transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
valid_set = Subset(
datasets.CIFAR10(root='./cifardata',
train=True,
transform=test_transf,
download=True), idx[:10000])
valid_loader = DataLoader(valid_set, batch_size=batch_size, shuffle=True)
m_t_losses = []
m_t_accs = []
m_v_losses = []
m_v_accs = []
for ii in range(repeats):
model = SiCNN_3(f_in, size, ratio, nratio, srange)
model.to(device)
train_loss = []
train_acc = []
valid_loss = []
valid_acc = []
for epoch in range(1, nb_epochs + 1):
train_l, train_a = train(model, train_loader, learning_rate,
criterion, epoch, batch_log, device)
train_l, train_a = test(model, train_loader, criterion, epoch,
batch_log, device)
valid_l, valid_a = test(model, valid_loader, criterion, epoch,
batch_log, device)
train_loss.append(train_l)
train_acc.append(train_a)
test_set = datasets.MNIST(root=root,
train=False,
transform=uniform,
download=True)
test_loader = DataLoader(dataset=test_set,
batch_size=batch_size,
shuffle=False,
num_workers=1,
pin_memory=True)
s_test_losses = []
s_test_accs = []
for ii in range(repeats):
model = SiCNN_3(f_in, size, ratio, nratio, srange, padding)
model.to(device)
for epoch in range(1, nb_epochs + 1):
train_l, train_a = train(model, train_loader, learning_rate,
criterion, epoch, batch_log, device)
train_l, train_a = test(model, train_loader, criterion, epoch,
batch_log, device)
test_l, test_a = test(model, test_loader, criterion, epoch, batch_log,
device)
s_test_losses.append(test_l)
s_test_accs.append(test_a)
pickle.dump(model, log)
}
log = open("cifar_gaussian_log_results.pickle", "wb")
scales = [0.40, 0.52, 0.64, 0.76, 0.88, 1.0, 1.12, 1.24, 1.36, 1.48, 1.60]
root = './cifardata'
if not os.path.exists(root):
os.mkdir(root)
criterion = nn.CrossEntropyLoss()
models = [
#kanazawa(f_in, ratio, nratio, srange=0),
#kanazawa(f_in, ratio, nratio, srange),
SiCNN_3(f_in, size, ratio, nratio, srange=0),
SiCNN_3(f_in, size, ratio, nratio, srange),
SiCNN_3(f_in, filter_size(size, ratio, nratio), 1 / ratio, nratio, srange),
SiCNN_3big(f_in, size, ratio, nratio, srange=0),
SiCNN_3(f_in, size, ratio=2**(1 / 3), nratio=6, srange=0),
SiCNN_3(f_in, size, ratio=2**(1 / 3), nratio=6, srange=srange),
SiCNN_3(f_in,
filter_size(size, 2**(1 / 3), 6),
ratio=2**(-1 / 3),
nratio=6,
srange=srange)
]
for m in range(len(models)):
locals()['test_losses_{0}'.format(m)] = []
locals()['test_accs_{0}'.format(m)] = []
shuffle=True,
num_workers=1,
pin_memory=True)
test_set = datasets.MNIST(root=root,
train=False,
transform=uniform,
download=True)
test_loader = DataLoader(dataset=test_set,
batch_size=batch_size,
shuffle=False)
criterion = nn.CrossEntropyLoss()
models = [
SiCNN_3(f_in=1, size=3, ratio=2**(2 / 3), nratio=3, srange=2),
SiCNN_3(1, filter_size(3, 2**(2 / 3), 3), 2**(-2 / 3), 3, 2),
SiCNN_3(1, 5, 2**(2 / 3), 3, 2),
SiCNN_3(1, filter_size(5, 2**(2 / 3), 3), 2**(-2 / 3), 3, 2, padding=1),
SiCNN_3(1, 5, 2**(1 / 3), 6, 2),
SiCNN_3(1, filter_size(5, 2**(1 / 3), 6), 2**(-1 / 3), 6, 2, padding=1)
]
pickle.dump(len(models), log)
for ii, model in enumerate(models):
print("model {}".format(ii))
model.to(device)
train_loss = []
train_acc = []
dataiter = iter(test_loader)
images, labels = dataiter.next()
print(' '.join('%5s' % classes[labels[j]] for j in range(4)))
img = torchvision.utils.make_grid(images)
img = img / 2 + 0.5 # unnormalize
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
"""
criterion = nn.CrossEntropyLoss()
models = [
SiCNN_3(f_in, size, ratio, nratio, srange, padding, nb_classes),
Model(f_in, size, ratio, nratio, srange, padding, nb_classes)
]
pickle.dump(len(models), log)
for model in models:
model.to(device)
train_loss=[]
train_acc = []
valid_loss = []
valid_acc = []
for epoch in range(1, nb_epochs + 1):
train_l, train_a = train(model, train_loader, learning_rate, criterion, epoch, batch_log, device)