def main(directory):
filenames = data.get_image_files(directory)
bs = 1000
batches = [
filenames[i * bs:(i + 1) * bs]
for i in range(int(len(filenames) / bs) + 1)
]
# compute mean and std
std(filenames, bs)
Us, evs = [], []
for batch in batches:
images = np.array([data.load_augment(f, 128, 128) for f in batch])
X = images.transpose(0, 2, 3, 1).reshape(-1, 3)
cov = np.dot(X.T, X) / X.shape[0]
U, S, V = np.linalg.svd(cov)
ev = np.sqrt(S)
Us.append(U)
evs.append(ev)
print('U')
print(np.mean(Us, axis=0))
print('eigenvalues')
print(np.mean(evs, axis=0))
def __getitem__(self, idx):
fname = self.input_path + '/' + self.file_list[idx]
w = 224
h = 224
aug_params = {
'zoom_range': (1 / 1.15, 1.15),
'rotation_range': (0, 360),
'shear_range': (0, 0),
'translation_range': (-20, 20),
'do_flip': True,
'allow_stretch': True,
}
sigma = 0.25
image = load_augment(fname,
w,
h,
aug_params=aug_params,
transform=None,
sigma=sigma,
color_vec=None)
#print('after', image.shape)
data = h5py.File(self.input_path + '/' + self.file_list[idx], 'r')
#image = data['image'].value
#print('before', image.shape)
target = float(data['target'].value)
#one_hot = np.zeros(5)
#one_hot[target] = 1
if self.transform is not None:
image = self.transform(image)
return image, torch.from_numpy(np.array([target]))
def compute_mean_std(files, batch_size=128):
"""Load images in files in batches and compute mean."""
m = np.zeros(3)
s = np.zeros(3)
s2 = np.zeros(3)
for i in range(0, len(files), batch_size):
print("done with {:>3} / {} images".format(i, len(files)))
images = np.array([data.load_augment(f,256,256) for f in files[i : i + batch_size]])
m += images.sum(axis=(0, 2, 3))
shape = images.shape
s2 += np.power(images, 2).sum(axis=(0, 2, 3))
n = len(files) * shape[2] * shape[3]
var = (s2 - m**2.0 / n) / (n - 1)
return (m / n).astype(np.float32),np.sqrt(var)
def main(filelist):
filenames = data.get_files(filelist)
bs = 1000
batches = [filenames[i * bs : (i + 1) * bs]
for i in range(int(len(filenames) / bs) + 1)]
Us, evs = [], []
for batch in batches:
images = np.array([data.load_augment(f, 256, 256) for f in batch])
X = images.transpose(0, 2, 3, 1).reshape(-1, 3)
cov = np.dot(X.T, X) / X.shape[0]
U, S, V = np.linalg.svd(cov)
ev = np.sqrt(S)
Us.append(U)
evs.append(ev)
print('U')
print(np.mean(Us, axis=0))
print('eigenvalues')
print(np.mean(evs, axis=0))
def load_shared(args):
i, array_name, fname, kwargs = args
array = SharedArray.attach(array_name)
array[i] = data.load_augment(fname, **kwargs)
def load_shared(args):
i, array_name, fname, kwargs = args
array = SharedArray.attach(array_name)
array[i] = data.load_augment(fname, **kwargs)
def load_shared(args):
i, array_name, fname, kwargs = args
np.random.seed(i) #comment this line if you want a non reproducible result
array = SharedArray.attach(array_name)
array[i] = data.load_augment(fname, **kwargs)