-
Notifications
You must be signed in to change notification settings - Fork 11
/
openmax.py
305 lines (250 loc) · 13 KB
/
openmax.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
"""
This is an implementation of OpenMax from
A. Bendale, T. Boult [Towards Open Set Deep Networks](http://vast.uccs.edu/~abendale/papers/0348.pdf). CVPR 2016
The implementation is checked against the original implementation https://github.com/abhijitbendale/OSDN
we do not rely on third party libraries to perform the calculations. Everything that is necessary for this
is already implemented here from scratch.
Notes:
- The original method was tested with AlexNet, as with all other methods we test this with ResNet and VGG.
- Using the terminology of the paper, here, we have only one channel. But technically, the authors
meant batch_size rather than "channels". The batches came from random crops in the Caffe python sample code.
This is not an efficient implementation. Specifically, OTModelWrapper.subnetwork_eval can be improved.
We decided to keep the structure to the original source code as much as possible.
"""
from __future__ import print_function
import numpy as np
from scipy.stats import exponweib
from tqdm import tqdm
from termcolor import colored
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from methods import AbstractMethodInterface, AbstractModelWrapper, SVMLoss
from methods.base_threshold import ProbabilityThreshold
from utils.iterative_trainer import IterativeTrainerConfig
from utils.logger import Logger
from datasets import MirroredDataset
import global_vars as Global
import scipy.stats
class LibNotMR(object):
"""
Instead of using LibMR (https://github.com/abhijitbendale/OSDN/tree/master/libMR) we implemented
the simple operations with Scipy. The output is checked against the original library for verification.
"""
def __init__(self, tailsize = 20):
self.tailsize = tailsize
self.min_val = None
self.translation = 10000 # this constant comes from the library.
# it only makes small numerical differences.
# we keep it for authenticity.
self.a = 1
self.loc = 0
self.c = None
self.scale = None
def fit_high(self, inputs):
inputs = inputs.numpy()
tailtofit = sorted(inputs)[-self.tailsize:]
self.min_val = np.min(tailtofit)
new_inputs = [i + self.translation - self.min_val for i in tailtofit]
params = scipy.stats.exponweib.fit(new_inputs, floc=0, f0=1)
self.c = params[1]
self.scale = params[3]
def w_score(self, inputs):
new_inputs = inputs + self.translation - self.min_val
new_score = scipy.stats.exponweib.cdf(new_inputs, a=self.a, c=self.c, loc=self.loc, scale=self.scale)
return new_score
def serialize(self):
return torch.FloatTensor([self.min_val, self.c, self.scale])
def deserialize(self, params):
self.min_val = params[0].item()
self.c = params[1].item()
self.scale = params[2].item()
def __str__(self):
return 'Weib: C=%.2f scale=%.2f min_val=%.2f'%(self.c, self.scale, self.min_val)
def distance_measure(ref, target):
ref = ref.unsqueeze(0)
target = target.unsqueeze(0)
euc_dist = torch.pairwise_distance(ref, target)
cos_dist = 1 - torch.nn.functional.cosine_similarity(ref, target)
query_dist = euc_dist.item()/200.0 + cos_dist.item() # the coefficients are taken from the original implementation.
return query_dist
class OTModelWrapper(AbstractModelWrapper):
""" The wrapper class for H.
"""
def __init__(self, base_model, mav, mr_models, alpharank = 10):
super(OTModelWrapper, self).__init__(base_model)
# Serialize the MR models.
n_classes = mav.size(0)
weib_params = torch.FloatTensor(n_classes, 3).fill_(0)
for cl in range(n_classes):
weib_params[cl] = mr_models[cl].serialize()
self.H = nn.Sequential(
nn.BatchNorm1d(n_classes+1), # Helps with faster convergence.
nn.Linear(n_classes+1, 1),
)
self.H.register_buffer('MAV', mav.clone())
self.H.register_buffer('mr_params', weib_params.clone())
self.alpharank = alpharank
self.reload_mr()
def reload_mr(self):
self.mr_models = []
n_classes = self.H.MAV.size(0)
for cl in range(n_classes):
mr_model = LibNotMR()
mr_model.deserialize(self.H.mr_params[cl])
self.mr_models.append(mr_model)
def set_eval_direct(self, eval_direct):
super(OTModelWrapper, self).set_eval_direct(eval_direct)
self.reload_mr()
def eval(self):
super(OTModelWrapper, self).eval()
self.reload_mr()
def subnetwork_eval(self, x):
base_output = self.base_model(x, softmax=False)
n_instances, n_classes = base_output.size()
alpharank = self.alpharank
output = base_output.new(n_instances, n_classes + 1).fill_(0)
for i in range(n_instances):
openmax_activation = []
openmax_unknown = []
activation = base_output[i]
alpha_weights = [((alpharank+1) - t)/float(alpharank) for t in range(1, alpharank+1)]
ranked_alpha = base_output.new_zeros(n_classes)
_, ranked_list = activation.sort()
for k in range(len(alpha_weights)):
ranked_alpha[ranked_list[k]] = alpha_weights[k]
for j in range(n_classes):
query_dist = distance_measure(self.H.MAV[j, :], activation)
wscore = base_output.new([self.mr_models[j].w_score(query_dist)])
modified_score = activation[j] * ( 1 - wscore*ranked_alpha[j] )
openmax_activation += [modified_score]
openmax_unknown += [activation[j] - modified_score]
openmax_activation = base_output.new(openmax_activation)
openmax_unknown = base_output.new(openmax_unknown)
activation_exp = openmax_activation.exp()
total_sum = activation_exp.sum().item() + openmax_unknown.sum().exp().item()
probs = activation_exp / total_sum
unks = openmax_unknown.sum().exp() / total_sum
output[i, 0:n_classes] = probs
output[i, n_classes] = unks
return output.detach()
def wrapper_eval(self, x):
output = self.H(x)
return output
def classify(self, x):
return (x > 0).long()
class OpenMax(ProbabilityThreshold):
def __init__(self, args):
super(OpenMax, self).__init__(args)
self.tailsize = 20 # default in the paper.
def method_identifier(self):
output = "OpenMax"
if len(self.add_identifier) > 0:
output = output + "/" + self.add_identifier
return output
def propose_H(self, dataset):
# Do the PBThresh propose_h first.
super(OpenMax, self).propose_H(dataset)
"""
For each class we calculate the mean of logits over _correct classifications_.
"""
n_classes = self.base_model.output_size()[1].item()
mav = torch.FloatTensor(n_classes, n_classes).fill_(0).to(self.args.device)
mav_count = torch.LongTensor(n_classes).fill_(0).to(self.args.device)
data_loader = DataLoader(dataset, batch_size=self.args.batch_size, num_workers=self.args.workers, pin_memory=True)
with torch.set_grad_enabled(False):
with tqdm(total=len(data_loader)) as pbar:
pbar.set_description('Calculating MAV')
for i, (image, label) in enumerate(data_loader):
pbar.update()
input, target = image.to(self.args.device), label.to(self.args.device)
prediction = self.base_model.forward(input, softmax=False)
_, max_ind = torch.max(prediction, dim=1)
for sub_ind in range(len(image)):
if max_ind[sub_ind].item() == label[sub_ind].item():
mav[label[sub_ind], :].add_(prediction[sub_ind])
mav_count[label[sub_ind]].add_(1)
assert (mav_count>0).all().item() == 1, 'Something wrong with the classes! Need at least one sample/class.'
mav.div_(mav_count.float().unsqueeze(1).expand_as(mav))
self.mav = mav # Store the MAV.
# In the original source code calculates three different distance measures
# Cosine, Euclidean, and a combination, however, the default mode of
# distance is set to the combination and the other two are not used.
distance_values = [[] for i in range(n_classes)]
with torch.set_grad_enabled(False):
with tqdm(total=len(data_loader)) as pbar:
pbar.set_description('Calculating the distances')
for i, (image, label) in enumerate(data_loader):
pbar.update()
input, target = image.to(self.args.device), label.to(self.args.device)
prediction = self.base_model.forward(input, softmax=False)
_, max_ind = torch.max(prediction, dim=1)
for sub_ind in range(len(image)):
if max_ind[sub_ind].item() == label[sub_ind].item():
query_dist = distance_measure(mav[label[sub_ind], :], prediction[sub_ind])
distance_values[label[sub_ind]].append(query_dist)
torch_values = [torch.FloatTensor(dv) for dv in distance_values]
self.weib_models = []
with tqdm(total=n_classes) as pbar:
pbar.set_description('Learning the Weibull model')
for cl in range(n_classes):
pbar.update()
mr_model = LibNotMR(tailsize = self.tailsize)
mr_model.fit_high(torch_values[cl])
self.weib_models.append(mr_model)
def get_H_config(self, dataset, will_train=True):
print("Preparing training D1+D2 (H)")
print("Mixture size: %s"%colored('%d'%len(dataset), 'green'))
# 80%, 20% for local train+test
train_ds, valid_ds = dataset.split_dataset(0.8)
if self.args.D1 in Global.mirror_augment:
print(colored("Mirror augmenting %s"%self.args.D1, 'green'))
new_train_ds = train_ds + MirroredDataset(train_ds)
train_ds = new_train_ds
# Initialize the multi-threaded loaders.
train_loader = DataLoader(train_ds, batch_size=self.args.batch_size, shuffle=True, num_workers=self.args.workers, pin_memory=True)
valid_loader = DataLoader(valid_ds, batch_size=self.args.batch_size, shuffle=True, num_workers=self.args.workers, pin_memory=True)
# Set up the criterion
# margin must be non-zero.
criterion = SVMLoss(margin=1.0).cuda()
# Set up the model
model = OTModelWrapper(self.base_model, self.mav, self.weib_models).to(self.args.device)
old_valid_loader = valid_loader
if will_train:
# cache the subnetwork for faster optimization.
from methods import get_cached
from torch.utils.data.dataset import TensorDataset
trainX, trainY = get_cached(model, train_loader, self.args.device)
validX, validY = get_cached(model, valid_loader, self.args.device)
new_train_ds = TensorDataset(trainX, trainY)
new_valid_ds = TensorDataset(validX, validY)
# Initialize the new multi-threaded loaders.
train_loader = DataLoader(new_train_ds, batch_size=2048, shuffle=True, num_workers=0, pin_memory=False)
valid_loader = DataLoader(new_valid_ds, batch_size=2048, shuffle=True, num_workers=0, pin_memory=False)
# Set model to direct evaluation (for cached data)
model.set_eval_direct(True)
# Set up the config
config = IterativeTrainerConfig()
base_model_name = self.base_model.__class__.__name__
if hasattr(self.base_model, 'preferred_name'):
base_model_name = self.base_model.preferred_name()
config.name = '_%s[%s](%s->%s)'%(self.__class__.__name__, base_model_name, self.args.D1, self.args.D2)
config.train_loader = train_loader
config.valid_loader = valid_loader
config.phases = {
'train': {'dataset' : train_loader, 'backward': True},
'test': {'dataset' : valid_loader, 'backward': False},
'testU': {'dataset' : old_valid_loader, 'backward': False},
}
config.criterion = criterion
config.classification = True
config.cast_float_label = True
config.stochastic_gradient = True
config.visualize = not self.args.no_visualize
config.model = model
config.optim = optim.Adagrad(model.H.parameters(), lr=1e-1, weight_decay=1.0/len(train_ds))
config.scheduler = optim.lr_scheduler.ReduceLROnPlateau(config.optim, patience=10, threshold=1e-1, min_lr=1e-8, factor=0.1, verbose=True)
config.logger = Logger()
config.max_epoch = 100
return config