-
Notifications
You must be signed in to change notification settings - Fork 3
/
eval_classify_mnist.py
261 lines (220 loc) · 9.45 KB
/
eval_classify_mnist.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
from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.optim.lr_scheduler import StepLR
from classify_mnist import Net
import numpy as np
import ipdb
import itertools
import matplotlib.pylab as plt
from scipy.stats import entropy
from tqdm import tqdm
from utils import tonp
def test(args, model, device, test_loader, tr=None):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
if tr is not None:
data = torch.stack([tr(x) for x in data])
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
def unstack_mnist(x, y):
x1 = x.unsqueeze(1).permute(0,1,3,2)
y1 = torch.from_numpy(np.array([y%10, y//10%10, y//100%10]))
return x1, y1
def plot_digits(x, y, fpath):
x = x[:25].detach().cpu().numpy()
y = y[:25].detach().cpu().numpy()
fig, axs = plt.subplots(5,5,figsize=(10,7))
for n, ax in enumerate(itertools.chain(*axs)):
ax.imshow(x[n][0])
ax.set_title(str(y[n]))
plt.tight_layout()
plt.savefig(fpath, bbox_inches='tight', pad_inches=0)
def compute_is_from_preds(preds, splits):
# Now compute the mean kl-div
N = len(preds)
split_scores = []
for k in range(splits):
part = preds[k * (N // splits): (k+1) * (N // splits), :]
py = np.mean(part, axis=0)
scores = []
for i in range(part.shape[0]):
pyx = part[i, :]
scores.append(entropy(pyx, py))
split_scores.append(np.exp(np.mean(scores)))
return np.mean(split_scores), np.std(split_scores)
def expand_pyx(ipyx):
"""
:inputs:
ipyx - (B,3,10)
:outputs:
pyx - (B,1000)
"""
pyx = np.zeros((len(ipyx), 1000))
# loop below is fast enough: 73328.02it/s
for n, _ipyx in tqdm(enumerate(ipyx)):
a,b,c = _ipyx
tmp = np.outer(a,b).ravel()
pyx[n] = np.outer(tmp,c).ravel()
assert np.abs(pyx.sum(-1).mean()-1) < 1e-5
return pyx
def compute_stackedmnist_stats(ipyx):
"""
:inputs:
ipyx - p(y|x), but represented as shape (B, 3, 10)
p(y|x) is actually pyx.prod(-1) where y \in [0,999]
:outputs:
n_modes - number of modes covered, int \in [1,1e3]
modes - e.g. [0,1,100,300]
brier - the Brier score, \in roughly [0,1]
IS - Inception Score, in this case in \in [1,1e3]
"""
ipyx = tonp(ipyx)
pyx = expand_pyx(ipyx)
predicted_modes = np.argmax(pyx,-1)
modes = list(set(predicted_modes))
n_modes = len(modes)
# mean Brier
predicted_onehot = np.eye(1000)[predicted_modes]
brier = ((1./len(pyx)) * # (1./pyx.shape[1]) *
np.sqrt(np.power(pyx-predicted_onehot,2).sum(-1)).sum())
IS = compute_is_from_preds(pyx, 1)[0]
return n_modes, modes, brier, IS
def test_stacked_mnist( model, device, test_loader, tr, no_label=True):
model.eval()
test_loss = 0
correct = 0
pyx = []
with torch.no_grad():
for data, target in test_loader:
data, target = zip(*[unstack_mnist(tr(x), y) for x, y in zip(data, target)])
data, target = torch.stack(data).to(device), torch.stack(target).to(device)
nB, nC, one, W, H = data.shape
data = data.reshape(nB*nC, one, W, H)
target = target.reshape(nB*nC)
output = model(data)
pyx.append(torch.softmax(output, -1).reshape(nB, nC, 10))
if not no_label:
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
# plot_digits(data,target,'tmp.png')
if not no_label:
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
#
pyx = torch.cat(pyx, 0)
n_modes, modes, brier, IS = compute_stackedmnist_stats(pyx)
print(n_modes,brier, IS)
return n_modes, brier, IS
def make_loader(x,y=None):
if y is None:
y = torch.zeros(len(x)).long()
dset = torch.utils.data.TensorDataset(x, y)
loader = torch.utils.data.DataLoader(dset, batch_size=1000,
shuffle=False, num_workers=1,
drop_last=False)
return loader
class EvalStackedMNIST(object):
"""docstring for EvalStackedMNIST"""
def __init__(self, device):
super(EvalStackedMNIST, self).__init__()
self.device = device
self.pretrained_classifier = Net().to(device)
sd = torch.load("mnist_cnn.pt")
self.pretrained_classifier.load_state_dict(sd)
self.tr = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(28),
transforms.ToTensor(),
transforms.Normalize((-.5,), (2,)), # undo previous normalize
transforms.Normalize((0.1307,), (0.3081,))
])
def eval_samples(self, samples):
loader = make_loader(samples)
# with torch.no_grad():
return test_stacked_mnist( self.pretrained_classifier, self.device, loader, self.tr)
def main():
# from train import check_dataset
import data
import ipdb
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=False,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = True
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('/scratch/gobi1/wangkuan/data/data/MNIST/MNIST', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('/scratch/gobi1/wangkuan/data/data/MNIST/MNIST', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
model = Net().to(device)
sd = torch.load("mnist_cnn.pt")
model.load_state_dict(sd)
test(args, model, device, train_loader)
# test(args, model, device, test_loader)
# # MNIST
# dat = data.load_data('mnist','/scratch/gobi1/wangkuan/data/data/MNIST/MNIST' ,
# device=device, imgsize=28, Ntrain=60000, Ntest=10000)
# test_dataset = torch.utils.data.TensorDataset(dat['X_test'],dat['Y_test'])
# test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=1000,
# shuffle=False, num_workers=1,
# drop_last=False)
# tr = transforms.Compose([
# transforms.ToPILImage(),
# transforms.Resize(28),
# transforms.ToTensor(),
# transforms.Normalize((-.5,), (2,)), # undo previous normalize
# transforms.Normalize((0.1307,), (0.3081,))
# ])
# test(args, model, device, test_loader, tr)
# StackedMNIST
dat = data.load_data('stackedmnist','data' ,
device=device, imgsize=64, Ntrain=60000, Ntest=10000)
eval_runner = EvalStackedMNIST(device)
print(eval_runner.eval_samples(dat['X_test']))
# test_loader = make_loader(dat['X_test'],dat['Y_test'])
# train_loader = make_loader(dat['X_train'],dat['Y_train'])
# r_loader = make_loader(dat['X_test'])
# test_stacked_mnist( model, device, test_loader, tr)
# test_stacked_mnist( model, device, r_loader, tr)
# # print("Training set")
# # test_stacked_mnist( model, device, train_loader, tr)
if __name__ == '__main__':
main()