def wave_tensor_data_backbone(batch_x, depth=-1, i_ref=-1, j_ref=-1):
batch_size, _ = batch_x.shape
FLAG_RAND_I = i_ref == -1
FLAG_RAND_J = j_ref == -1
FLAG_DEPTH = depth == -1
wave_tensor = init_wave_tensor(batch_size)
for num_batch in range(batch_size):
image = mnist_reshape_32(batch_x[num_batch])
w = WaveImage(image=image)
if FLAG_RAND_I:
i_ref = np.random.randint(16)
if FLAG_RAND_J:
j_ref = np.random.randint(16)
if FLAG_DEPTH:
depth = 1 + np.random.randint(6)
pow2_i = calc_pow2(i_ref)
pow2_j = calc_pow2(j_ref)
for h in range(6 - depth, 6):
data_h = w.get_data()[h]
if h == 0:
wave_tensor[h][num_batch][0][0][0] = data_h[(0, 0)] #/ 4**4
else:
u = (pow2_i[h - 1], pow2_j[h - 1])
#for u in data_h:
# wave_tensor[h][num_batch][u[0]][u[1]][:] = 0
wave_tensor[h][num_batch][u[0]][u[1]][:] = data_h[
u] #/ 4 ** (5 - h)
return wave_tensor
def wave_tensor_data_backbone(batch_x, depth=-1, i_ref=-1, j_ref=-1):
tab_depth = [1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 6]
batch_size, _ = batch_x.shape
FLAG_RAND_I = i_ref == -1
FLAG_RAND_J = j_ref == -1
FLAG_DEPTH = depth == -1
wave_tensor = init_wave_tensor(batch_size)
for num_batch in range(batch_size):
image = mnist_reshape_32(batch_x[num_batch])
w = WaveImage(image=image)
if FLAG_RAND_I:
#i_ref = np.random.randint(16)
# RND-PARTS
liste_u_full = []
for i in range(16):
for j in range(16):
liste_u_full += [(i, j)]
indices_perm = np.arange(256, dtype='int')
np.random.shuffle(indices_perm)
nb_u = np.random.randint(256)
liste_u_full = np.array(liste_u_full)
liste_u = liste_u_full[indices_perm]
liste_u = liste_u[:nb_u]
if FLAG_RAND_J:
#j_ref = np.random.randint(16)
pass # RND-PARTS
if FLAG_DEPTH:
# indice_depth = np.random.randint(21)
# depth = tab_depth[indice_depth] # BIASED
# depth = 1 + np.random.randint(6) # UNIF
# depth = 6 # BASE
#indice_depth = np.random.randint(256) #FULL
#if indice_depth < 1:
# depth = 6
#elif indice_depth < 4:
# depth = 4
#elif indice_depth < 16:
# depth = 3
#elif indice_depth < 64:
# depth = 2
#else:
# depth = 1
depth = 6 # RND-PARTS
for u in liste_u:
i_ref = u[0]
j_ref = u[1]
pow2_i = calc_pow2(i_ref)
pow2_j = calc_pow2(j_ref)
for h in range(6 - depth, 6):
data_h = w.get_data()[h]
if h == 0:
wave_tensor[h][num_batch][0][0][0] = data_h[(0,
0)] #/ 4**4
else:
u = (pow2_i[h - 1], pow2_j[h - 1])
#for u in data_h:
# wave_tensor[h][num_batch][u[0]][u[1]][:] = 0
wave_tensor[h][num_batch][u[0]][u[1]][:] = data_h[
u] #/ 4 ** (5 - h)
return wave_tensor
def wave_tensor_data(batch_x):
batch_size, _ = batch_x.shape
wave_tensor = {}
for h in range(6):
if h == 0:
h_size = 1
wave_tensor[h] = np.zeros((batch_size, h_size, h_size, 1))
else:
h_size = 2**(h - 1)
wave_tensor[h] = np.zeros((batch_size, h_size, h_size, 3))
for num_batch in range(batch_size):
image = mnist_reshape_32(batch_x[num_batch])
w = WaveImage(image=image)
for h in range(w.get_h_max()):
data_h = w.get_data()[h]
if h == 0:
wave_tensor[h][num_batch][0][0][0] = data_h[(0, 0)]
else:
for u in data_h:
wave_tensor[h][num_batch][u[0]][u[1]][:] = data_h[u]
return wave_tensor
# ## Creation de la base d'apprentissage
# In[12]:
file_name = "mnist-waveimage-train-mu-Sigma-rho.pkl"
if not os.path.isfile(file_name):
B_train = []
for i in range(len(mnist.train.images)):
if i % 1000 == 0:
sys.stdout.write('\rstep %d' % i)
sys.stdout.flush()
c = mnist.train.labels[i]
x_ref = mnist.train.images[i]
image = mnist_reshape_32(x_ref)
w = WaveImage(image=image)
data = w.get_data()
for h in range(w.get_h_max()):
data_h = w.get_data()[h]
for u in data_h:
v = data_h[u]
B_train += [(v, (c, h, u))]
### Dictionnaire (Base d'apprentissage)
Data_train = [[], [], [], [], [], [], [], [], [], []]
for c in range(10):
Data_train[c] = [{}, {}, {}, {}, {}, {}]
for d in B_train: