-
Notifications
You must be signed in to change notification settings - Fork 4
/
path_analysis.py
202 lines (172 loc) · 8.86 KB
/
path_analysis.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
import random
import matplotlib as mpl
import numpy as np
import torch
from torchvision import datasets, models, transforms
from tqdm import tqdm
from src import Pruner
import os
# ### Setup Imagenet
#
# ImageNet as of Oct2019 can no longer be downloaded using pytorch.
# https://github.com/pytorch/vision/issues/1453
# To download ImageNet, see http://image-net.org/.
imagenet_dir = '/home/ashkan/data/ILSVRC2012/'
transform = transforms.Compose([transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
imagenet = datasets.ImageNet(imagenet_dir, download=False, split='val', transform=transform)
classes = imagenet.classes
mpl.rcParams['figure.dpi'] = 1200
def dead_path_analysis(dead_path, path):
return (np.dot(path, dead_path)/sum(path))
def iou_path_analysis(path1, path2):
return (np.dot(path1, path2)/(sum(path1+path2)-np.dot(path1, path2)))
def get_path(attribution_name, data, model, model_sparsity_threshold):
class_id = model(data)
model.eval()
if attribution_name == "NC":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_neuron_mct(model_sparsity_threshold, debug=False)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "IntGrad":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_integrad(model_sparsity_threshold, debug=False)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "GreedyNC":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_greedy(1, 99)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "Random":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_random(model_sparsity_threshold, debug=False)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "CDRP":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_dgr(model_sparsity_threshold, debug=False)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "CDRP-R":
pruner = Pruner.Pruner(model, data, device)
pruner.prune_dgr(model_sparsity_threshold, debug=False, r=True)
path = pruner.pruned_activations_mask
pruner.remove_handles()
elif attribution_name == "DeadsPath":
pruner = Pruner.Pruner(model, data, device)
pruner.dead_neurons_path()
path = pruner.pruned_activations_mask
pruner.remove_handles()
layer_paths = []
for i in range(len(path)):
layer_paths.append(path[i].clone().cpu().detach().numpy().reshape(-1))
path = np.hstack((layer_paths[i] for i in range(len(layer_paths))))
return path, layer_paths
def analyze_paths(model, dataloader, images, classes, model_sparsity_threshold):
global id
base_dir = './paths_logs_'+str(model_sparsity_threshold)+'/'
print(base_dir)
if not os.path.isdir(base_dir):
os.makedirs(base_dir)
model = model.to(device)
model.eval()
name_methods = ["CDRP", "NC", "IntGrad", "GreedyNC", "Random", "CDRP-R"]
num_methods = len(name_methods)
dataiter = iter(dataloader)
i = 0
iou = [[0 for j in range(num_methods)] for k in range(num_methods)]
dead_portions = [0 for j in range(num_methods)]
samples_ious = np.zeros((num_samples, num_methods, num_methods))
samples_deads = np.zeros((num_samples, num_methods))
layer_based_iou = None
layer_based_dead = None
samples_layer_ious = None
samples_layer_deads = None
n_layers = None
for chosen in tqdm(range(num_samples)):
data, _ = dataiter.next()
data = data.to(device)
output = model(data.clone())
output = torch.nn.functional.softmax(output.detach(), dim=1)
predicted_logit = output.data.max(1)[1].item()
predicted_prob = output.data.max(1)[0].item()
paths = [None for j in range(num_methods)]
layer_paths = [None for j in range(num_methods)]
new_iou = [[0 for j in range(num_methods)] for k in range(num_methods)]
new_deads = [0 for j in range(num_methods)]
for j in range(num_methods):
paths[j], layer_paths[j] = get_path(name_methods[j], data.clone(), model, model_sparsity_threshold)
if n_layers is None:
n_layers = len(layer_paths[j])
layer_based_iou = [[[0 for l in range(n_layers)] for m in range(num_methods)] for k in range(num_methods)]
layer_based_dead = [[0 for l in range(n_layers)] for k in range(num_methods)]
samples_layer_ious = np.zeros((num_samples, num_methods, num_methods, n_layers))
samples_layer_deads = np.zeros((num_samples, num_methods, n_layers))
new_layer_ious = [[[0 for l in range(n_layers)] for j in range(num_methods)] for k in range(num_methods)]
new_layer_dead = [[0 for l in range(n_layers)] for k in range(num_methods)]
for j in range(num_methods):
for k in range(j, num_methods):
new_iou[j][k] = iou_path_analysis(paths[j], paths[k])
iou[j][k] += new_iou[j][k]
for l in range(n_layers):
new_layer_ious[j][k][l] = iou_path_analysis(layer_paths[j][l], layer_paths[k][l])
layer_based_iou[j][k][l] += new_layer_ious[j][k][l]
#print(name_methods[j], "vs.", name_methods[k], iou_path_analysis(paths[j], paths[k]))
dead_path, l_dead_path = get_path("DeadsPath", data.clone(), model, 0)
#print("Dead Neuron Analysis")
for j in range(num_methods):
new_deads[j] = dead_path_analysis(dead_path, paths[j])
dead_portions[j] += new_deads[j]
for l in range(n_layers):
new_layer_dead[j][l] = dead_path_analysis(l_dead_path[l], layer_paths[j][l])
layer_based_dead[j][l] += new_layer_dead[j][l]
#print("Dead Portion of", name_methods[j], dead_path_analysis(dead_path, paths[j]))
samples_ious[chosen] = np.asarray(new_iou)
samples_deads[chosen] = np.asarray(new_deads)
samples_layer_ious[chosen] = np.asarray(new_layer_ious)
samples_layer_deads[chosen] = np.asarray(new_layer_dead)
if chosen % 50 == 0:
saved_iou = np.asarray(iou)/(chosen+1)
saved_dead = np.asarray(dead_portions)/(chosen+1)
saved_layer_ious = np.asarray(layer_based_iou)/(chosen+1)
saved_layer_deads = np.asarray(layer_based_dead)/(chosen+1)
print(saved_iou)
print(saved_dead)
# print(samples_ious)
# print(samples_deads)
print(saved_layer_ious)
print(saved_layer_deads)
np.save(base_dir + '/jaccards'+str(chosen)+'.npy', saved_iou)
np.save(base_dir + '/deads'+str(chosen)+'.npy', saved_dead)
np.save(base_dir + '/layerwise_jaccards'+str(chosen)+'.npy', saved_layer_ious)
np.save(base_dir + '/layerwise_dead'+str(chosen)+'.npy', saved_layer_deads)
np.save(base_dir + '/samples_jaccards'+str(chosen)+'.npy', samples_ious)
np.save(base_dir + '/samples_deads'+str(chosen)+'.npy', samples_deads)
np.save(base_dir + '/samples_layerwise_jaccards'+str(chosen)+'.npy', samples_layer_ious)
np.save(base_dir + '/samples_layerwise_deads'+str(chosen)+'.npy', samples_layer_deads)
i += 1
iou = np.asarray(iou)/num_samples
dead_portions = np.asarray(dead_portions)/num_samples
layer_based_iou = np.asarray(layer_based_iou)/num_samples
layer_based_dead = np.asarray(layer_based_dead)/num_samples
# print(layer_based_iou)
np.save(base_dir + '/jaccards'+str('final')+'.npy', iou)
np.save(base_dir + '/deads'+str('final')+'.npy', dead_portions)
np.save(base_dir + '/layerwise_jaccards'+str('final')+'.npy', layer_based_iou)
np.save(base_dir + '/layerwise_deads'+str('final')+'.npy', layer_based_dead)
np.save(base_dir + '/samples_jaccards'+str('final')+'.npy', samples_ious)
np.save(base_dir + '/samples_deads'+str('final')+'.npy', samples_deads)
np.save(base_dir + '/samples_layerwise_jaccards'+str('final')+'.npy', samples_layer_ious)
np.save(base_dir + '/samples_layerwise_deads'+str('final')+'.npy', samples_layer_deads)
num_samples = 1000
indices = random.sample(range(0, len(imagenet)), num_samples)
dataset = torch.utils.data.Subset(imagenet, indices)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True)
model = models.vgg16(pretrained=True)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model_sparsity_threshold = 80 # 90, 99
analyze_paths(model, dataloader, indices, classes, model_sparsity_threshold)