
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()
models.append({key: tf.Variable(val.eval()) for key, val in approx_variables.iteritems()})
In [4]:
# Computes the accuracy of our model
accuracy, disagreement = experiment.get_metrics_ensemble(model_variables, models, num_samples=10)
tf.global_variables_initializer().run()
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}))
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'])))
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 Full_Ensemble_Adam.csv.
In [6]:
import pandas as pd
results = pd.DataFrame(columns=('run', 'samples', 'acc', 'TP', 'FN', 'TN', 'FP'))
for run in range(5):
models = []
num_models = 15
optimizer = tf.train.AdamOptimizer(0.005)
inference_dict = {model_variables[key]: val for key, val in approx_variables.iteritems()}
for _ in range(num_models):
inference = ed.MAP(inference_dict, data={y_: model_variables['y']})
n_iter=1000
inference.initialize(n_iter=n_iter, optimizer=optimizer)
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()
models.append({key: tf.Variable(val.eval()) for key, val in approx_variables.iteritems()})
for num_samples in range(15):
accuracy, disagreement = experiment.get_metrics_ensemble(model_variables, models,
num_samples=num_samples + 1)
tf.global_variables_initializer().run()
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('Full_Ensemble_Adam.csv', index=False)