-
Notifications
You must be signed in to change notification settings - Fork 13
/
test_uncertainty.py
264 lines (241 loc) · 11.5 KB
/
test_uncertainty.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
import torch
import numpy as np
import random
from torch.autograd import Variable
import torch.nn as nn
import torch.optim
import json
import torch.utils.data.sampler
import os
import glob
import time
import configs
import backbone
import data.feature_loader as feat_loader
from data.datamgr import SetDataManager
from methods.baselinetrain import BaselineTrain
from methods.baselinefinetune import BaselineFinetune
from methods.protonet import ProtoNet
from methods.DKT import DKT
from methods.matchingnet import MatchingNet
from methods.relationnet import RelationNet
from methods.maml import MAML
from io_utils import model_dict, get_resume_file, parse_args, get_best_file , get_assigned_file
def _set_seed(seed, verbose=True):
if(seed!=0):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if(verbose): print("[INFO] Setting SEED: " + str(seed))
else:
if(verbose): print("[INFO] Setting SEED: None")
class ECELoss(nn.Module):
""" Calculates the Expected Calibration Error of a model.
(This isn't necessary for temperature scaling, just a cool metric).
The input to this loss is the logits of a model, NOT the softmax scores.
This divides the confidence outputs into equally-sized interval bins.
In each bin, we compute the confidence gap:
bin_gap = | avg_confidence_in_bin - accuracy_in_bin |
We then return a weighted average of the gaps, based on the number
of samples in each bin.
Adapted from: https://github.com/gpleiss/temperature_scaling
See: Naeini, Mahdi Pakdaman, Gregory F. Cooper, and Milos Hauskrecht.
"Obtaining Well Calibrated Probabilities Using Bayesian Binning." AAAI.
2015.
"""
def __init__(self, n_bins=15):
"""
n_bins (int): number of confidence interval bins
"""
super(ECELoss, self).__init__()
bin_boundaries = torch.linspace(0, 1, n_bins + 1)
self.bin_lowers = bin_boundaries[:-1]
self.bin_uppers = bin_boundaries[1:]
def calibrate(self, logits, labels, iterations=50, lr=0.01):
temperature_raw = torch.ones(1, requires_grad=True, device="cuda")
nll_criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.LBFGS([temperature_raw], lr=lr, max_iter=iterations)
softplus = nn.Softplus() #temperature must be > zero, Softplus could be used
def closure():
if torch.is_grad_enabled(): optimizer.zero_grad()
#loss = nll_criterion(logits / softplus(temperature_raw.expand_as(logits)), labels)
loss = nll_criterion(logits / temperature_raw.expand_as(logits), labels)
if loss.requires_grad: loss.backward()
return loss
optimizer.step(closure)
return temperature_raw
def forward(self, logits, labels, temperature=1.0, onevsrest=False):
logits_scaled = logits / temperature
if(onevsrest):
softmaxes = torch.sigmoid(logits_scaled) / torch.sum(torch.sigmoid(logits_scaled), dim=1, keepdim=True)
else:
softmaxes = torch.softmax(logits_scaled, dim=1)
confidences, predictions = torch.max(softmaxes, 1)
accuracies = predictions.eq(labels)
ece = torch.zeros(1, device=logits.device)
for bin_lower, bin_upper in zip(self.bin_lowers, self.bin_uppers):
# Calculated |confidence - accuracy| in each bin
in_bin = confidences.gt(bin_lower.item()) * confidences.le(bin_upper.item())
prop_in_bin = in_bin.float().mean()
if prop_in_bin.item() > 0:
accuracy_in_bin = accuracies[in_bin].float().mean()
avg_confidence_in_bin = confidences[in_bin].mean()
ece += torch.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin
return ece
def get_logits_targets(params):
acc_all = []
iter_num = 600
few_shot_params = dict(n_way = params.test_n_way , n_support = params.n_shot)
if params.dataset in ['omniglot', 'cross_char']:
assert params.model == 'Conv4' and not params.train_aug ,'omniglot only support Conv4 without augmentation'
params.model = 'Conv4S'
if params.method == 'baseline':
model = BaselineFinetune( model_dict[params.model], **few_shot_params )
elif params.method == 'baseline++':
model = BaselineFinetune( model_dict[params.model], loss_type = 'dist', **few_shot_params )
elif params.method == 'protonet':
model = ProtoNet( model_dict[params.model], **few_shot_params )
elif params.method == 'DKT':
model = DKT(model_dict[params.model], **few_shot_params)
elif params.method == 'matchingnet':
model = MatchingNet( model_dict[params.model], **few_shot_params )
elif params.method in ['relationnet', 'relationnet_softmax']:
if params.model == 'Conv4':
feature_model = backbone.Conv4NP
elif params.model == 'Conv6':
feature_model = backbone.Conv6NP
elif params.model == 'Conv4S':
feature_model = backbone.Conv4SNP
else:
feature_model = lambda: model_dict[params.model]( flatten = False )
loss_type = 'mse' if params.method == 'relationnet' else 'softmax'
model = RelationNet( feature_model, loss_type = loss_type , **few_shot_params )
elif params.method in ['maml' , 'maml_approx']:
backbone.ConvBlock.maml = True
backbone.SimpleBlock.maml = True
backbone.BottleneckBlock.maml = True
backbone.ResNet.maml = True
model = MAML( model_dict[params.model], approx = (params.method == 'maml_approx') , **few_shot_params )
if params.dataset in ['omniglot', 'cross_char']: #maml use different parameter in omniglot
model.n_task = 32
model.task_update_num = 1
model.train_lr = 0.1
else:
raise ValueError('Unknown method')
model = model.cuda()
checkpoint_dir = '%s/checkpoints/%s/%s_%s' %(configs.save_dir, params.dataset, params.model, params.method)
if params.train_aug:
checkpoint_dir += '_aug'
if not params.method in ['baseline', 'baseline++'] :
checkpoint_dir += '_%dway_%dshot' %( params.train_n_way, params.n_shot)
#modelfile = get_resume_file(checkpoint_dir)
if not params.method in ['baseline', 'baseline++'] :
if params.save_iter != -1:
modelfile = get_assigned_file(checkpoint_dir,params.save_iter)
else:
modelfile = get_best_file(checkpoint_dir)
if modelfile is not None:
tmp = torch.load(modelfile)
model.load_state_dict(tmp['state'])
else:
print("[WARNING] Cannot find 'best_file.tar' in: " + str(checkpoint_dir))
split = params.split
if params.save_iter != -1:
split_str = split + "_" +str(params.save_iter)
else:
split_str = split
if params.method in ['maml', 'maml_approx', 'DKT']: #maml do not support testing with feature
if 'Conv' in params.model:
if params.dataset in ['omniglot', 'cross_char']:
image_size = 28
else:
image_size = 84
else:
image_size = 224
datamgr = SetDataManager(image_size, n_eposide = iter_num, n_query = 15 , **few_shot_params)
if params.dataset == 'cross':
if split == 'base':
loadfile = configs.data_dir['miniImagenet'] + 'all.json'
else:
loadfile = configs.data_dir['CUB'] + split +'.json'
elif params.dataset == 'cross_char':
if split == 'base':
loadfile = configs.data_dir['omniglot'] + 'noLatin.json'
else:
loadfile = configs.data_dir['emnist'] + split +'.json'
else:
loadfile = configs.data_dir[params.dataset] + split + '.json'
novel_loader = datamgr.get_data_loader( loadfile, aug = False)
if params.adaptation:
model.task_update_num = 100 #We perform adaptation on MAML simply by updating more times.
model.eval()
logits_list = list()
targets_list = list()
for i, (x,_) in enumerate(novel_loader):
logits = model.get_logits(x).detach()
targets = torch.tensor(np.repeat(range(params.test_n_way), model.n_query)).cuda()
logits_list.append(logits) #.cpu().detach().numpy())
targets_list.append(targets) #.cpu().detach().numpy())
else:
novel_file = os.path.join( checkpoint_dir.replace("checkpoints","features"), split_str +".hdf5")
cl_data_file = feat_loader.init_loader(novel_file)
logits_list = list()
targets_list = list()
n_query = 15
n_way = few_shot_params['n_way']
n_support = few_shot_params['n_support']
class_list = cl_data_file.keys()
for i in range(iter_num):
#----------------------
select_class = random.sample(class_list,n_way)
z_all = []
for cl in select_class:
img_feat = cl_data_file[cl]
perm_ids = np.random.permutation(len(img_feat)).tolist()
z_all.append( [ np.squeeze( img_feat[perm_ids[i]]) for i in range(n_support+n_query) ] ) # stack each batch
z_all = torch.from_numpy(np.array(z_all))
model.n_query = n_query
logits = model.set_forward(z_all, is_feature = True).detach()
targets = torch.tensor(np.repeat(range(n_way), n_query)).cuda()
logits_list.append(logits)
targets_list.append(targets)
#----------------------
return torch.cat(logits_list, 0), torch.cat(targets_list, 0)
def main():
params = parse_args('test')
seed = params.seed
repeat = params.repeat
# 1. Find the value of temperature (calibration)
print("Calibration: finding temperature hyperparameter...")
ece_module = ECELoss()
temperature_list = list()
for _ in range(repeat):#repeat):
_set_seed(0) # random seed
logits, targets = get_logits_targets(parse_args('test'))
temperature = ece_module.calibrate(logits, targets, iterations=300, lr=0.01).item()
if(temperature>0): temperature_list.append(temperature)
print("Calibration: temperature", temperature, "; mean temperature", np.mean(temperature_list))
# Filtering invalid temperatures (e.g. temp<0)
if(len(temperature_list)>0):temperature = np.mean(temperature_list)
else: temperature = 1.0
# 2. Use the temperature to record the ECE
# repeat the test N times changing the seed in range [seed, seed+repeat]
ece_list = list()
for i in range(seed, seed+repeat):
if(seed!=0): _set_seed(i)
else: _set_seed(0)
logits, targets = get_logits_targets(parse_args('test'))
#ece = ece_module.forward(logits, targets, temperature, onevsrest=params.method=='DKT').item()
ece = ece_module.forward(logits, targets, temperature, onevsrest=False).item()
ece_list.append(ece)
print("ECE:", np.mean(ece_list), "+-", np.std(ece_list))
# 3. Print the final ECE (averaged over all seeds)
print("-----------------------------")
print('Seeds = %d | Overall ECE = %4.4f +- %4.4f' %(repeat, np.mean(ece_list), np.std(ece_list)))
print("-----------------------------")
if __name__ == '__main__':
main()