def test_one_hop(ext_test_imgs):
# a = time.time()
# n_samp = ext_test_imgs.shape[0]
# num_cores = int(multiprocessing.cpu_count() / 2)
#
# test_one_hop_attri_saab = Parallel(n_jobs=num_cores, backend='multiprocessing')(delayed(concatenate)(ext_test_imgs, k, 2, True) for k in range(n_samp))
# print(type(test_one_hop_attri_saab))
# test_one_hop_attri_saab = np.array(test_one_hop_attri_saab)
# print(test_one_hop_attri_saab.shape)
# b = time.time()
# print("time for concatenate 1st test:", b - a)
#
# # Saab dimension reduction from 27 dim to 12 dim
# test_one_hop_attri = Saab_Getfeature(test_one_hop_attri_saab, pca_params_1)
# print("Test 1st hop attribute shape:", test_one_hop_attri.shape)
# gt = np.zeros((ext_test_imgs.shape[0], ext_test_imgs.shape[1], ext_test_imgs.shape[2]))
# test_one_hop_attri = PixelHop_Unit(ext_test_imgs, gt, dilate=np.array([1]), num_AC_kernels=27, pad='reflect',
# weight_root=save_dir_matFiles + 'weight',
# getK=True, Pass_Ener_thrs=0.2, energy_percent=0.02, useDC=False, stride=None,
# getcov=0, split_spec=1, hopidx=1)
test_one_hop_attri_saab = PixelHop_Unit(ext_test_imgs, dilate=0, window_size=2, idx_list=None, getK=False, pad='reflect',
weight_root=save_dir_matFiles + 'weight',
Pass_Ener_thrs=0.1, Leaf_Ener_thrs=0.0001, num_kernels=None, PCA_ener_percent=0.99,
useDC=False, stride=None, getcov=0, split_spec=0, hopidx=1)
test_one_hop_attri_biased = PixelHop_fit(weight_name=save_dir_matFiles + 'weight',feature_ori=test_one_hop_attri_saab,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
return test_one_hop_attri_biased
def one_hop(extended_imgs):
# # 8 nearest neighbors
# a = time.time()
# n_samp = extended_imgs.shape[0]
#
# num_cores = int(multiprocessing.cpu_count()/2)
#
# one_hop_attri_saab = Parallel(n_jobs=num_cores, backend='multiprocessing')(delayed(concatenate)(extended_imgs, k, 2, False) for k in range(n_samp))
# print(type(one_hop_attri_saab))
#
# one_hop_attri_saab = np.array(one_hop_attri_saab)
# print(one_hop_attri_saab.shape)
#
# b = time.time()
# print("time for concatenate 1st :", b-a)
# # Saab dimension reduction from 27 dim to 12 dim
# pca_params_1 = Saab_Getkernel(one_hop_attri_saab, "27")
# c = time.time()
# print("time for kernel:", c-b)
#
# one_hop_attri = Saab_Getfeature(one_hop_attri_saab, pca_params_1)
# d = time.time()
# print("time for feature:", d-c)
one_hop_attri_saab = PixelHop_Unit(extended_imgs,
getK=True,
idx_list=None,
dilate=1,
window_size=4,
pad='reflect',
weight_root=save_dir_matFiles +
'weight',
Pass_Ener_thrs=0.1,
Leaf_Ener_thrs=0.0001,
num_kernels=None,
PCA_ener_percent=0.99,
useDC=False,
stride=None,
getcov=0,
split_spec=0,
hopidx=1)
print("Hop-1 feature ori shape:", one_hop_attri_saab.shape)
one_hop_attri_response = PixelHop_fit(weight_name=save_dir_matFiles +
'weight',
feature_ori=one_hop_attri_saab,
split_spec=0,
hopidx=1,
Pass_Ener_thrs=0.1)
print("one hop attribute shape:", one_hop_attri_response.shape)
return one_hop_attri_response
def test_second_hop(test_one_hop_attri_response, idx_stop_list):
# a = time.time()
# n_samp = test_one_hop_attri.shape[0]
# num_cores = int(multiprocessing.cpu_count() / 2)
#
# test_sec_hop_attri_saab = Parallel(n_jobs=num_cores, backend='multiprocessing')(delayed(concatenate)(test_one_hop_attri, k, 3, False) for k in range(n_samp))
# print(type(test_sec_hop_attri_saab))
# test_sec_hop_attri_saab = np.array(test_sec_hop_attri_saab)
# print(test_sec_hop_attri_saab.shape)
# b = time.time()
# print("time for concatenate 2nd :", b - a)
#
# # Saab dimension reduction from 27 dim to 12 dim
# test_sec_hop_attri = Saab_Getfeature(test_sec_hop_attri_saab, pca_params_2)
# print("Test 2nd hop attribute shape:", test_sec_hop_attri.shape)
all_nodes = np.arange(test_one_hop_attri_response.shape[-1]).tolist()
for i in idx_stop_list:
if i in all_nodes:
all_nodes.remove(i)
intermediate_idx = all_nodes
test_sec_hop_attri_saab = PixelHop_Unit(test_one_hop_attri_response,
dilate=7,
window_size=1,
idx_list=intermediate_idx,
getK=False,
pad='reflect',
weight_root=save_dir_matFiles +
'weight',
Pass_Ener_thrs=0.1,
Leaf_Ener_thrs=0.0001,
num_kernels=None,
PCA_ener_percent=0.95,
useDC=False,
stride=None,
getcov=0,
split_spec=1,
hopidx=2)
print("sec hop neighborhood construction shape:",
test_sec_hop_attri_saab.shape)
test_sec_hop_attri_response = PixelHop_fit(
weight_name=save_dir_matFiles + 'weight',
feature_ori=test_sec_hop_attri_saab,
split_spec=1,
hopidx=2,
Pass_Ener_thrs=0.1)
print("sec hop attri shape:", test_sec_hop_attri_response.shape)
return test_sec_hop_attri_response
def test_one_hop(ext_test_imgs):
test_one_hop_attri_saab = PixelHop_Unit(ext_test_imgs, dilate=1, window_size=window_rad,
idx_list=None, getK=False, pad='reflect',
weight_root=os.path.join(save_dir_matFiles, 'weight'),
Pass_Ener_thrs=0.1, Leaf_Ener_thrs=0.0001, num_kernels=None,
PCA_ener_percent=0.99,
useDC=False, stride=None, getcov=0, split_spec=0, hopidx=1)
test_one_hop_attri_biased = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=test_one_hop_attri_saab,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
return test_one_hop_attri_biased
def test_forth_hop(test_third_hop_attri_response):
test_fourth_hop_attri_saab = PixelHop_Unit(test_third_hop_attri_response, dilate=1, window_size=window_rad,
idx_list=None, getK=False, pad='reflect',
weight_root=os.path.join(save_dir_matFiles, 'weight'),
Pass_Ener_thrs=0.1, Leaf_Ener_thrs=0.0001, num_kernels=None,
PCA_ener_percent=0.97,
useDC=False, stride=None, getcov=0, split_spec=1, hopidx=4)
# print("Hop-3 feature ori shape:", test_third_hop_attri_saab.shape)
test_fourth_hop_attri_response = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=test_fourth_hop_attri_saab,
split_spec=1, hopidx=4, Pass_Ener_thrs=0.1)
# print("third hop attri shape:", test_third_hop_attri_response.shape)
return test_fourth_hop_attri_response
def third_hop(response, energy_prev, test_input):
print("---- Hop 3 ----")
sec_hop_total = []
sec_hop_energy = []
sec_hop_test = []
for c in range(response.shape[-1]):
cur_response = response[:, :, :, c][:, :, :, None]
cur_test = test_input[:, :, :, c][:, :, :, None]
cur_attri_saab, cur_saab = PixelHop_Unit(cur_response, getK=True, idx_list=None,
dilate=2, window_size=window_rad, pad='reflect',
weight_root=os.path.join(save_dir_matFiles, 'weight'),
Pass_Ener_thrs=0.1, Leaf_Ener_thrs=0.6, num_kernels=None,
PCA_ener_percent=1,
useDC=False, stride=None, getcov=0, split_spec=0, hopidx=1)
cur_response = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=cur_attri_saab,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
cur_test = Shrink(cur_test)
cur_test_context = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=cur_test,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
print("sec hop attribute shape:", cur_response.shape)
energy = cur_saab['leaf0'].tree['ac_energy']
energy = energy / np.sum(energy) * energy_prev[c]
engr_th = 0.0005
high_engr_channels = np.where((energy > engr_th))
to_append = np.squeeze(cur_response[:, :, :, high_engr_channels])
if len(to_append.shape) == 3:
to_append = to_append[:, :, :, None]
if to_append.shape[-1] == 0:
to_append = np.squeeze(cur_response[:, :, :, :3])
sec_hop_total.append(to_append.copy())
sec_hop_energy.append(energy[high_engr_channels].copy() if len(high_engr_channels) != 0 else energy[:3].copy())
to_append_test = np.squeeze(cur_test_context[:, :, :, high_engr_channels])
if len(to_append_test.shape) == 3:
to_append_test = to_append_test[:, :, :, None]
if to_append_test.shape[-1] == 0:
to_append_test = np.squeeze(cur_test_context[:, :, :, :3])
sec_hop_test.append(to_append_test.copy())
print("CURRENT C:", c)
sec_hop_total = np.concatenate(sec_hop_total, axis=3)
sec_hop_energy = np.concatenate(sec_hop_energy, axis=0)
sec_hop_test = np.concatenate(sec_hop_test, axis=3)
return sec_hop_total, sec_hop_energy, sec_hop_test
def one_hop(response, test_response):
print("---- Hop 1 ----")
one_hop_attri_saab, all_saab = PixelHop_Unit(response, getK=True, idx_list=None,
dilate=2, window_size=window_rad, pad='reflect',
weight_root=os.path.join(save_dir_matFiles, 'weight'),
Pass_Ener_thrs=0.1, Leaf_Ener_thrs=0.6, num_kernels=None,
PCA_ener_percent=1,
useDC=False, stride=None, getcov=0, split_spec=0, hopidx=1)
print("Hop-1 feature ori shape:", one_hop_attri_saab.shape)
one_hop_attri_response = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=one_hop_attri_saab,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
test_response = Shrink(test_response)
test_context = PixelHop_trans(weight_name=os.path.join(save_dir_matFiles, 'weight'),
feature_ori=test_response,
split_spec=0, hopidx=1, Pass_Ener_thrs=0.1)
print("one hop attribute shape:", one_hop_attri_response.shape)
kernels = all_saab['leaf0'].tree['kernel']
energy = all_saab['leaf0'].tree['ac_energy']
energy = energy / np.sum(energy)
kernel_5 = np.array(
[[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0],
]
)
high_frez_pow = []
thres_high = 0.7
engr_th_high = 0.005
engr_th_low = 0.02
for each_kernel in kernels:
each_kernel = np.reshape(each_kernel, (5, 5))
each_fft = np.fft.fftshift(np.fft.fft2(each_kernel))
high_frez_pow.append(1 - np.sum(np.abs(each_fft * kernel_5)) / np.sum(np.abs(each_fft)))
high_frez_pow = np.array(high_frez_pow)
high_frez_channels = np.where((high_frez_pow > thres_high) & (energy > engr_th_high))
low_frez_channels = np.where((high_frez_pow < thres_high) & (energy > engr_th_low))
return (one_hop_attri_response[:, :, :, high_frez_channels], one_hop_attri_response[:, :, :, low_frez_channels],
energy[high_frez_channels], energy[low_frez_channels],
test_context[:, :, :, high_frez_channels], test_context[:, :, :, low_frez_channels])