concentrate speed

feature_model.py

@@ -8,6 +8,7 @@
 from matrix_compute import get_gxyz
 from filter_func import get_normalized_gaussian
 from kmeans_vector import KMeans_Vector
+from preprocessing import get_array_data
 
 import filter_constant as C
 
@@ -40,9 +41,8 @@ def get_features(patchX, patchY, patchZ, weight):
 
  trace, fa, mode = get_lamda_u(l1, l2, l3)
 
- # return angle_p, angle_t, math.log(trace), fa, mode
- # return angle_p, angle_t, math.log(trace)/4, fa, mode/2, index1, sign
- return v1[0], v1[1], v1[2], math.log(trace), fa, mode, index1, sign
+ # return v1[0], v1[1], v1[2], math.log(trace), fa, mode, index1, sign
+ return v1[0], v1[1], v1[2], trace*2, fa, mode, index1, sign
 
 
 @njit
@@ -72,8 +72,8 @@ def init_buckets(Q_TOTAL):
 
 def k_means_modeling(quantization):
 
- with open('./arrays/qua', 'rb') as p:
- quantization = pickle.load(p)
+ # with open('./arrays/qua', 'rb') as p:
+ #  quantization = pickle.load(p)
 
  kmeans_angle = KMeans_Vector(n_clusters=C.Q_ANGLE, verbose=True, max_iter=30, n_init=1)
  kmeans_angle.fit(quantization[:, :3])
@@ -84,24 +84,24 @@ def k_means_modeling(quantization):
  return kmeans_angle, kmeans_tensor
 
 
-def make_kmeans_model():
+def make_kmeans_model(file_list):
  G_WEIGHT = get_normalized_gaussian()
 
  MAX_POINTS = 15000000
  patchNumber = 0
  point_space = np.zeros((MAX_POINTS, 6))
 
- # for file_idx, file in enumerate(file_list):
- #  print('\r', end='')
- #  print('' * 60, end='')
- #  print('\r Making Point Space: '+ file.split('\\')[-1] + str(MAX_POINTS) + ' patches (' + str(100*patchNumber/MAX_POINTS) + '%)')
+ for file_idx, file in enumerate(file_list):
+ print('\r', end='')
+ print('' * 60, end='')
+ print('\r Making Point Space: '+ file.split('\\')[-1] + str(MAX_POINTS) + ' patches (' + str(100*patchNumber/MAX_POINTS) + '%)')
 
- #  im_HR, im_LR = get_array_data(file, training=True)
- #  im_GX, im_GY, im_GZ = np.gradient(im_LR)
+ im_HR, im_LR = get_array_data(file, training=True)
+ im_GX, im_GY, im_GZ = np.gradient(im_LR)
 
- #  point_space, patchNumber = make_point_space(im_LR, im_GX, im_GY, im_GZ, patchNumber, G_WEIGHT, point_space, MAX_POINTS)
- #  if patchNumber > MAX_POINTS / 2:
- #  break
+ point_space, patchNumber = make_point_space(im_LR, im_GX, im_GY, im_GZ, patchNumber, G_WEIGHT, point_space, MAX_POINTS)
+ if patchNumber > MAX_POINTS / 2:
+ break
 
  quantization = point_space[0:patchNumber, :]
 

filter_constant.py

@@ -5,7 +5,7 @@
 RESULT_DIR = "./result/"
 
 Q_ANGLE = 3#4
-Q_TENSOR = 341#170#128
+Q_TENSOR = 170#341#170#128
 
 PATCH_SIZE = 11
 PATCH_HALF = PATCH_SIZE // 2

test.py

@@ -10,7 +10,6 @@
 
 import filter_constant as C
 
-from crop_black import *
 from feature_model import *
 from filter_func import *
 from kmeans_vector import KMeans_Vector
@@ -75,8 +74,6 @@ def make_hr_yz(i1, result_image, im_LR, jS, h, iS):
  if im_LR[i1, j1, k1] == 0:
  continue
 
- # patch1 = patch.ravel()
- # print(iS[cnt][0])
  patch = np.transpose(patch, iS[cnt][0])
 
  if iS[cnt][1][0] < 0:
@@ -89,8 +86,7 @@ def make_hr_yz(i1, result_image, im_LR, jS, h, iS):
  elif iS[cnt][1][1] < 0 and iS[cnt][1][2] < 0:
  patch = np.flip(patch, axis=0)
 
- patch1 = np.append(patch, 1).astype(np.float32)
- result_image[i1, j1, k1] = np.dot(patch1, h[jS[cnt]])
+ result_image[i1, j1, k1] = (patch * h_comb[jS[cnt]]).sum() + h_bias[jS[cnt]]
  cnt += 1
 
  return result_image
@@ -109,6 +105,8 @@ def make_hr_yz(i1, result_image, im_LR, jS, h, iS):
  G_WEIGHT = get_normalized_gaussian()
 
  h = np.load('./arrays/h_{}x_{}.npy'.format(C.R, C.Q_TOTAL))
+ h_comb = h[:, :-1].reshape(h.shape[0], C.PATCH_SIZE, C.PATCH_SIZE, C.PATCH_SIZE)
+ h_bias = h[:, -1]
  kmeans = load_kmeans_model()
 
  filestart = time.time()

train.py

@@ -9,7 +9,6 @@
 
 import filter_constant as C
 
-from crop_black import *
 from feature_model import *
 from filter_func import *
 from kmeans_vector import KMeans_Vector
@@ -106,8 +105,8 @@ def train_qv(im_LR, im_HR, w, kmeans, Q, V, count):
 
  G_WEIGHT = get_normalized_gaussian()
 
- # kmeans = make_kmeans_model()
- kmeans = load_kmeans_model()
+ kmeans = make_kmeans_model(file_list)
+ # kmeans = load_kmeans_model()
 
  start = time.time()
 

util.py

@@ -3,10 +3,7 @@
 import pickle
 import nibabel as nib
 
-from crop_black import *
 from filter_func import *
-from get_lr import *
-from hashtable import *
 from matrix_compute import *
 from util import *
 
@@ -24,14 +21,6 @@ def make_dataset(dir):
  return images
 
 
-def ask_save_qv(Q, V, finished_files):
- try:
- a = input_timer("\r Enter to save >> ", 10)
- save_qv(Q, V, finished_files)
- except TimeoutError as e:
- pass
-
-
 def save_qv(Q, V, finished_files, count):
  print('\rSaving QVF...', end='', flush=True)
  np.savez('./arrays/QVF_{}'.format(C.R), Q=Q, V=V, finished_files=np.array(finished_files), count=count)