def _get_gauss_params(self):
'''
Returns the mean and covariance for the gaussian model on the whole
patch (i.e., window to sample plus padding around it)
'''
means = np.zeros((3, self.patchSize * self.patchSize))
covs = np.zeros((3, self.patchSize * self.patchSize,
self.patchSize * self.patchSize))
path_mean = self.path_folder + '{}_means{}_indep'.format(
"test", self.patchSize)
path_cov = self.path_folder + '{}_covs{}_indep'.format(
"test", self.patchSize)
# check if values are already precomputed and saved; otherwise do so first
if os.path.exists(path_mean + '.npy') and os.path.exists(path_cov +
'.npy'):
means = np.load(path_mean + '.npy')
covs = np.load(path_cov + '.npy')
else:
for c in [0, 1, 2]:
#net = utlC.get_caffenet(self.netname)
# get the imagenet data
X, _, _ = utlD.get_imagenet_data()
# get samples for fitting the distribution
patchesMat = np.empty((0, self.patchSize * self.patchSize),
dtype=np.float)
for i in range(int(self.num_samples_fit / X.shape[0]) + 1):
# get a random (upper left) position of the patch
idx = random.sample(
range((self.image_dims[0] - self.patchSize) *
(self.image_dims[1] - self.patchSize)), 1)[0]
idx = np.unravel_index(
idx, (self.image_dims[0] - self.patchSize,
self.image_dims[1] - self.patchSize))
idx = [idx[0], idx[1]]
# get the patch from all the images in X, from the given channel
patch = X[:, c, idx[0]:idx[0] + self.patchSize,
idx[1]:idx[1] + self.patchSize]
patchesMat = np.vstack(
(patchesMat,
patch.reshape(
(X.shape[0], self.patchSize * self.patchSize))))
# compute the mean and covariance of the collected samples
means[c] = np.mean(patchesMat, axis=0)
covs[c] = np.cov(patchesMat.T)
# save the mean and the covariance
np.save(path_mean, means)
np.save(path_cov, covs)
return means, covs
def save_minmax_values(netname):
'''
When X.npy is updated, this can be executed to also update the min/max
values of the data (which is being used to cut off the values in the
sampler so that we don't have overflowing values)
'''
#net = utlC.get_caffenet(netname)
X, _, _ = utlD.get_imagenet_data()
minMaxVals = np.zeros((2, 3, X.shape[-1], X.shape[-1]))
minMaxVals[0] = np.min(X, axis=0)
minMaxVals[1] = np.max(X, axis=0)
path_folder = './gaussians/'
if not os.path.exists(path_folder):
os.makedirs(path_folder)
np.save(path_folder + '{}_minMaxVals'.format(netname), minMaxVals)
# settings for sampling
sampl_style = 'conditional' # choose: conditional / marginal
num_samples = 10
padding_size = 2 # important for conditional sampling,
# l = win_size+2*padding_size in alg 1
# (see paper)
# set the batch size - the larger, the faster computation will be
# (if caffe crashes with memory error, reduce the batch size)
batch_size = 30
# ------------------------ SET-UP ------------------------
# get the data
X_test, X_test_im, X_filenames = utlD.get_imagenet_data(net=model)
# get the label names of the 1000 ImageNet classes
classnames = utlD.get_imagenet_classnames()
if not test_indices:
test_indices = [i for i in range(X_test.shape[0])]
# make folder for saving the results if it doesn't exist
path_results = './results/'
if not os.path.exists(path_results):
os.makedirs(path_results)
# ------------------------ EXPERIMENTS ------------------------
# change the batch size of the network to the given value