-
Notifications
You must be signed in to change notification settings - Fork 1
/
test_batch_LIME.py
322 lines (272 loc) · 12.3 KB
/
test_batch_LIME.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
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Feb 11 11:25:55 2021
@author:
"""
import torch
import os
import importlib
import open3d as o3d
import sys
import numpy as np
import time
import matplotlib.pyplot as plt
import random
from lime import lime_3d
num_drop_steps = 10
num_samples = int(10**2)
num_features = 20
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
ROOT_DIR = BASE_DIR
sys.path.append(os.path.join(ROOT_DIR, 'models'))
def detach_cpu(tensor):
return tensor.detach().numpy()
def detach_gpu(tensor):
return tensor.detach().cpu().numpy()
def take_second(elem):
return elem[1]
def take_first(elem):
return elem[0]
def gen_pc_data(ori_data,segments,explain,label,filename):
basic_path = "visu/"
color = np.zeros([ori_data.shape[0],3])
max_contri = 0
min_contri = 0
for k in explain[label]:
if k[1] > 0 and k[1] > max_contri:
max_contri = k[1]
elif k[1] < 0 and k[1] < min_contri:
min_contri = k[1]
if max_contri > 0:
positive_color_scale = 1/max_contri
else:
positive_color_scale = 0
if min_contri < 0:
negative_color_scale = 1/min_contri
else:
negative_color_scale = 0
ex_sorted = sorted(explain[label],key=take_first,reverse=False)
for i in range(segments.shape[0]):
if ex_sorted[segments[i]][1] > 0:
color[i][0] = ex_sorted[segments[i]][1] * positive_color_scale
elif ex_sorted[segments[i]][1] < 0:
color[i][2] = ex_sorted[segments[i]][1] * negative_color_scale
else:
color[i] = [0,0,0]
pc_colored = np.concatenate((ori_data,color),axis=1)
pc = o3d.geometry.PointCloud()
pc.points = o3d.utility.Vector3dVector(pc_colored[:,0:3])
pc.colors = o3d.utility.Vector3dVector(pc_colored[:,3:6])
o3d.io.write_point_cloud(basic_path+filename, pc)
print("Generate point cloud", filename, "successful!")
def reverse_points(points,segments,explain,start='positive',percentage=0.05):
res = np.copy(points)
if start == 'positive':
explain.sort(key=take_second,reverse=True)
elif start == 'negative':
explain.sort(key=take_second,reverse=False)
else:
print('Wrong start input!')
return res
to_rev_list = explain
num_of_rev = int(len(to_rev_list)*percentage)
for i in range(num_of_rev):
segment_to_rev = to_rev_list[i][0]
rev_points_index = np.argwhere(segments==segment_to_rev)
for p in rev_points_index:
res[p] = [0,0,0]
res = np.delete(res,np.argwhere(res==[0,0,0]),0)
return res
def reverse_points_random(points,percentage=0.05):
res = np.copy(points)
num_of_rev = int(points.shape[0]*percentage)
index_to_drop = np.random.choice(np.arange(points.shape[0]),size=num_of_rev,replace=False)
for i in range(num_of_rev):
res[index_to_drop[i]] = [0,0,0]
res = np.delete(res,np.argwhere(res==[0,0,0]),0)
return res
def sampling(points, sample_size):
points = np.unique(points, axis=0)
num_p = points.shape[0]
index = range(num_p)
sampled_index = random.sample(index,sample_size)
sampled = points[sampled_index]
print(np.unique(sampled,axis=0).shape)
return sampled
def get_pred(points,classifier):
points = torch.from_numpy(points)
points = points.transpose(2, 1)
points = points.float()
if torch.cuda.is_available() == True:
points = points.cuda()
classifier = classifier.cuda()
classifier = classifier.eval()
pred, bf_sftmx , _ = classifier(points)
pred_choice = pred.data.max(1)[1]
return pred_choice, SHAPE_NAMES[pred_choice], bf_sftmx
test_file = 'data/modelnet40_normal_resampled/modelnet40_test_lime.txt'
num_class = 40
target_class = 0 #Plane
experiment_dir = 'log/classification/pointnet_cls_msg'
data_dir = 'data/modelnet40_normal_resampled'
model_name = os.listdir(experiment_dir+'/logs')[0].split('.')[0]
MODEL = importlib.import_module(model_name)
SHAPE_NAMES = [line.rstrip() for line in \
open(os.path.join(BASE_DIR, 'data/shape_names.txt'))]
classifier = MODEL.get_model(num_class,normal_channel=False)
if torch.cuda.is_available() == True:
checkpoint = torch.load(str(experiment_dir) + '/checkpoints/best_model.pth')
else:
checkpoint = torch.load(str(experiment_dir) + '/checkpoints/best_model.pth',map_location=torch.device('cpu'))
print(classifier.load_state_dict(checkpoint['model_state_dict']))
layer_name = list(classifier.state_dict().keys())
selected_layer = 'fc3'
total_num_ins = 0
batch_pos_recorder = []
batch_neg_recorder = []
batch_rdm_recorder = []
R2 = 0
Mean_dis = 0
Weighted_mean_dis = 0
L1_loss = 0
weighted_L1_loss = 0
L2_loss = 0
weighted_L2_loss = 0
adjusted_R2 = 0
avg_time = 0
for line in open(test_file):
print("Processing number ", total_num_ins+1, "instances......\n")
if total_num_ins >= 3:
break
ori_class = line[:line.rfind('_')]
cur_file = data_dir + '/' + ori_class + '/' + line[:-1] + '.txt'
prototype = np.expand_dims((sampling(np.loadtxt(cur_file,delimiter=',')[:,0:3], 1024)),0)
cur_cls_num, pred_class, logits = get_pred(prototype,classifier)
save_name = str(total_num_ins) + '_' + ori_class + '_' + pred_class + '.ply'
if ori_class == pred_class: #Count sum of correctly classified instances
total_num_ins += 1
l = detach_gpu(cur_cls_num)[0]
points_for_exp = np.asarray(prototype.copy().squeeze())
if torch.cuda.is_available() == True:
classifier = classifier.cuda()
predict_fn = classifier.eval()
explainer = lime_3d.LimeImageExplainer(random_state=0,num_cluster=num_features,kernel_width=0.1)
tmp = time.time()
explaination = explainer.explain_instance(points_for_exp, predict_fn, top_labels=1, num_features=num_features, num_samples=num_samples, random_seed=0)
print ('Time consuming: ',time.time() - tmp,'s')
avg_time += time.time() - tmp
#temp, mask = explaination.get_image_and_mask(l, positive_only=False, negative_only=False, num_features=100, hide_rest=True)
#gen_pc_data(points_for_exp,explaination.segments,explaination.local_exp,l,(str(start_idx)+'_'+SHAPE_NAMES[pred_val[i-start_idx]]+'_gt_is_'+SHAPE_NAMES[l]+'_'+'_lime.ply'))
gen_pc_data(points_for_exp,explaination.segments,explaination.local_exp,l,'LIME/'+str(num_samples)+'_'+str(num_features)+'/correct/'+save_name)
scores_pos = [detach_gpu(logits[0][l])]
scores_neg = [detach_gpu(logits[0][l])]
scores_rdm = [detach_gpu(logits[0][l])]
for p in range(1,num_drop_steps+1):
print("Flipping steps: ", str(p) +" of "+ str(num_drop_steps))
res_pos = reverse_points(points_for_exp.copy(),explaination.segments,explaination.local_exp[l],start='positive',percentage=0.05*p)
_, cur_class_name, cur_logits = get_pred(np.expand_dims(res_pos,0),classifier)
cur_score = cur_logits[0][l]
scores_pos.append(detach_gpu(cur_score))
res_neg = reverse_points(points_for_exp.copy(),explaination.segments,explaination.local_exp[l],start='negative',percentage=0.05*p)
_, cur_class_name, cur_logits = get_pred(np.expand_dims(res_neg,0),classifier)
cur_score = cur_logits[0][l]
scores_neg.append(detach_gpu(cur_score))
res_rdm = reverse_points_random(points_for_exp.copy(),0.05*p)
_, cur_class_name, cur_logits = get_pred(np.expand_dims(res_rdm,0),classifier)
cur_score = cur_logits[0][l]
scores_rdm.append(detach_gpu(cur_score))
batch_pos_recorder.append(np.asarray(scores_pos))
batch_neg_recorder.append(np.asarray(scores_neg))
batch_rdm_recorder.append(np.asarray(scores_rdm))
elif ori_class != pred_class:
total_num_ins += 1
l = detach_gpu(cur_cls_num)[0]
points_for_exp = np.asarray(prototype.copy().squeeze())
if torch.cuda.is_available() == True:
classifier = classifier.cuda()
predict_fn = classifier.eval()
explainer = lime_3d.LimeImageExplainer(random_state=0,num_cluster=num_features,kernel_width=0.1)
tmp = time.time()
explaination = explainer.explain_instance(points_for_exp, predict_fn, top_labels=1, num_features=num_features, num_samples=num_samples, random_seed=0)
print ('Time consuming: ',time.time() - tmp,'s')
avg_time += time.time() - tmp
#temp, mask = explaination.get_image_and_mask(l, positive_only=False, negative_only=False, num_features=100, hide_rest=True)
#gen_pc_data(points_for_exp,explaination.segments,explaination.local_exp,l,(str(start_idx)+'_'+SHAPE_NAMES[pred_val[i-start_idx]]+'_gt_is_'+SHAPE_NAMES[l]+'_'+'_lime.ply'))
gen_pc_data(points_for_exp,explaination.segments,explaination.local_exp,l,'LIME/'+str(num_samples)+'_'+str(num_features)+'/wrong/'+save_name)
print("Local fidelity score: ", explaination.score)
Mean_dis += explaination.mean_diff
Weighted_mean_dis += explaination.weighted_mean_diff
L1_loss += explaination.L1_loss
weighted_L1_loss += explaination.weighted_L1_loss
L2_loss += explaination.L2_loss
weighted_L2_loss += explaination.weighted_L2_loss
R2 += explaination.score
adjusted_R2 += explaination.adjusted_R2
Mean_dis /= total_num_ins
Weighted_mean_dis /= total_num_ins
L1_loss /= total_num_ins
weighted_L1_loss /= total_num_ins
L2_loss /= total_num_ins
weighted_L2_loss /= total_num_ins
R2 /= total_num_ins
adjusted_R2 /= total_num_ins
avg_time /= total_num_ins
print("Average Mean Distance: ", Mean_dis)
print("Weighted Mean Distance: ", Weighted_mean_dis)
print("Average L1 loss: ", L1_loss)
print("Average weighted L1 loss: ", weighted_L1_loss)
print("Average L2 loss: ", L2_loss)
print("Average weighted L2 loss: ", weighted_L2_loss)
print("Average R2 score: ", R2)
print("Average adjusted R2: ", adjusted_R2)
print("Average processing time: ", avg_time)
batch_pos_recorder = np.asarray(batch_pos_recorder)
batch_neg_recorder = np.asarray(batch_neg_recorder)
batch_rdm_recorder = np.asarray(batch_rdm_recorder)
tmp_nor = np.concatenate((batch_pos_recorder,batch_neg_recorder,batch_rdm_recorder),axis=1)
np.save('visu/LIME/recorder'+str(num_features)+'_'+str(num_samples)+'_'+'.npy', tmp_nor)
for i in range(batch_pos_recorder.shape[0]):
bias = np.min(tmp_nor[i])
batch_pos_recorder[i] -= bias
batch_neg_recorder[i] -= bias
batch_rdm_recorder[i] -= bias
tmp_nor[i] -= bias
nor = np.max(tmp_nor[i])
batch_pos_recorder[i] /= nor
batch_neg_recorder[i] /= nor
batch_rdm_recorder[i] /= nor
res_mean_pos = np.mean(batch_pos_recorder,axis=0)
res_max_pos = np.max(batch_pos_recorder,axis=0)
res_min_pos = np.min(batch_pos_recorder,axis=0)
res_mean_neg = np.mean(batch_neg_recorder,axis=0)
res_max_neg = np.max(batch_neg_recorder,axis=0)
res_min_neg = np.min(batch_neg_recorder,axis=0)
res_mean_rdm = np.mean(batch_rdm_recorder,axis=0)
res_max_rdm = np.max(batch_rdm_recorder,axis=0)
res_min_rdm = np.min(batch_rdm_recorder,axis=0)
#Draw pictures
x = np.arange(0, (num_drop_steps+1)*5, 5)
flipping_figure = plt.figure(1)
plt.subplots_adjust(left=None, bottom=None, right=None, top=None,
wspace=0.3, hspace=0.35)
ax = plt.subplot(111, facecolor='lightgrey')
plt.ylim(0,1.1)
ax.plot(x,res_mean_pos,'-r',linewidth=1,label='Positive')
ax.plot(x,res_mean_neg,'-b',linewidth=1,label='Negative')
ax.plot(x,res_mean_rdm,'-g',linewidth=1,label='Random')
ax.grid()
plt.xlabel("Percentage of clusters reversed",fontsize=13)
plt.ylabel("Average Scores",fontsize=13)
plt.savefig('visu/LIME_Quan/oneline_lime_'+str(num_features)+'_'+str(num_samples)+'_'+'.jpg')
ax.plot(x,res_max_pos,'-',color='#FFC0CB',linewidth=0.5)
ax.plot(x,res_min_pos,'-',color='#FFC0CB',linewidth=0.5)
ax.fill_between(x, res_min_pos, res_max_pos, facecolor='#FFC0CB', alpha=0.3)
ax.plot(x,res_max_neg,'-',color='#6495ED',linewidth=0.5)
ax.plot(x,res_min_neg,'-',color='#6495ED',linewidth=0.5)
ax.fill_between(x, res_min_neg, res_max_neg, facecolor='#6495ED', alpha=0.3)
ax.plot(x,res_max_rdm,'-',color='#32CD32',linewidth=0.5)
ax.plot(x,res_min_rdm,'-',color='#32CD32',linewidth=0.5)
ax.fill_between(x, res_min_rdm, res_max_rdm, facecolor='#32CD32', alpha=0.3)
#ax.legend(fontsize=13)
plt.savefig('visu/LIME_Quan/lime_'+str(num_features)+'_'+str(num_samples)+'_'+'.jpg')