def transformation_cifar10_vs_tinyimagenet():
_, (x_test, y_test) = load_cifar10()
x_test_out = load_tinyimagenet('/home/izikgo/Imagenet_resize/Imagenet_resize/')
transformer = Transformer(8, 8)
n = 16
k = 8
base_mdl = create_wide_residual_network(x_test.shape[1:], 10, n, k)
transformations_cls_out = Activation('softmax')(dense(transformer.n_transforms)(base_mdl.get_layer(index=-3).output))
mdl = Model(base_mdl.input, [base_mdl.output, transformations_cls_out])
mdl.load_weights('cifar10_WRN_doublehead-transformations_{}-{}.h5'.format(n, k))
scores_mdl = Model(mdl.input, mdl.output[1])
x_test_all = np.concatenate((x_test, x_test_out))
preds = np.zeros((len(x_test_all), transformer.n_transforms))
for t in range(transformer.n_transforms):
preds[:, t] = scores_mdl.predict(transformer.transform_batch(x_test_all, [t] * len(x_test_all)),
batch_size=128)[:, t]
labels = np.concatenate((np.ones(len(x_test)), np.zeros(len(x_test_out))))
scores = preds.mean(axis=-1)
save_roc_pr_curve_data(scores, labels, 'cifar10-vs-tinyimagenet_transformations.npz')
def osoc_test(hyper_para):
transformer = Transformer(8, 8)
C = wrn.WideResNet(28, num_classes=10, dropout_rate=0,
widen_factor=10).cuda()
C.load_state_dict(torch.load(hyper_para.experiment_name + '.pth'))
single_class_ind = hyper_para.inclass[0]
(x_train, y_train), (x_test, y_test) = load_cifar10()
C.eval()
score = []
lbl = []
features = []
correct = 0
total = 0
dset = CustomDataset(x_test, y_test)
testLoader = DataLoader(dset, batch_size=128)
for i, (inputs0, labels) in enumerate(testLoader):
inputs = inputs0.permute(0, 3, 2, 1)
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
features += f.detach().cpu().tolist()
val, ind = torch.max(act, dim=1)
#score += val.detach().cpu().tolist()
CC = wrn.WideResNet(28,
num_classes=72,
dropout_rate=0,
widen_factor=10).cuda()
val, ind, labels = (val.detach().cpu().tolist(),
ind.detach().cpu().tolist(),
labels.detach().cpu().tolist())
for ind, (ii, gt) in enumerate(zip(ind, labels)):
gt = gt[0]
CC.load_state_dict(
torch.load('saved/' + hyper_para.experiment_name + '_' +
str(gt) + '.pth'))
x_test0 = transformer.transform_batch(
np.expand_dims(inputs0[ind, :, :, :].detach().cpu().numpy(),
0), [0])
x_test0 = torch.tensor(x_test0).permute(0, 3, 2, 1).cuda()
act, f = CC(x_test0)
act = act.detach().cpu().tolist()
score += [act[0][0]]
if gt in hyper_para.inclass:
total += 1
if ii == gt:
correct += 1
lbl.append(1)
else:
lbl.append(0)
break
fpr, tpr, thresholds = metrics.roc_curve(lbl, score)
AUC = metrics.auc(fpr, tpr)
ACC = float(correct) / total
print('AUROC: ' + str(AUC) + '\t Accuracy: ' + str(ACC))
def __init__(self, rules, values):
wx.grid.PyGridTableBase.__init__(self)
# crear transformador
transformer = Transformer(rules)
# cargar datos y transformarlos
self.data = [[value, transformer.transform_value(value)] for value in values]
self.column_labels = [u"Valor Original", u"Valor Transformado"]
def _build(self):
self.params = {}
self.latent_values = None
# Build the transformer
beta_warp = BetaWarp(self.num_dims)
beta_alpha, beta_beta = beta_warp.hypers
self.params['beta_alpha'] = beta_alpha
self.params['beta_beta'] = beta_beta
transformer = Transformer(self.num_dims)
transformer.add_layer(beta_warp)
# Build the component kernels
input_kernel = Matern52(self.num_dims)
ls = input_kernel.hypers
self.params['ls'] = ls
# Now apply the transformation.
transform_kernel = TransformKernel(input_kernel, transformer)
# Add some perturbation for stability
stability_noise = Noise(self.num_dims)
# Finally make a noisy version if necessary
# In a classifier GP the notion of "noise" is really just the scale.
if self.noiseless:
self._kernel = SumKernel(transform_kernel, stability_noise)
else:
scaled_kernel = Scale(transform_kernel)
self._kernel = SumKernel(scaled_kernel, stability_noise)
amp2 = scaled_kernel.hypers
self.params['amp2'] = amp2
# Build the mean function (just a constant mean for now)
self.mean = Hyperparameter(
initial_value = 0.0,
prior = priors.Gaussian(0.0,1.0),
name = 'mean'
)
self.params['mean'] = self.mean
# Buld the latent values. Empty for now until the GP gets data.
self.latent_values = Hyperparameter(
initial_value = np.array([]),
name = 'latent values'
)
# Build the samplers
to_sample = [self.mean] if self.noiseless else [self.mean, amp2]
self._samplers.append(SliceSampler(*to_sample, compwise=False, thinning=self.thinning))
self._samplers.append(WhitenedPriorSliceSampler(ls, beta_alpha, beta_beta, compwise=True, thinning=self.thinning))
self.latent_values_sampler = EllipticalSliceSampler(self.latent_values, thinning=self.ess_thinning)
def _build(self):
# Build the transformer
beta_warp = BetaWarp(self.num_dims)
transformer = Transformer(self.num_dims)
transformer.add_layer(beta_warp)
# Build the component kernels
input_kernel = Matern52(self.num_dims)
stability_noise_kernel = Noise(self.num_dims) # Even if noiseless we use some noise for stability
scaled_input_kernel = Scale(input_kernel)
sum_kernel = SumKernel(scaled_input_kernel, stability_noise_kernel)
noise_kernel = Noise(self.num_dims)
# The final kernel applies the transformation.
self._kernel = TransformKernel(sum_kernel, transformer)
# Finally make a noisy version if necessary
if not self.noiseless:
self._kernel_with_noise = SumKernel(self._kernel, noise_kernel)
# Build the mean function (just a constant mean for now)
self.mean = Hyperparameter(
initial_value = 0.0,
prior = priors.Gaussian(0.0,1.0),
name = 'mean'
)
# Get the hyperparameters to sample
ls = input_kernel.hypers
amp2 = scaled_input_kernel.hypers
beta_alpha, beta_beta = beta_warp.hypers
self.params = {
'mean' : self.mean,
'amp2' : amp2,
'ls' : ls,
'beta_alpha' : beta_alpha,
'beta_beta' : beta_beta
}
# Build the samplers
if self.noiseless:
self._samplers.append(SliceSampler(self.mean, amp2, compwise=False, thinning=self.thinning))
else:
noise = noise_kernel.hypers
self.params.update({'noise' : noise})
self._samplers.append(SliceSampler(self.mean, amp2, noise, compwise=False, thinning=self.thinning))
self._samplers.append(SliceSampler(ls, beta_alpha, beta_beta, compwise=True, thinning=self.thinning))
def os_test_ens(testLoader, hyper_para, C, isauc):
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0,
widen_factor=10).cuda()
C.load_state_dict(
torch.load('saved/' + hyper_para.experiment_name + '_' +
str(hyper_para.inclass[0]) + '.pth'))
single_class_ind = hyper_para.inclass[0]
(x_train, y_train), (x_test, y_test) = load_cifar10()
transformer = Transformer(8, 8)
glabels = y_test.flatten() == single_class_ind
C.cuda().eval()
scores = np.array([[]])
features = []
correct = 0
total = 0
preds = np.zeros((len(x_test), transformer.n_transforms))
for t in [0]: #range(72):
score = []
x_test0 = transformer.transform_batch(x_test, [t] * len(x_test))
dset = CustomDataset(x_test0, [t] * len(x_test))
testLoader = DataLoader(dset, batch_size=128, shuffle=False)
for i, (inputs, labels) in enumerate(testLoader):
inputs = inputs.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
features += f.detach().cpu().tolist()
#act = torch.nn.functional.softmax(act, dim=1)
score += act[:, t].detach().cpu().tolist()
preds[:, t] = list(score)
fpr, tpr, thresholds = metrics.roc_curve(glabels, score)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
scores = np.sum(((preds)), 1)
fpr, tpr, thresholds = metrics.roc_curve(glabels, scores)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
return ([0, 0])
class RotateDataset(Dataset):
def __init__(self, x_tensor):
self.x = x_tensor
self.transformer = Transformer(8, 8)
def __getitem__(self, index):
trans_id = np.random.randint(self.transformer.n_transforms, size=1)[0]
#return(self.transformer.transform_sample(self.x[index], trans_id), trans_id)
return ((self.transformer.transform_batch(
np.expand_dims(self.x[index], 0), [trans_id]))[0], trans_id)
def __len__(self):
return len(self.x)
class RotateDataset(Dataset):
def __init__(self, x_tensor, lbl):
self.x = x_tensor
self.y = lbl
self.perm = np.random.permutation(len(x_tensor))
self.transformer = Transformer(8, 8)
def __getitem__(self, index):
index = self.perm[index]
trans_id = np.random.randint(self.transformer.n_transforms, size=1)[0]
#return(self.transformer.transform_sample(self.x[index], trans_id), trans_id)
return ((self.transformer.transform_batch(
np.expand_dims(self.x[index], 0),
[trans_id]))[0], trans_id, self.y[index])
def __len__(self):
return len(self.x)
def train_cifar10_transformations():
(x_train, y_train), _ = load_cifar10()
transformer = Transformer(8, 8)
def data_gen(x, y, batch_size):
while True:
ind_permutation = np.random.permutation(len(x))
for b_start_ind in range(0, len(x), batch_size):
batch_inds = ind_permutation[b_start_ind:b_start_ind + batch_size]
x_batch = x[batch_inds]
y_batch = y[batch_inds].flatten()
if K.image_data_format() == 'channels_first':
x_batch = np.transpose(x_batch, (0, 2, 3, 1))
y_t_batch = np.random.randint(0, transformer.n_transforms, size=len(x_batch))
x_batch = transformer.transform_batch(x_batch, y_t_batch)
if K.image_data_format() == 'channels_first':
x_batch = np.transpose(x_batch, (0, 3, 1, 2))
yield (x_batch, [to_categorical(y_batch, num_classes=10), to_categorical(y_t_batch, num_classes=transformer.n_transforms)])
n = 16
k = 8
base_mdl = create_wide_residual_network(x_train.shape[1:], 10, n, k)
transformations_cls_out = Activation('softmax')(dense(transformer.n_transforms)(base_mdl.get_layer(index=-3).output))
mdl = Model(base_mdl.input, [base_mdl.output, transformations_cls_out])
mdl.compile(SGDTorch(lr=.1, momentum=0.9, nesterov=True), 'categorical_crossentropy', ['acc'])
lr_cb = LearningRateScheduler(lambda e: 0.1 * (0.2 ** (e >= 160 and 3 or e >= 120 and 2 or e >= 60 and 1 or 0)))
batch_size = 128
mdl.fit_generator(
generator=data_gen(x_train, y_train, batch_size=batch_size),
steps_per_epoch=len(x_train) // batch_size,
epochs=200,
callbacks=[lr_cb]
)
mdl.save_weights('cifar10_WRN_doublehead-transformations_{}-{}.h5'.format(n, k))
def _transformations_experiment(dataset_load_fn, dataset_name,
single_class_ind, gpu_q):
# gpu_to_use = gpu_q.get()
# os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
if dataset_name in ['cats-vs-dogs']:
transformer = Transformer(16, 16)
n, k = (16, 8)
else:
transformer = Transformer(8, 8)
n, k = (10, 4)
mdl = create_wide_residual_network(x_train.shape[1:],
transformer.n_transforms, n, k)
mdl.compile('adam', 'categorical_crossentropy', ['acc'])
# get inliers of specific class
x_train_task = x_train[y_train.flatten() == single_class_ind]
# [0_i, ..., (N_transforms-1)_i, ..., ..., 0_N_samples, ...,
# (N_transforms-1)_N_samples] shape: (N_transforms*N_samples,)
transformations_inds = np.tile(np.arange(transformer.n_transforms),
len(x_train_task))
x_train_task_transformed = transformer.transform_batch(
np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
batch_size = 128
mdl.fit(x=x_train_task_transformed,
y=to_categorical(transformations_inds),
batch_size=batch_size,
epochs=int(np.ceil(200 / transformer.n_transforms)))
scores = np.zeros((len(x_test), ))
matrix_evals = np.zeros(
(len(x_test), transformer.n_transforms, transformer.n_transforms))
observed_data = x_train_task
for t_ind in range(transformer.n_transforms):
observed_dirichlet = mdl.predict(transformer.transform_batch(
observed_data, [t_ind] * len(observed_data)),
batch_size=1024)
log_p_hat_train = np.log(observed_dirichlet).mean(axis=0)
alpha_sum_approx = calc_approx_alpha_sum(observed_dirichlet)
alpha_0 = observed_dirichlet.mean(axis=0) * alpha_sum_approx
mle_alpha_t = fixed_point_dirichlet_mle(alpha_0, log_p_hat_train)
x_test_p = mdl.predict(transformer.transform_batch(
x_test, [t_ind] * len(x_test)),
batch_size=1024)
matrix_evals[:, :, t_ind] += x_test_p
scores += dirichlet_normality_score(mle_alpha_t, x_test_p)
scores /= transformer.n_transforms
matrix_evals /= transformer.n_transforms
scores_simple = np.trace(matrix_evals, axis1=1, axis2=2)
scores_entropy = get_entropy(matrix_evals)
scores_xH = get_xH(transformer, matrix_evals)
labels = y_test.flatten() == single_class_ind
save_results_file(dataset_name,
single_class_ind,
scores=scores,
labels=labels,
experiment_name='transformations')
save_results_file(dataset_name,
single_class_ind,
scores=scores_simple,
labels=labels,
experiment_name='transformations-simple')
save_results_file(dataset_name,
single_class_ind,
scores=scores_entropy,
labels=labels,
experiment_name='transformations-entropy')
save_results_file(dataset_name,
single_class_ind,
scores=scores_xH,
labels=labels,
experiment_name='transformations-xH')
mdl_weights_name = '{}_transformations_{}_{}_weights.h5'.format(
dataset_name, get_class_name_from_index(single_class_ind,
dataset_name),
datetime.datetime.now().strftime('%Y-%m-%d-%H%M'))
mdl_weights_path = os.path.join(RESULTS_DIR, dataset_name,
mdl_weights_name)
mdl.save_weights(mdl_weights_path)
def __init__(self, x_tensor, lbl):
self.x = x_tensor
self.y = lbl
self.perm = np.random.permutation(len(x_tensor))
self.transformer = Transformer(8, 8)
def octest0(hyper_para):
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0,
widen_factor=10).cuda()
C.load_state_dict(
torch.load(hyper_para.experiment_name + '_' +
str(hyper_para.inclass[0]) + '.pth'))
single_class_ind = hyper_para.inclass[0]
if hyper_para.source == 'mnist':
(x_train, y_train), (x_test, y_test) = load_mnist()
elif hyper_para.source == 'svhn':
(x_train, y_train), (x_test, y_test) = load_svhn()
elif hyper_para.source == 'amazon':
(x_train, y_train), (x_test, y_test) = load_amazon()
elif hyper_para.source == 'dslr':
(x_train, y_train), (x_test, y_test) = load_dslr()
transformer = Transformer(8, 8)
glabels = y_test.flatten() == single_class_ind
print(hyper_para.source)
print(len(x_test))
print(np.sum(glabels))
C.cuda().eval()
scores = np.array([[]])
features = []
correct = 0
total = 0
preds = np.zeros((len(x_test), transformer.n_transforms))
for t in range(1):
score = []
dset = CustomDataset(x_test, y_test)
testLoader = DataLoader(dset, batch_size=128, shuffle=False)
transformer = Transformer(8, 8)
for i, (inputs, labels) in enumerate(testLoader):
inputs = torch.tensor(
transformer.transform_batch(inputs.detach().cpu().numpy(),
[t] * len(inputs)))
inputs = inputs.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
features += f.detach().cpu().tolist()
#act = torch.nn.functional.softmax(act, dim=1)
score += act[:, t].detach().cpu().tolist()
preds[:, t] = list(score)
fpr, tpr, thresholds = metrics.roc_curve(glabels, score)
AUC = metrics.auc(fpr, tpr)
#print('AUROC: ' + str(AUC) )
scores = np.sum(((preds)), 1)
fpr, tpr, thresholds = metrics.roc_curve(glabels, scores)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
file1 = open("AUC.txt", "a") #append mode
file1.write(hyper_para.source + "\t" + hyper_para.method + "\t" +
str(hyper_para.inclass[0]) + "\t" +
hyper_para.experiment_name + "\t" + str(AUC) + "\n")
file1.close()
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0,
widen_factor=10).cuda()
C.load_state_dict(
torch.load(hyper_para.experiment_name + '_' +
str(hyper_para.inclass[0]) + '.pth'))
single_class_ind = hyper_para.inclass[0]
if hyper_para.target == 'mnist':
(x_train, y_train), (x_test, y_test) = load_mnist()
elif hyper_para.target == 'svhn':
(x_train, y_train), (x_test, y_test) = load_svhn()
elif hyper_para.target == 'amazon':
(x_train, y_train), (sx_test, y_test) = load_amazon()
elif hyper_para.target == 'dslr':
(x_train, y_train), (x_test, y_test) = load_dslr()
transformer = Transformer(8, 8)
C.cuda().eval()
scores = np.array([[]])
features = []
correct = 0
total = 0
preds = np.zeros((len(x_test) + len(x_train), transformer.n_transforms))
for t in range(1):
score = []
dset = CustomDataset(np.concatenate((x_test, x_train), 0),
np.concatenate((y_test, y_train), 0))
testLoader = DataLoader(dset, batch_size=128, shuffle=False)
transformer = Transformer(8, 8)
for i, (inputs, labels) in enumerate(testLoader):
inputs = torch.tensor(
transformer.transform_batch(inputs.detach().cpu().numpy(),
[t] * len(inputs)))
inputs = inputs.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
features += f.detach().cpu().tolist()
#act = torch.nn.functional.softmax(act, dim=1)
score += act[:, t].detach().cpu().tolist()
preds[:, t] = list(score)
#print('AUROC: ' + str(AUC) )
glabels = (np.concatenate((y_test, y_train))).flatten() == single_class_ind
scores = np.sum(((preds)), 1)
fpr, tpr, thresholds = metrics.roc_curve(glabels, scores)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
file1 = open("AUC.txt", "a") #append mode
file1.write(hyper_para.target + "\t" + hyper_para.method + "\t" +
str(hyper_para.inclass[0]) + "\t" +
hyper_para.experiment_name + "\t" + str(AUC) + "\n")
file1.close()
def oceval(hyper_para):
import sklearn
for single_class_ind in hyper_para.inclass:
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0, widen_factor=10)
C.load_state_dict(
torch.load(hyper_para.experiment_name + '_' +
str(hyper_para.inclass[0]) + '.pth'))
C.cuda()
if hyper_para.source == 'mnist':
(x_train, y_train), (x_test, y_test) = load_mnist()
(x_train2, y_train2), (x_test2, y_test2) = load_svhn()
elif hyper_para.source == 'svhn':
(x_train, y_train), (x_test, y_test) = load_svhn()
(x_train2, y_train2), (x_test2, y_test2) = load_mnist()
elif hyper_para.source == 'amazon':
(x_train, y_train), (x_test, y_test) = load_amazon()
(x_train2, y_train2), (x_test2, y_test2) = load_dslr()
elif hyper_para.source == 'dslr':
(x_train, y_train), (x_test, y_test) = load_dslr()
(x_train2, y_train2), (x_test2, y_test2) = load_amazon()
if hyper_para.method == 'justsource':
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
domain_lbl = [0] * len(x_train_task)
tst_lbl = [0] * len(x_test)
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=512, shuffle=True)
else: #if hyper_para.method == 'balancedsourcetarget':
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_train_task2 = x_train2[y_train2.flatten() == single_class_ind]
x_train_task2 = x_train_task2[0:hyper_para.target_n]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
x_test2 = x_train_task2[int(len(x_train_task2) * 0.8):]
x_train_task2 = x_train_task2[0:int(len(x_train_task2) * 0.8)]
domain_lbl = [0] * len(x_train_task) + [1] * (len(
x_train_task2) * int(len(x_train_task) / len(x_train_task2)))
x_train_task2 = np.tile(
x_train_task2,
(int(len(x_train_task) / len(x_train_task2)), 1, 1, 1))
x_train_task = np.concatenate((x_train_task, x_train_task2))
tst_lbl = [0] * len(x_test) + [1] * (
len(x_test2) * int(len(x_test) / len(x_test2)))
x_test2 = np.tile(x_test2,
(int(len(x_test) / len(x_test2)), 1, 1, 1))
x_test = np.concatenate((x_test, x_test2))
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=512, shuffle=True)
transformer = Transformer(8, 8)
dset = RotateDataset(x_train_task, domain_lbl)
trainLoader = DataLoader(dset, batch_size=512, shuffle=True)
np.set_printoptions(threshold=sys.maxsize)
correct = 0
total = 0
preds = []
target = []
for t in range(10):
for i, (inputs, labels, dlbls) in enumerate(testLoader):
inputs = inputs.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
_, ind = torch.max(act, 1)
target += labels.detach().cpu().tolist()
preds += ind.detach().cpu().tolist()
import matplotlib
import matplotlib.pyplot as plt
cm = sklearn.metrics.confusion_matrix(target, preds, normalize='true')
#plt.imshow(cm)
#plt.show()
cm = np.diag(cm)
cm = cm / np.sum(cm) * 72
return cm
def octrain(hyper_para):
for single_class_ind in hyper_para.inclass:
hyper_para.C = wrn.WideResNet(28,
num_classes=72,
dropout_rate=0,
widen_factor=10)
if hyper_para.source == 'mnist':
(x_train, y_train), (x_test, y_test) = load_mnist()
(x_train2, y_train2), (x_test2, y_test2) = load_svhn()
elif hyper_para.source == 'svhn':
(x_train, y_train), (x_test, y_test) = load_svhn()
(x_train2, y_train2), (x_test2, y_test2) = load_mnist()
elif hyper_para.source == 'amazon':
(x_train, y_train), (x_test, y_test) = load_amazon()
(x_train2, y_train2), (x_test2, y_test2) = load_dslr()
elif hyper_para.source == 'dslr':
(x_train, y_train), (x_test, y_test) = load_dslr()
(x_train2, y_train2), (x_test2, y_test2) = load_amazon()
if hyper_para.method == 'justsource':
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
domain_lbl = [0] * len(x_train_task)
tst_lbl = [0] * len(x_test)
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=16, shuffle=True)
else:
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_train_task2 = x_train2[y_train2.flatten() == single_class_ind]
x_train_task2 = x_train_task2[0:hyper_para.target_n]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
x_test2 = x_train_task2[int(len(x_train_task2) * 0.8):]
x_train_task2 = x_train_task2[0:int(len(x_train_task2) * 0.8)]
domain_lbl = [0] * len(x_train_task)
x_train_task2 = np.tile(
x_train_task2,
(int(len(x_train_task) / len(x_train_task2)), 1, 1, 1))
tst_lbl = [0] * len(x_test) + [1] * (
len(x_test2) * int(len(x_test) / len(x_test2)))
x_test2 = np.tile(x_test2,
(int(len(x_test) / len(x_test2)), 1, 1, 1))
x_test = np.concatenate((x_test, x_test2))
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=16, shuffle=True)
transformer = Transformer(8, 8)
dset = RotateDataset(x_train_task, domain_lbl)
trainLoader = DataLoader(dset, batch_size=512, shuffle=True)
'''transformations_inds = np.tile(np.arange(transformer.n_transforms), len(x_train_task))
x_train_task_transformed = transformer.transform_batch(np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
dset = CustomDataset(x_train_task_transformed,transformations_inds )
trainLoader = DataLoader(dset, batch_size=32, shuffle=True)'''
# define networks
C = hyper_para.C
C = torch.nn.DataParallel(C)
# define loss functions
ce_criterion = nn.CrossEntropyLoss()
# define optimizer
optimizer_c = optim.Adam(C.parameters(),
lr=hyper_para.lr,
betas=(0.9, 0.999))
# turn on the train mode
C.train(mode=True)
# initialization of auxilary variables
running_tl = 0.0
running_cc = 0.0
running_rc = 0.0
running_ri = 0.0
# if gpu use cuda
best_acc = 0
for i in range(100): #int(100*hyper_para.source_n/len(x_train_task))):
acct = 0
nelt = 0
acc0t = 0
nel0t = 0
acc1t = 0
nel1t = 0
if hyper_para.gpu:
C.cuda()
for idx, (inputs, labels, dlbls) in enumerate(trainLoader):
inputs = inputs.permute(0, 3, 2, 1)
t1 = time.time()
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels.cuda()
dlbls = dlbls.cuda()
act, fea = C(inputs)
_, ind = torch.max(act, 1)
if True:
loss_cc = ce_criterion(act, labels)
optimizer_c.zero_grad()
loss = loss_cc
loss.backward()
optimizer_c.step()
running_cc += loss_cc.data
t2 = time.time()
torch.save(
C.module.state_dict(),
hyper_para.experiment_name + '_' + str(single_class_ind) + '.pth')
w = oceval(hyper_para)
# finetune
if hyper_para.source == 'mnist':
(x_train, y_train), (x_test, y_test) = load_mnist()
(x_train2, y_train2), (x_test2, y_test2) = load_svhn()
elif hyper_para.source == 'svhn':
(x_train, y_train), (x_test, y_test) = load_svhn()
(x_train2, y_train2), (x_test2, y_test2) = load_mnist()
elif hyper_para.source == 'amazon':
(x_train, y_train), (x_test, y_test) = load_amazon()
(x_train2, y_train2), (x_test2, y_test2) = load_dslr()
elif hyper_para.source == 'dslr':
(x_train, y_train), (x_test, y_test) = load_dslr()
(x_train2, y_train2), (x_test2, y_test2) = load_amazon()
if hyper_para.method == 'justsource':
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
domain_lbl = [0] * len(x_train_task)
tst_lbl = [0] * len(x_test)
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=16, shuffle=True)
else: #if hyper_para.method == 'balancedsourcetarget':
x_train_task = x_train[y_train.flatten() == single_class_ind]
x_train_task2 = x_train2[y_train2.flatten() == single_class_ind]
x_train_task2 = x_train_task2[0:hyper_para.target_n]
x_test = x_train_task[int(len(x_train_task) * 0.8):]
x_train_task = x_train_task[0:int(len(x_train_task) * 0.8)]
x_test2 = x_train_task2[int(len(x_train_task2) * 0.8):]
x_train_task2 = x_train_task2[0:int(len(x_train_task2) * 0.8)]
domain_lbl = [0] * len(x_train_task) + [1] * (len(
x_train_task2) * int(len(x_train_task) / len(x_train_task2)))
x_train_task2 = np.tile(
x_train_task2,
(int(len(x_train_task) / len(x_train_task2)), 1, 1, 1))
x_train_task = np.concatenate((x_train_task, x_train_task2))
tst_lbl = [0] * len(x_test) + [1] * (
len(x_test2) * int(len(x_test) / len(x_test2)))
x_test2 = np.tile(x_test2,
(int(len(x_test) / len(x_test2)), 1, 1, 1))
x_test = np.concatenate((x_test, x_test2))
dset2 = RotateDataset(x_test, tst_lbl)
testLoader = DataLoader(dset2, batch_size=16, shuffle=True)
transformer = Transformer(8, 8)
dset = RotateDataset(x_train_task, domain_lbl)
trainLoader = DataLoader(dset, batch_size=512, shuffle=True)
'''transformations_inds = np.tile(np.arange(transformer.n_transforms), len(x_train_task))
x_train_task_transformed = transformer.transform_batch(np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
dset = CustomDataset(x_train_task_transformed,transformations_inds )
trainLoader = DataLoader(dset, batch_size=32, shuffle=True)'''
if hyper_para.method == 'dd':
Dnet = FCDiscriminator().cuda()
optimizer_D = optim.Adam(Dnet.parameters(),
lr=hyper_para.lr,
betas=(0.9, 0.999))
Dnet.train()
# define networks
C = hyper_para.C
C = torch.nn.DataParallel(C)
# define loss functions
print(w)
ce_criterion = nn.CrossEntropyLoss(weight=torch.Tensor(w).cuda())
# define optimizer
optimizer_c = optim.Adam(C.parameters(),
lr=hyper_para.lr,
betas=(0.9, 0.999))
# turn on the train mode
C.train(mode=True)
# initialization of auxilary variables
running_tl = 0.0
running_cc = 0.0
running_rc = 0.0
running_ri = 0.0
# if gpu use cuda
best_acc = 0
for i in range(500): #int(100*hyper_para.source_n/len(x_train_task))):
acct = 0
nelt = 0
acc0t = 0
nel0t = 0
acc1t = 0
nel1t = 0
if hyper_para.gpu:
C.cuda()
for idx, (inputs, labels, dlbls) in enumerate(trainLoader):
inputs = inputs.permute(0, 3, 2, 1)
t1 = time.time()
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels.cuda()
dlbls = dlbls.cuda()
act, fea = C(inputs)
_, ind = torch.max(act, 1)
if hyper_para.method == 'dd':
# Train Discriminator
Dnet.train()
d = Dnet(fea)
_, mind = torch.max(d, 1)
acc = torch.mean(torch.eq(mind, dlbls).cpu().float())
optimizer_D.zero_grad()
optimizer_c.zero_grad()
loss_d = nnfunc.binary_cross_entropy_with_logits(
d.squeeze(), dlbls.float())
loss_cc = ce_criterion(act, labels)
loss = loss_cc + loss_d
loss.backward()
optimizer_c.step()
optimizer_D.step()
else:
loss_cc = ce_criterion(act, labels)
optimizer_c.zero_grad()
loss = loss_cc
loss.backward()
optimizer_c.step()
running_cc += loss_cc.data
t2 = time.time()
acct += torch.sum(torch.eq(ind, labels)).cpu().numpy()
nelt += ind.shape[0]
acc0t += torch.sum(
torch.eq(ind[dlbls == 0],
labels[dlbls == 0])).cpu().numpy()
nel0t += (dlbls == 0).shape[0]
acc1t += torch.sum(
torch.eq(ind[dlbls == 1],
labels[dlbls == 1])).cpu().numpy()
nel1t += (dlbls == 1).shape[0]
if idx % 10 == 0:
line = hyper_para.BLUE + '[' + str(format(i + 1, '8d')) + '/' + str(
format(int(hyper_para.iterations), '8d')) + ']' + hyper_para.ENDC + \
hyper_para.GREEN + ' loss_cc: ' + hyper_para.ENDC + str(
format(running_cc / hyper_para.stats_frequency, '1.8f')) + \
hyper_para.YELLOW + ' time (min): ' + hyper_para.ENDC + str(int((t2 - t1) * 20.0))
print(line)
acc = 0
nel = 0
acc0 = 0
nel0 = 0
acc1 = 0
nel1 = 0
for idx, (inputs, labels, dlbls) in enumerate(testLoader):
inputs = inputs.permute(0, 3, 2, 1)
t1 = time.time()
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels.cuda()
dlbls = dlbls.cuda()
act, _ = C(inputs)
_, ind = torch.max(act, 1)
acc += torch.sum(torch.eq(ind, labels)).cpu().numpy()
nel += ind.shape[0]
acc0 += torch.sum(
torch.eq(ind[dlbls == 0],
labels[dlbls == 0])).cpu().numpy()
nel0 += (dlbls == 0).shape[0]
acc1 += torch.sum(
torch.eq(ind[dlbls == 1],
labels[dlbls == 1])).cpu().numpy()
nel1 += (dlbls == 1).shape[0]
print(['Val', acc / nel, acc0 / nel0, acc1 / nel1, nel0, nel1])
if acc / nel >= best_acc:
torch.save(
C.module.state_dict(), hyper_para.experiment_name +
'_' + str(single_class_ind) + '.pth')
best_acc = acc / nel
running_cc = 0.0
print([
'Train', acct / (nel0t + nel1t), acc0t / nel0t, acc1t / nel1t,
nel0t, nel1t
])
def _transformations_experiment(dataset_load_fn, dataset_name,
single_class_ind, gpu_q):
gpu_to_use = gpu_q.get()
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
if dataset_name in ['cats-vs-dogs']:
transformer = Transformer(16, 16)
n, k = (16, 8)
else:
transformer = Transformer(8, 8)
n, k = (10, 4)
#mdl = create_wide_residual_network(x_train.shape[1:], transformer.n_transforms, n, k)
'''mdl.compile('adam',
'categorical_crossentropy',
['acc'])'''
x_train_task = x_train[y_train.flatten() == single_class_ind]
#x_train_task = x_train_task[0:15][:][:][:]
transformations_inds = np.tile(np.arange(transformer.n_transforms),
len(x_train_task))
x_train_task_transformed = transformer.transform_batch(
np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
batch_size = 32
'''mdl.fit(x=x_train_task_transformed, y=to_categorical(transformations_inds),
batch_size=batch_size, epochs=int(np.ceil(200/transformer.n_transforms)))
mdl_weights_name = '{}_transformations_{}_weights.h5'.format(dataset_name,
get_class_name_from_index(single_class_ind, dataset_name))
#mdl.load_weights('cifar10/'+mdl_weights_name)
#################################################################################################
# simplified normality score
#################################################################################################
preds = np.zeros((len(x_test), transformer.n_transforms))
for t in [1]:#range(1):
preds[:, t] = mdl.predict(transformer.transform_batch(x_test, [t] * len(x_test)),
batch_size=batch_size)[:, t]
labels = y_test.flatten() == single_class_ind
scores = preds.mean(axis=-1)
#print("Accuracy : "+ str(scores))
#################################################################################################'''
dset = CustomDataset(x_train_task_transformed, transformations_inds)
trainLoader = DataLoader(dset, batch_size=32, shuffle=True)
ce_criterion = nn.CrossEntropyLoss()
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0,
widen_factor=10).cuda()
C = nn.DataParallel(C)
optimizer_c = optim.Adam(C.parameters(), lr=0.001, betas=(0.9, 0.999))
C.train(mode=True)
'''for epoch in range(int(1)):
for (inputs, labels) in trainLoader:
inputs = inputs.permute(0,3,2,1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act,_ = C(inputs)
loss_cc = ce_criterion(act, labels)
optimizer_c.zero_grad()
loss_cc.backward()
optimizer_c.step()
#running_cc += loss_cc.data
running_cc = 0.0
torch.save(C.cpu().state_dict(), 'out.pth')'''
C = C.cpu()
C.load_state_dict(torch.load('out.pth'))
(x_train, y_train), (x_test, y_test) = load_cifar10()
transformer = Transformer(8, 8)
glabels = y_test.flatten() == single_class_ind
C.cuda().eval()
scores = np.array([[]])
features = []
correct = 0
total = 0
preds = np.zeros((len(x_test), transformer.n_transforms))
for t in [6, 6]: #range(72):
score = []
x_test0 = transformer.transform_batch(x_test, [t] * len(x_test))
dset = CustomDataset(x_test0, [t] * len(x_test))
testLoader = DataLoader(dset, batch_size=128, shuffle=False)
for i, (inputs, labels) in enumerate(testLoader):
inputs = inputs.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act, f = C(inputs)
features += f.detach().cpu().tolist()
#act = torch.nn.functional.softmax(act, dim=1)
score += act[:, t].detach().cpu().tolist()
preds[:, t] = list(score)
fpr, tpr, thresholds = metrics.roc_curve(glabels, score)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
scores = np.sum(((preds)), 1)
fpr, tpr, thresholds = metrics.roc_curve(glabels, scores)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
def calc_approx_alpha_sum(observations):
N = len(observations)
f = np.mean(observations, axis=0)
return (N * (len(f) - 1) *
(-psi(1))) / (N * np.sum(f * np.log(f)) -
np.sum(f * np.sum(np.log(observations), axis=0)))
def inv_psi(y, iters=5):
# initial estimate
cond = y >= -2.22
x = cond * (np.exp(y) + 0.5) + (1 - cond) * -1 / (y - psi(1))
for _ in range(iters):
x = x - (psi(x) - y) / polygamma(1, x)
return x
def fixed_point_dirichlet_mle(alpha_init, log_p_hat, max_iter=1000):
alpha_new = alpha_old = alpha_init
for _ in range(max_iter):
alpha_new = inv_psi(psi(np.sum(alpha_old)) + log_p_hat)
if np.sqrt(np.sum((alpha_old - alpha_new)**2)) < 1e-9:
break
alpha_old = alpha_new
return alpha_new
def dirichlet_normality_score(alpha, p):
return np.sum((alpha - 1) * np.log(p), axis=-1)
'''scores = np.zeros((len(x_test),))
observed_data = x_train_task
for t_ind in range(transformer.n_transforms):
observed_dirichlet = mdl.predict(transformer.transform_batch(observed_data, [t_ind] * len(observed_data)),
batch_size=64)
log_p_hat_train = np.log(observed_dirichlet).mean(axis=0)
alpha_sum_approx = calc_approx_alpha_sum(observed_dirichlet)
alpha_0 = observed_dirichlet.mean(axis=0) * alpha_sum_approx
mle_alpha_t = fixed_point_dirichlet_mle(alpha_0, log_p_hat_train)
x_test_p = mdl.predict(transformer.transform_batch(x_test, [t_ind] * len(x_test)),
batch_size=64)
scores += dirichlet_normality_score(mle_alpha_t, x_test_p)'''
scores /= transformer.n_transforms
labels = y_test.flatten() == single_class_ind
r = (roc_auc_score(labels, scores))
f = open("guru99.txt", "a")
f.write(str(r) + "\n")
f.close()
'''res_file_name = '{}_transformations_{}_{}.npz'.format(dataset_name,
def octest(hyper_para):
CC = wrn.WideResNet(28, num_classes=10, dropout_rate=0,
widen_factor=10).cuda()
CC.load_state_dict(torch.load(hyper_para.experiment_name + '.pth'))
C = wrn.WideResNet(28, num_classes=72, dropout_rate=0,
widen_factor=10).cuda()
C.cuda().eval()
CC.cuda().eval()
muc = {}
stdc = {}
mu = {}
std = {}
for cname in hyper_para.inclass:
single_class_ind = cname
C.load_state_dict(
torch.load('saved/' + hyper_para.experiment_name + '_' +
str(cname) + '.pth'))
(x_train, y_train), (x_test, y_test) = load_cifar10()
transformer = Transformer(8, 8)
x_train_task0 = x_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
y_train_task = y_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
for t in range(72):
x_train_task = transformer.transform_batch(x_train_task0,
[t] * len(y_train_task))
dset = CustomDataset(x_train_task, y_train_task)
trainLoader = DataLoader(dset, batch_size=128, shuffle=True)
features = []
for inputs0, labels in trainLoader:
inputs = inputs0.permute(0, 3, 2, 1).cuda()
act, f = C(inputs)
features += act[:, t].detach().cpu().tolist()
if t == 0:
totfeatures = features
else:
totfeatures += features
mu[str(cname)] = np.mean(totfeatures)
std[str(cname)] = np.sqrt(np.var(totfeatures))
features = {}
if True:
(x_train, y_train), (x_test, y_test) = load_cifar10()
x_train_task = x_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
y_train_task = y_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
dset = CustomDataset(x_train_task, y_train_task)
trainLoader = DataLoader(dset, batch_size=128, shuffle=True)
for inputs0, labels0 in trainLoader:
inputs = inputs0.permute(0, 3, 2, 1).cuda()
act, f = CC(inputs)
val, ind = torch.max(act, dim=1)
val, ind, labels0 = (val.detach().cpu().tolist(),
ind.detach().cpu().tolist(),
labels0.detach().cpu().tolist())
for idx, (ii, gt) in enumerate(zip(ind, labels0)):
gt = gt[0]
if ii == gt and gt in hyper_para.inclass:
if str(ii) not in features.keys():
features[str(ii)] = [
act[idx, ii].detach().cpu().tolist()
]
else:
features[str(ii)] += [
act[idx, ii].detach().cpu().tolist()
]
print(np.shape(features[str(ii)]))
for k in features.keys():
muc[str(k)] = np.mean(features[str(k)])
stdc[str(k)] = np.sqrt(np.var(features[str(k)]))
(x_train, y_train), (x_test, y_test) = load_cifar10()
transformer = Transformer(8, 8)
scores = np.array([[]])
features = []
lbl = []
correct = 0
total = 0
preds = np.zeros((len(x_test), transformer.n_transforms))
dset = CustomDataset(x_test, y_test)
testLoader = DataLoader(dset, batch_size=128)
score = []
for i, (inputs0, labels) in enumerate(testLoader):
inputs = inputs0.permute(0, 3, 2, 1)
if True:
inputs = inputs.cuda()
labels = labels.cuda()
act0, f = CC(inputs)
features += f.detach().cpu().tolist()
val, ind = torch.max(act0, dim=1)
val, ind, labels = (val.detach().cpu().tolist(),
ind.detach().cpu().tolist(),
labels.detach().cpu().tolist())
#act = torch.nn.functional.softmax(act, dim=1)
#score += val
for idx, (ii, gt) in enumerate(zip(ind, labels)):
C.load_state_dict(
torch.load('saved/' + hyper_para.experiment_name + '_' +
str(ii) + '.pth'))
gt = gt[0]
score_temp = []
for t in range(72):
x_test0 = transformer.transform_batch(
torch.unsqueeze(inputs0[idx, :, :, :],
0).detach().cpu().numpy(), [t])
inputs = torch.tensor(x_test0).permute(0, 3, 2, 1).cuda()
act, _ = C(inputs)
act = act[:, t]
act = act.detach().cpu().tolist()
if t == 0:
score_temp = act[0]
else:
score_temp += act[t]
score += [(score_temp - mu[str(ii)]) / (std[str(ii)]) +
(val[idx] - muc[str(ii)]) / (stdc[str(ii)])]
if gt in hyper_para.inclass:
total += 1
if ii == gt:
correct += 1
lbl.append(1)
else:
lbl.append(0)
fpr, tpr, thresholds = metrics.roc_curve(lbl, score)
AUC = metrics.auc(fpr, tpr)
print('AUROC: ' + str(AUC))
return ([0, 0])
from tqdm import tqdm
from scripts.detached_transformer_od_hits import calc_approx_alpha_sum, fixed_point_dirichlet_mle, dirichlet_normality_score, plot_histogram_disc_loss_acc_thr
if __name__ == "__main__":
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # dynamically grow the memory used on the GPU
sess = tf.Session(config=config)
set_session(sess)
single_class_ind = 1
(x_train, y_train), (x_test, y_test) = load_fashion_mnist()
print(x_train.shape)
print(x_test.shape)
transformer = Transformer(8, 8)
n, k = (10, 4)
mdl = create_wide_residual_network(input_shape=x_train.shape[1:],
num_classes=transformer.n_transforms,
depth=n,
widen_factor=k)
mdl.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['acc'])
print(mdl.summary())
# get inliers of specific class
x_train_task = x_train[y_train.flatten() == single_class_ind]
print(x_train_task.shape)
def octrain(hyper_para):
for single_class_ind in [0]: #range(10):
hyper_para.C = wrn.WideResNet(28,
num_classes=72,
dropout_rate=0,
widen_factor=10)
(x_train, y_train), (x_test, y_test) = load_cifar10()
x_train_task = x_train[y_train.flatten() == single_class_ind]
dset = RotateDataset(x_train_task)
trainLoader = DataLoader(dset, batch_size=512, shuffle=True)
transformer = Transformer(8, 8)
'''transformations_inds = np.tile(np.arange(transformer.n_transforms), len(x_train_task))
x_train_task_transformed = transformer.transform_batch(np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
dset = CustomDataset(x_train_task_transformed,transformations_inds )
trainLoader = DataLoader(dset, batch_size=32, shuffle=True)'''
# define networks
C = hyper_para.C
C = torch.nn.DataParallel(C)
# define loss functions
ce_criterion = nn.CrossEntropyLoss()
# define optimizer
optimizer_c = optim.Adam(C.parameters(),
lr=hyper_para.lr,
betas=(0.9, 0.999))
# turn on the train mode
C.train(mode=True)
# initialization of auxilary variables
running_tl = 0.0
running_cc = 0.0
running_rc = 0.0
running_ri = 0.0
# if gpu use cuda
for i in range(100):
if hyper_para.gpu:
C.cuda()
for idx, (inputs, labels) in enumerate(trainLoader):
inputs = inputs.permute(0, 3, 2, 1)
t1 = time.time()
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels.cuda()
act, _ = C(inputs)
loss_cc = ce_criterion(act, labels)
optimizer_c.zero_grad()
loss_cc.backward()
optimizer_c.step()
running_cc += loss_cc.data
t2 = time.time()
if idx % 10 == 0:
line = hyper_para.BLUE + '[' + str(format(i + 1, '8d')) + '/' + str(
format(int(hyper_para.iterations), '8d')) + ']' + hyper_para.ENDC + \
hyper_para.GREEN + ' loss_cc: ' + hyper_para.ENDC + str(
format(running_cc / hyper_para.stats_frequency, '1.8f')) + \
hyper_para.YELLOW + ' time (min): ' + hyper_para.ENDC + str(int((t2 - t1) * 20.0))
print(line)
running_cc = 0.0
torch.save(
C.module.state_dict(),
hyper_para.experiment_name + '_' + str(single_class_ind) + '.pth')
def __init__(self, x_tensor):
self.x = x_tensor
self.transformer = Transformer(8, 8)
def _transformations_experiment(dataset_load_fn, dataset_name, single_class_ind, gpu_q):
gpu_to_use = gpu_q.get()
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_to_use
(x_train, y_train), (x_test, y_test) = dataset_load_fn()
if dataset_name in ['cats-vs-dogs']:
transformer = Transformer(16, 16)
n, k = (16, 8)
else:
transformer = Transformer(8, 8)
n, k = (10, 4)
mdl = create_wide_residual_network(x_train.shape[1:], transformer.n_transforms, n, k)
mdl.compile('adam',
'categorical_crossentropy',
['acc'])
x_train_task = x_train[y_train.flatten() == single_class_ind]
transformations_inds = np.tile(np.arange(transformer.n_transforms), len(x_train_task))
x_train_task_transformed = transformer.transform_batch(np.repeat(x_train_task, transformer.n_transforms, axis=0),
transformations_inds)
batch_size = 128
mdl.fit(x=x_train_task_transformed, y=to_categorical(transformations_inds),
batch_size=batch_size, epochs=int(np.ceil(200/transformer.n_transforms))
)
#################################################################################################
# simplified normality score
#################################################################################################
# preds = np.zeros((len(x_test), transformer.n_transforms))
# for t in range(transformer.n_transforms):
# preds[:, t] = mdl.predict(transformer.transform_batch(x_test, [t] * len(x_test)),
# batch_size=batch_size)[:, t]
#
# labels = y_test.flatten() == single_class_ind
# scores = preds.mean(axis=-1)
#################################################################################################
def calc_approx_alpha_sum(observations):
N = len(observations)
f = np.mean(observations, axis=0)
return (N * (len(f) - 1) * (-psi(1))) / (
N * np.sum(f * np.log(f)) - np.sum(f * np.sum(np.log(observations), axis=0)))
def inv_psi(y, iters=5):
# initial estimate
cond = y >= -2.22
x = cond * (np.exp(y) + 0.5) + (1 - cond) * -1 / (y - psi(1))
for _ in range(iters):
x = x - (psi(x) - y) / polygamma(1, x)
return x
def fixed_point_dirichlet_mle(alpha_init, log_p_hat, max_iter=1000):
alpha_new = alpha_old = alpha_init
for _ in range(max_iter):
alpha_new = inv_psi(psi(np.sum(alpha_old)) + log_p_hat)
if np.sqrt(np.sum((alpha_old - alpha_new) ** 2)) < 1e-9:
break
alpha_old = alpha_new
return alpha_new
def dirichlet_normality_score(alpha, p):
return np.sum((alpha - 1) * np.log(p), axis=-1)
scores = np.zeros((len(x_test),))
observed_data = x_train_task
for t_ind in range(transformer.n_transforms):
observed_dirichlet = mdl.predict(transformer.transform_batch(observed_data, [t_ind] * len(observed_data)),
batch_size=1024)
log_p_hat_train = np.log(observed_dirichlet).mean(axis=0)
alpha_sum_approx = calc_approx_alpha_sum(observed_dirichlet)
alpha_0 = observed_dirichlet.mean(axis=0) * alpha_sum_approx
mle_alpha_t = fixed_point_dirichlet_mle(alpha_0, log_p_hat_train)
x_test_p = mdl.predict(transformer.transform_batch(x_test, [t_ind] * len(x_test)),
batch_size=1024)
scores += dirichlet_normality_score(mle_alpha_t, x_test_p)
scores /= transformer.n_transforms
labels = y_test.flatten() == single_class_ind
res_file_name = '{}_transformations_{}_{}.npz'.format(dataset_name,
get_class_name_from_index(single_class_ind, dataset_name),
datetime.now().strftime('%Y-%m-%d-%H%M'))
res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name)
save_roc_pr_curve_data(scores, labels, res_file_path)
mdl_weights_name = '{}_transformations_{}_{}_weights.h5'.format(dataset_name,
get_class_name_from_index(single_class_ind, dataset_name),
datetime.now().strftime('%Y-%m-%d-%H%M'))
mdl_weights_path = os.path.join(RESULTS_DIR, dataset_name, mdl_weights_name)
mdl.save_weights(mdl_weights_path)
gpu_q.put(gpu_to_use)
def os_train(hyper_para):
C = wrn.WideResNet(28, num_classes=10, dropout_rate=0,
widen_factor=10).cuda()
single_class_ind = hyper_para.inclass[0]
(x_train, y_train), (x_test, y_test) = load_cifar10()
transformer = Transformer(8, 8)
x_train_task = x_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
y_train_task = y_train[[
i in hyper_para.inclass for i in y_train.flatten()
]]
print(np.shape(y_train_task))
print(np.shape(y_train))
y_train_task = y_train_task.astype(int)
#x_train_task = x_train_task[0:15][:][:][:]
dset = CustomDataset(x_train_task, y_train_task)
trainLoader = DataLoader(dset, batch_size=128, shuffle=True)
# define networks
# define loss functions
ce_criterion = nn.CrossEntropyLoss()
# define optimizer
optimizer_c = optim.Adam(C.parameters(),
lr=hyper_para.lr,
betas=(0.9, 0.999))
# turn on the train mode
C.train(mode=True)
# initialization of auxilary variables
running_tl = 0.0
running_cc = 0.0
running_rc = 0.0
running_ri = 0.0
# if gpu use cuda
for i in range(int(20)):
if hyper_para.gpu:
C.cuda()
for iii, (inputs, labels) in enumerate(trainLoader):
inputs = inputs.permute(0, 3, 2, 1)
t1 = time.time()
if hyper_para.gpu:
inputs = inputs.cuda()
labels = labels[:, 0].cuda()
act, _ = C(inputs)
loss_cc = ce_criterion(act, labels)
optimizer_c.zero_grad()
loss_cc.backward()
optimizer_c.step()
running_cc += loss_cc.data
t2 = time.time()
if iii % 50 == 0:
line = hyper_para.BLUE + '[' + str(format(i + 1, '8d')) + '/' + str(
format(int(hyper_para.iterations), '8d')) + ']' + hyper_para.ENDC + \
hyper_para.GREEN + ' loss_cc: ' + hyper_para.ENDC + str(
format(running_cc / hyper_para.stats_frequency, '1.8f')) + \
hyper_para.YELLOW + ' time (min): ' + hyper_para.ENDC + str(int((t2 - t1) * 20.0))
print(line)
running_cc = 0.0
torch.save(C.state_dict(), hyper_para.experiment_name + '.pth')
