def process_patch_features(model, patch, cuda):
patch = np.moveaxis(patch, -1, 0)
patch = np.expand_dims(patch, axis=0)
patch = clip_and_scale_image(patch, 'rgb')
# Embed tile
patch = torch.from_numpy(patch).float()
patch = Variable(patch)
if cuda: patch = patch.cuda()
z = model.encode(patch)
if cuda: z = z.cpu()
z = z.data.numpy()
return z
def get_emax_features(img_names,
model,
z_dim,
cuda,
bands=5,
patch_size=50,
patch_per_img=36,
save=True,
verbose=False,
npy=True):
model.eval()
X = np.zeros((len(img_names), z_dim))
for k in range(len(img_names)):
img_name = img_names[k]
if verbose:
print("Getting features for: " + img_name)
img = load_landsat(img_name, bands, bands_only=True, is_npy=npy)
img_shape = img.shape
offset_row = (img_shape[0] - 300) // 2
offset_col = (img_shape[1] - 300) // 2
output = np.zeros((36, z_dim))
for i in range(6):
for j in range(6):
start_row = i * 50 + offset_row
end_row = (i + 1) * 50 + offset_row
start_col = j * 50 + offset_col
end_col = (j + 1) * 50 + offset_col
# print("Rows: " + str((start_row,end_row)))
# print("Cols: " + str((start_col,end_col)))
patch = img[start_row:end_row, start_col:end_col, :]
patch = np.moveaxis(patch, -1, 0)
patch = np.expand_dims(patch, axis=0)
patch = clip_and_scale_image(patch, 'landsat')
# Embed tile
patch = torch.from_numpy(patch).float()
patch = Variable(patch)
if cuda: patch = patch.cuda()
z = model.encode(patch)
if cuda: z = z.cpu()
z = z.data.numpy()
output[(i + 1) * (j + 1) - 1, :] += z.reshape(-1)
output_emax = np.zeros((1, z_dim))
for i in range(36):
output_emax[0][i] = np.amax(output[:, i])
X[k] = output_emax
return X
def __getitem__(self, index):
# Get tile
path = self.list_ids.iloc[index].patch_dir
name = os.path.basename(path)
tile = load_patch(path)
# randomly sample 50x50 section of the tile
# x_idx = np.random.randint(low=0, high=120-50)
# y_idx = np.random.randint(low=0, high=120-50)
# tile = tile[x_idx:x_idx+50,y_idx:y_idx+50,:]
tile = tile[35:85, 35:85, :]
# Scale
tile = clip_and_scale_image(tile, img_type='landsat')
# Rearrange to PyTorch order
tile = np.moveaxis(tile, -1, 0)
tile = np.expand_dims(tile, axis=0)
# Load data and get label
y = multihot(self.y[name], self.labs)
return tile, y
def __call__(self, sample):
a, n, d = (clip_and_scale_image(sample['anchor'], self.img_type),
clip_and_scale_image(sample['neighbor'], self.img_type),
clip_and_scale_image(sample['distant'], self.img_type))
sample = {'anchor': a, 'neighbor': n, 'distant': d}
return sample