ann-info-flow/plot_utils.py at main · milesmahon/ann-info-flow · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
#!/usr/bin/env python3

from __future__ import print_function, division

import numpy as np
import matplotlib.pyplot as plt
import os
import sys
import joblib
from param_utils import init_params


def init_plots(vis):
    # Initialize the plotting axes and add them to the namespace
    vis.fig = plt.figure(figsize=(15, 5))

    vis.fig.canvas.toolbar_visible = False
    vis.fig.canvas.header_visible = False
    vis.fig.canvas.footer_visible = False
    vis.fig.canvas.resizable = False

    vis.ax_weights = vis.fig.add_subplot(1, 3, 1, aspect='equal')
    vis.ax_weights.set_axis_off()
    vis.ax_weights.set_title('ANN Weights')

    vis.ax_bias_flows = vis.fig.add_subplot(1, 3, 2, aspect='equal')
    vis.ax_bias_flows.set_axis_off()
    vis.ax_bias_flows.set_title('Weighted Bias Flows')

    vis.ax_acc_flows = vis.fig.add_subplot(1, 3, 3, aspect='equal')
    vis.ax_acc_flows.set_axis_off()
    vis.ax_acc_flows.set_title('Weighted Accuracy Flows')

    plt.tight_layout()


def plot_ann(layer_sizes, weights, plot_params=None, ax=None, flow_type='bias', info_method='', label_name=''):
    """
    Plots a visualization of a trained neural network with given weights.

    Based on a gist by Colin Raffel
    (https://gist.github.com/craffel/2d727968c3aaebd10359)
    """

    if ax is None:
        plt.figure(figsize=(6, 6))
        ax = plt.gca()
        plt.axis('off')
    if plot_params is None:
        plot_params = (0.1, 0.9, 0.9, 0.1)
    left, right, top, bottom = plot_params

    num_layers = len(layer_sizes)
    v_spacing = (top - bottom) / max(layer_sizes)
    h_spacing = (right - left) / (num_layers - 1)

    # Plot circles for neurons
    for n, layer_size in enumerate(layer_sizes):
        layer_top = v_spacing * (layer_size - 1) / 2 + (top + bottom) / 2
        for m in range(layer_size):
            circle = plt.Circle((n*h_spacing + left, layer_top - m*v_spacing),
                                h_spacing/4, color='w', ec='k', zorder=4)
            ax.add_artist(circle)

    # Normalize edge weights to lie between -1 and +1
    # max_weight = max(np.abs(w).max() for w in weights)
    # normalized_weights = [w / max_weight for w in weights]
    normalized_weights = [w / 0.5 for w in weights]  # MM normalizing to .5, since that's around the max we see across both tasks
    # normalized_weights = weights  # MM normalizing to 1, since max entropy for this task is 1
    alpha0 = 0.1  # MM originally 0.3
    t0, tmax = (0.5, 5)

    # Plot lines for edges
    for n, (layer_size_a, layer_size_b) in enumerate(zip(layer_sizes[:-1], layer_sizes[1:])):
        layer_top_a = v_spacing*(layer_size_a - 1)/2 + (top + bottom)/2
        layer_top_b = v_spacing*(layer_size_b - 1)/2 + (top + bottom)/2
        for m in range(layer_size_a):
            for o in range(layer_size_b):
                weight = normalized_weights[n].reshape((layer_size_b, layer_size_a))[o, m]
                if flow_type == 'acc':
                    color = 'C0'
                elif flow_type == 'bias':
                    color = 'C1'
                else:
                    color = 'C2'
                alpha = alpha0 + abs(weight) * (1 - alpha0)
                thickness = t0 + abs(weight) * (tmax - t0)

                line = plt.Line2D([n*h_spacing + left, (n + 1)*h_spacing + left],
                                  [layer_top_a - m*v_spacing, layer_top_b - o*v_spacing],
                                  color=color, alpha=alpha, linewidth=thickness)
                ax.add_artist(line)
    if label_name != '':
        ax.set_title(label_name, fontsize=18)
    elif flow_type == 'bias':
        ax.set_title("Bias flow visualization\n(Dataset: Synthetic, MI est: %s)" % info_method ,fontsize=18)
    else:
        ax.set_title("Accuracy flow visualization\n(Dataset: Synthetic, MI est: %s)" % info_method ,fontsize=18)

    return ax


if __name__ == '__main__':

    if len(sys.argv) == 1:
        print("Please provide the dataset (adult, tinyscm), MI estimation method (corr,linear-svm,kernel-svm), and run number.")
        exit()

    if len(sys.argv) > 1:
        dataset = sys.argv[1]
    else:
        params = init_params()
        dataset = params.dataset

    if len(sys.argv) > 2:
        subfolder = sys.argv[2]
    else:
        subfolder = ''

    if len(sys.argv) > 3:
        run = int(sys.argv[3])
    else:
        run = 0

    #subfolder = 'linear-svm'
    results_dir = 'results-%s' % dataset
    results_subfolder = os.path.join(results_dir, subfolder)

    params=init_params(dataset=dataset)
    nets=joblib.load(params.annfile)
    layer_sizes = nets[run].layer_sizes

    #weights = nets[0].get_weights() # Plot just the network weights.
    bias_flows=joblib.load(results_subfolder + '/analyzed-data.pkl')[run][2][:-1]
    weights = [abs(w) for w in bias_flows]
    plot_ann(layer_sizes, weights, flow_type='bias', info_method=subfolder)

    #weights = nets[0].get_weights()
    acc_flows=joblib.load(results_subfolder + '/analyzed-data.pkl')[run][5][:-1]
    weights = [abs(w) for w in acc_flows]
    plot_ann(layer_sizes, weights, flow_type='acc', info_method=subfolder)
    plt.show()