def sample_fn_csgm_kde(present, num_samples, compute_labels=True):
'''
present = which segments to choose from...
'''
data = np.zeros((num_samples, n_features))
data[:, present] = 1
print(num_samples)
# now generate some images
_, mask = image_utils.create_mask(None,
segments, {'feature': present},
rgb=True)
target = self.get_target(image, mask)
if self.present != present:
self.sampler = csgm.CSGM(target,
mask,
self.G,
num_samples,
bandwidth=0.5)
self.present = present
BS = 64
#raw_data = np.zeros((num_samples,28,28))
raw_data = data
_, _, backgrounds = self.sampler.sample(num_samples)
current_batch = backgrounds + target
current_preds = classifier_fn(current_batch)
current_preds_max = np.argmax(current_preds, axis=1)
labels = (current_preds_max == true_label).astype(int)
return raw_data, data, labels
def sample_fn_csgm(present, num_samples, compute_labels=True):
'''
present = which segments to choose from...
'''
if not compute_labels:
data = np.random.randint(0, 2,
num_samples * n_features).reshape(
(num_samples, n_features))
data[:, present] = 1
return [], data, []
data = np.zeros((num_samples, n_features))
data[:, present] = 1
# now generate some images
_, mask = image_utils.create_mask(None, segments,
{'feature': present})
target = self.get_target(image, mask)
#plt.imshow(target)
BS = 64
#raw_data = np.zeros((num_samples,28,28))
raw_data = data
labels = np.zeros((num_samples)).astype(int)
for j in range(0, num_samples, BS):
n_s = min(num_samples, j + BS) - j
if self.true_sampling:
_, _, backgrounds = csgm.reconstruct_batch_threshold(
target,
mask,
np.sum(mask),
self.G,
n_s,
self.sample_p(n_s),
lr=1e-1 if self.batch_norm else 1e-2)
else:
_, _, backgrounds = csgm.reconstruct_batch(
target,
mask,
np.sum(mask),
self.G,
n_s,
lr=1e-1 if self.batch_norm else 1e-2,
threshold=self.threshold)
current_batch = np.zeros((n_s, 28, 28))
for i in range(len(backgrounds)):
temp = copy.deepcopy(target)
curr = backgrounds[i] ## should be (28,28)
temp += curr
current_batch[i] = np.expand_dims(temp, 0)
current_preds = classifier_fn(current_batch)
current_preds_max = np.argmax(current_preds, axis=1)
current_labels = (current_preds_max == true_label).astype(int)
labels[j:j + n_s] = current_labels
return raw_data, data, labels
def sample_fn_stein(present, num_samples, compute_labels=True):
'''
present = which segments to choose from...
'''
data = np.zeros((num_samples, n_features))
data[:, present] = 1
print(num_samples)
# now generate some images
_, mask = image_utils.create_mask(None,
segments, {'feature': present},
rgb=True)
target = self.get_target(image, mask)
# first time we called it
if self.stein is None:
# setup for stein's
G = dcgan.ProbGenerator(self.G, mask, target)
K = svgd.RBF()
X = torch.randn(num_samples, 100).cuda()
adam = torch.optim.Adam([X], lr=1e-1)
self.stein = svgd.SVGD(G, K, adam)
# train...
print("Starting Stein Training")
X = self.stein.train(X, num_iter=1000)
print("Trained!")
BS = 64
collected = 0
labels = np.zeros((num_samples)).astype(int)
raw_data = data
print("Predicting...")
while collected < num_samples:
X = self.stein.sample(BS)
if X.shape[0] == 0:
# print("Bad samples")
continue
#print("KDE found", X.shape[0], "good samples.")
backgrounds = X.view(-1, 28,
28).data.cpu().numpy() * (1 - target)
end = min(backgrounds.shape[0], num_samples - collected)
backgrounds = backgrounds[:end]
current_batch = backgrounds + target
current_preds = classifier_fn(current_batch)
current_preds_max = np.argmax(current_preds, axis=1)
current_labels = (current_preds_max == true_label).astype(int)
labels[collected:collected + end] = current_labels
collected += end
print("Collected", num_samples, "!")
return raw_data, data, labels