In [1]:
%matplotlib inline
%load_ext autoreload
%autoreload 2
from __future__ import division
from __future__ import print_function
from __future__ import absolute_import
from __future__ import unicode_literals
import numpy as np
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
sns.set(color_codes=True)
np.random.seed(42)
In [2]:
f = np.load('../output/spam_results.npz')
orig_results = f['orig_results']
flipped_results = f['flipped_results']
fixed_influence_loo_results = f['fixed_influence_loo_results']
fixed_loss_results = f['fixed_loss_results']
fixed_random_results = f['fixed_random_results']
In [3]:
sns.set_style('white')
fontsize=14
x = np.arange(0, 0.35, 0.05)
metric_idx = 0
flip_mean = np.mean(flipped_results, axis=1)
flip_error = np.std(flipped_results, axis=1) / np.sqrt(flipped_results.shape[1])
fig, axs = plt.subplots(1, 2, figsize=(7,3.5))
flip_idx = 1
print("Fraction of flips: %s" % ((flip_idx + 1) * 0.05))
num_train_examples = 4137
num_flips = int(num_train_examples / 20) * (flip_idx + 1)
for ax, metric_idx in zip(axs, [2, 0]):
for results, label in [
(fixed_influence_loo_results, 'Influence'),
(fixed_loss_results, 'Loss'),
(fixed_random_results, 'Random')]:
mean = np.mean(results, axis=2)[flip_idx, :, metric_idx]
error = np.std(results, axis=2)[flip_idx, :, metric_idx] / np.sqrt(results.shape[2])
mean = np.insert(mean, 0, flip_mean[flip_idx, metric_idx])
error = np.insert(error, 0, flip_error[flip_idx, metric_idx])
if metric_idx == 0:
mean /= num_flips
error /= num_flips
ax.errorbar(
x,
mean,
error,
label=label)
ax.set_xlim((0, 0.30))
ax.set_xticks(x)
ax.set_xlabel('Fraction of train data checked', fontsize=fontsize)
if metric_idx > 0:
ax.plot(
x,
[orig_results[metric_idx - 1]] * len(x),
'--',
color='k',
label='Clean data')
ax.legend(loc=(0.52,0.30), fontsize=fontsize-3)
ax.set_ylim((0.885, 0.98))
ax.set_ylabel('Test accuracy', fontsize=fontsize)
else:
ax.set_ylim((0, 1))
ax.set_ylabel('Fraction of flips fixed', fontsize=fontsize)
plt.tight_layout()
plt.subplots_adjust(wspace=0.28)
# plt.savefig('../figs/fig-spam.png', dpi=600, bbox_inches='tight')
