img, gt = img[30:115, 24:94, :], gt[30:115, 24:94]
PATCH_SIZE = 13
REG_Coef_, NEIGHBORING_, RO_ = 1e2, 30, 0.4
REG_Coef_K, NEIGHBORING_K, RO_K, GAMMA = 1e3, 30, 0.8, 10
if im_ == 'PaviaU':
img, gt = img[150:350, 100:200, :], gt[150:350, 100:200]
REG_Coef_, NEIGHBORING_, RO_ = 1e3, 20, 0.6
REG_Coef_K, NEIGHBORING_K, RO_K, GAMMA = 6 * 1e4, 30, 0.8, 100
n_row, n_column, n_band = img.shape
x_img = minmax_scale(img.reshape(n_row * n_column, n_band)).reshape((n_row, n_column, n_band))
print('original img shape: ', x_img.shape)
# # reduce spectral bands using PCA
pca = PCA(n_components=nb_comps)
img = minmax_scale(pca.fit_transform(img.reshape(n_row * n_column, n_band))).reshape(n_row, n_column, nb_comps)
x_patches, y_ = p.get_HSI_patches_rw(img, gt, (PATCH_SIZE, PATCH_SIZE))
print('reduced img shape: ', img.shape)
print('x_patch tensor shape: ', x_patches.shape)
n_samples, n_width, n_height, n_band = x_patches.shape
x_patches_2d = np.reshape(x_patches, (n_samples, -1))
y = p.standardize_label(y_)
# # reorder samples according to gt
x_patches_2d, y = order_sam_for_diag(x_patches_2d, y)
# # normalize data
x_patches_2d = normalize(x_patches_2d)
print('final sample shape: %s, labels: %s' % (x_patches_2d.shape, np.unique(y)))
N_CLASSES = np.unique(y).shape[0] # Indian : 8 KSC : 10 SalinasA : 6 PaviaU : 8
# ========================
EPOCH = 100
LEARNING_RATE = 0.0002
REG_GRAPH = 0.001 # beta
REG_TASK = 100. # alpha
WEIGHT_DECAY = 0.001 # lambda
VERBOSE_TIME = 10
n_row, n_column, n_band = img.shape
img = minmax_scale(img.reshape(n_row * n_column,
n_band)).reshape(img.shape)
# perform PCA
# pca = PCA(n_components=nb_comps)
# img = pca.fit_transform(img.reshape(n_row * n_column, n_band)).reshape((n_row, n_column, nb_comps))
# print('pca shape: %s, percentage: %s' % (img.shape, np.sum(pca.explained_variance_ratio_)))
x_patches, y_ = p.get_HSI_patches_rw(img, gt,
(NEIGHBORING_SIZE, NEIGHBORING_SIZE))
x_patches = normalize(x_patches.reshape(x_patches.shape[0],
-1)).reshape(x_patches.shape)
print('img shape:', img.shape)
print('img_patches_nonzero:', x_patches.shape)
n_samples, n_width, n_height, n_band = x_patches.shape
y = p.standardize_label(y_)
x_patches, y = order_sam_for_diag(x_patches, y)
print('x_patches shape: %s, labels: %s' % (x_patches.shape, np.unique(y)))
N_CLASSES = np.unique(y).shape[
0] # wuhan : 5 Pavia : 6 Indian : 8 KSC : 10 SalinasA : 6 PaviaU : 8
"""
=======================================
Clustering
======================================