-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluation_real.py
223 lines (197 loc) · 11 KB
/
evaluation_real.py
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
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
#!/usr/bin/env python
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib
from helper import *
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
from mpl_toolkits.axes_grid1 import make_axes_locatable
np.set_printoptions(precision=3)
path = "results_real/"
def plot_confusion_matrix(eval_matrix, model_name):
w,h = get_fig_dim(width=487,fraction=0.7)
fig, ax = plt.subplots(figsize=(w,h))
print("Confusion Matrix...",model_name)
labels = ["airplane", "automobile", "bird", "cat", "deer", "dog","frog","horse","ship","truck"]
#cm = confusion_matrix( eval_matrix[:,4], eval_matrix[:,5], labels= np.arange(10, dtype=int), normalize='true')
ConfusionMatrixDisplay.from_predictions(eval_matrix[:,4], eval_matrix[:,5],normalize='true', display_labels=labels, values_format='.2f', xticks_rotation='vertical', cmap='YlGnBu', ax=ax, colorbar=False)
ax.set_xlabel("Model's Prediction of Expert's Label")
ax.set_ylabel("Expert's Label Prediction")
fig.tight_layout()
plt.savefig(path+"CM_"+model_name+".pdf")
plt.show()
def plot_expert_acc(trained, baseline, baseline_name, n_experts,ax):
acc_trained = np.array([np.mean(trained[:,4]==trained[:,5], where= trained[:,2]==exp) for exp in range(n_experts)])
acc_baseline = np.array([np.mean(baseline[:,4]==baseline[:,5], where= baseline[:,2]==exp) for exp in range(n_experts)])
print("scenario 1 - #experts not displayed: ", np.arange(n_experts,dtype=int)[np.isnan(acc_trained)])
print("scenario 1 - #experts Gumbel-Max SI-SCM is more accurate for than ",baseline_name,": ", np.sum(acc_trained>acc_baseline))
H, x_edges, y_edges = np.histogram2d(acc_trained, acc_baseline, bins=(np.arange(0, 1.1, 0.025),np.arange(0, 1.1, 0.025)))
#plt.title(" Accuracy per Expert SI-SCM vs. "+ baseline_name)
X, Y = np.meshgrid(x_edges, y_edges)
im = ax.pcolormesh(X, Y, H, cmap='YlGnBu', vmin=0, vmax=6)
ax.axline((1,1),slope=1, ls="--", color="red")
ax.plot(np.nanmean(acc_baseline), np.nanmean(acc_trained), 'rx', markersize=8)
ax.set_xlabel('Accuracy '+baseline_name)
ax.set_ylabel('Accuracy G.-M. SI-SCM')
return im
def plot_expert_acc_same_group(trained, baseline, baseline_name, n_experts, ax):
acc_trained = np.array([np.mean(trained[:,4]==trained[:,5], where= (trained[:,2]==exp) & (trained[:,6]==1)) for exp in range(n_experts)])
acc_baseline = np.array([np.mean(baseline[:,4]==baseline[:,5], where= (baseline[:,2]==exp) & (baseline[:,6]==1)) for exp in range(n_experts)])
print("scenario 2 - #experts not displayed: ", np.arange(n_experts,dtype=int)[np.isnan(acc_trained)])
print("scenario 2 - #experts Gumbel-Max SI-SCM is more accurate for than ",baseline_name,": ", np.sum(acc_trained>acc_baseline))
H, x_edges, y_edges = np.histogram2d(acc_trained, acc_baseline, bins=(np.arange(0, 1.1, 0.025),np.arange(0, 1.1, 0.025)))
X, Y = np.meshgrid(x_edges, y_edges)
im = ax.pcolormesh(X, Y, H, cmap='YlGnBu', vmin=0, vmax=6)
ax.axline((1,1),slope=1, ls="--", color="red")
ax.plot(np.nanmean(acc_baseline), np.nanmean(acc_trained), 'rx', markersize=8)
ax.set_xlabel('Accuracy '+baseline_name)
ax.set_ylabel('Accuracy G.-M. SI-SCM')
return im
def plot_expert_acc_diffgroup(trained, baseline, baseline_name, n_experts):
acc_trained = np.array([np.mean(trained[:,4]==trained[:,5], where= (trained[:,2]==exp) & (trained[:,6]==0)) for exp in range(n_experts)])
acc_baseline = np.array([np.mean(baseline[:,4]==baseline[:,5], where= (baseline[:,2]==exp) & (baseline[:,6]==0)) for exp in range(n_experts)])
H, x_edges, y_edges = np.histogram2d(acc_trained, acc_baseline, bins=(np.arange(0, 1.1, 0.025),np.arange(0, 1.1, 0.025)))
X, Y = np.meshgrid(x_edges, y_edges)
plt.pcolormesh(X, Y, H, cmap='YlGnBu')
plt.axline((1,1),slope=1, ls="--", color="red")
plt.plot(np.nanmean(acc_baseline), np.nanmean(acc_trained), 'rx', markersize=8)
plt.xlabel('Accuracy '+baseline_name )
plt.ylabel('Accuracy Gumbel-Max SI-SCM')
plt.colorbar()
plt.savefig(path+"hist2d_NGE_"+baseline_name)
plt.show()
def plot_expert_acc_same_vs_diff_group( baseline, baseline_name, n_experts,ax):
acc_group = np.array([np.mean(baseline[:,4]==baseline[:,5], where= (baseline[:,2]==exp) & (baseline[:,6]==1)) for exp in range(n_experts)])
acc_not_group = np.array([np.mean(baseline[:,4]==baseline[:,5], where= (baseline[:,2]==exp) & (baseline[:,6]==0)) for exp in range(n_experts)])
print("scenario 2 vs.3 - #experts ", baseline_name," more accurate in 2 than 3: ", np.sum(acc_group>acc_not_group))
H, x_edges, y_edges = np.histogram2d(acc_group, acc_not_group, bins=(np.arange(0, 1.1, 0.025),np.arange(0, 1.1, 0.025)))
X, Y = np.meshgrid(x_edges, y_edges)
im =ax.pcolormesh(X, Y, H, cmap='YlGnBu', vmin=0, vmax=7)
ax.axline((1,1),slope=1, ls="--", color="red")
ax.plot(np.nanmean(acc_not_group), np.nanmean(acc_group), 'rx', markersize=8)
ax.set_ylabel(r"Accuracy for h,h' $\in \psi$")
ax.set_xlabel(r"Accuracy for h,h' $\in \mathcal{H}$")
return im
def print_overall_acc(trained, untrained, nb):
trained_acc = np.mean(trained[:,4]==trained[:,5])
trained_acc_g = np.mean(trained[:,4]==trained[:,5], where= trained[:,6]==1)
trained_acc_ng = np.mean(trained[:,4]==trained[:,5], where= trained[:,6]==0)
untrained_acc = np.mean(untrained[:,4]==untrained[:,5])
untrained_acc_g = np.mean(untrained[:,4]==untrained[:,5], where= untrained[:,6]==1)
untrained_acc_ng = np.mean(untrained[:,4]==untrained[:,5], where= untrained[:,6]==0)
nb_acc = np.mean(nb[:,4]==nb[:,5])
nb_acc_g = np.mean(nb[:,4]==nb[:,5], where= nb[:,6]==1)
nb_acc_ng = np.mean(nb[:,4]==nb[:,5], where= nb[:,6]==0)
print("Model\t: GM-SI-SCM \t GNB \t GNB+CNB")
print("Acc scenario 1\t:", trained_acc, untrained_acc, nb_acc)
print("Acc scenario 2\t:", trained_acc_g, untrained_acc_g, nb_acc_g)
print("Acc scnario 3 \t:", trained_acc_ng, untrained_acc_ng, nb_acc_ng)
def main():
latexify(font_size=10)
#read result files
siscm_psi_results = pd.read_csv(path+"evaluation_results_SISCM_M(Psi).csv").to_numpy()
siscm_H_results = pd.read_csv(path+"evaluation_results_SISCM_M(H).csv").to_numpy()
nb_baseline_results = pd.read_csv(path+"evaluation_results_nb_baseline.csv").to_numpy()
#gnb_results = pd.read_csv(path+"evaluation_results_proba_models.csv").to_numpy()
#Meta-data about the real experiment
n_experts = int(np.max(siscm_psi_results[:,2])+1)
n_data_test = int(np.max(siscm_psi_results[:,0])+1)
disagreement = [np.mean((siscm_H_results[:,4]!=siscm_H_results[:,3]), where=siscm_H_results[:,2]==exp) for exp in range(n_experts)]
print("# data test: ",n_data_test)
print("# experts: ",n_experts)
print("Disagreement ratio: ",np.nanmean(disagreement))
n_pred = siscm_psi_results.shape[0]
n_pred_group = int(np.sum(siscm_psi_results[:,6]==1))
n_pred_notgroup = int(np.sum(siscm_psi_results[:,6]==0))
print("#Pred: ", n_pred)
print("#Pred Group: ", n_pred_group)
print("#Pred not Group: ", n_pred_notgroup)
##################
#print overall accuracy results of the three models
print_overall_acc(siscm_psi_results, siscm_H_results, nb_baseline_results)
##################
#plot group sizes of SI-SCM M(Psi) as vertical barplot
groups = pd.read_csv(path+"SI-SCM_groups.csv",header=None).to_numpy()
group_sizes = np.sum(~np.isnan(groups), axis=1).T
w,h = get_fig_dim(width=487,fraction=0.7)
fig, axes = plt.subplots(figsize=(w,h))
axes.set_ylabel("Mutually Similar Groups")
axes.set_xlabel("Number of Experts in each Group")
# get rid of the frame
axes.spines['right'].set_visible(False)
axes.spines['top'].set_visible(False)
axes.spines['bottom'].set_visible(False)
# remove all the ticks and directly label each bar with respective value
axes.xaxis.set_ticks_position('none')
axes.xaxis.set_ticks([])
axes.yaxis.set_ticks(np.arange(group_sizes.shape[0])+1)
axes.barh(y=np.arange(group_sizes.shape[0])+1, width=group_sizes)
axes.bar_label(axes.containers[0], padding=3)
fig.tight_layout()
plt.savefig(path+"groups_hist.pdf")
plt.show()
###################
#plot confusion matrices
plot_confusion_matrix(siscm_psi_results, "Gumbel-Max SI-SCM")
plot_confusion_matrix(siscm_H_results, "GNB")
plot_confusion_matrix( nb_baseline_results, "GNB+CNB")
###################
#plot per expert accuracy
w,h = get_fig_dim(width=487,fraction=0.4)
fig, axes = plt.subplots(figsize=(w,h))
im =plot_expert_acc(siscm_psi_results, siscm_H_results, "GNB", n_experts,axes)
axes.set(aspect=1)
#plt.colorbar(im, location='right',shrink=0.7)
fig.tight_layout()
plt.savefig(path+"compare_untrained_sc1.pdf")
plt.show()
fig, axes = plt.subplots(figsize=(w,h))
im = plot_expert_acc(siscm_psi_results, nb_baseline_results, "GNB+CNB", n_experts, axes)
axes.set(aspect=1)
#plt.colorbar(im, location='right',shrink=0.7)
fig.tight_layout()
plt.savefig(path+"compare_naivebayes_sc1.pdf")
plt.show()
w, _ = get_fig_dim(width=487,fraction=0.48)
fig, axes = plt.subplots(figsize=(w,h))
axes.set(aspect=1)
#plt.colorbar(im, location='right',shrink=0.7)
divider = make_axes_locatable(axes)
ax_cb = divider.new_horizontal(size='5%', pad='10%')
fig = axes.get_figure()
fig.add_axes(ax_cb)
im =plot_expert_acc_same_group(siscm_psi_results, siscm_H_results, "GNB", n_experts,axes)
matplotlib.colorbar.ColorbarBase(ax_cb, cmap='YlGnBu', norm=matplotlib.colors.Normalize(vmin=0, vmax=6), orientation='vertical')#, ticks=[0,1,2,3,4])
fig.tight_layout()
plt.savefig(path+"compare_untrained_sc2.pdf")
plt.show()
fig, axes = plt.subplots(figsize=(w,h))
axes.set(aspect=1)
#plt.colorbar(im, location='right',shrink=0.7)
divider = make_axes_locatable(axes)
ax_cb = divider.new_horizontal(size='5%', pad='10%')
fig = axes.get_figure()
fig.add_axes(ax_cb)
im = plot_expert_acc_same_group(siscm_psi_results, nb_baseline_results, "GNB+CNB", n_experts,axes)
matplotlib.colorbar.ColorbarBase(ax_cb, cmap='YlGnBu', norm=matplotlib.colors.Normalize(vmin=0, vmax=6), orientation='vertical')#, ticks=[0,1,2,3,4])
fig.tight_layout()
plt.savefig(path+"compare_naivebayes_sc2.pdf")
plt.show()
###########
#plot GNB+CNB baseline results in groups vs across groups
w,h = get_fig_dim(width=487,fraction=0.5)
fig, axes = plt.subplots(figsize=(w,h))
axes.set(aspect=1)
divider = make_axes_locatable(axes)
ax_cb = divider.new_horizontal(size='5%', pad='10%')
fig = axes.get_figure()
fig.add_axes(ax_cb)
im = plot_expert_acc_same_vs_diff_group( nb_baseline_results, "GNB+CNB", n_experts, axes)
#plt.colorbar(im, location='right',shrink=0.7)
matplotlib.colorbar.ColorbarBase(ax_cb, cmap='YlGnBu', norm=matplotlib.colors.Normalize(vmin=0, vmax=7), orientation='vertical')
fig.tight_layout()
plt.savefig(path+"nb_baseline.pdf")
plt.show()
###################
if __name__ == "__main__":
main()