def generate_representor(data_dir, dictionary_path, subclass):
vocabulary_dict = load_vocabulary(dictionary_path)
shape_vocabulary = np.shape(vocabulary_dict[0])
vocabulary_size = shape_vocabulary[0]
representers = []
patches, coding_labeles, labeles = extract_patches_multidir(
data_dir, subclasses=[subclass], return_flag=True)
for case_index, cur_patches in enumerate(patches):
cur_case_representor = np.zeros([10, vocabulary_size])
patches_coding_labeles = {}
for patch_index, cur_patch in enumerate(cur_patches):
cur_coding_label = coding_labeles[case_index][patch_index]
if cur_coding_label not in patches_coding_labeles.keys():
patches_coding_labeles[cur_coding_label] = []
patches_coding_labeles[cur_coding_label].append(cur_patch)
for key in patches_coding_labeles.keys():
cur_patches_coding_label = patches_coding_labeles[key]
cur_vocabulary = vocabulary_dict[key]
distance_arr = cal_distance(cur_patches_coding_label,
cur_vocabulary)
for i in range(len(distance_arr)):
min_index = np.argmin(distance_arr[i])
cur_case_representor[int(key), min_index] += 1
representers.append(cur_case_representor.flatten())
return representers, labeles
def generate_representor_version2(data_dir, dictionary_path, subclass, vectorizer=None):
save_dir = '/home/give/PycharmProjects/ICPR2018/LeaningBased/BoVW-NGram/patches'
save_dir = os.path.join(save_dir, subclass)
paths = glob(os.path.join(save_dir, '*.npy'))
kmeans_model = joblib.load(dictionary_path)
shape_vocabulary = np.shape(kmeans_model.cluster_centers_)
vocabulary_size = shape_vocabulary[0]
if len(paths) == 0:
patches, _, labeles = extract_patches_multidir(data_dir, subclasses=[subclass], return_flag=True,
patch_size=3)
else:
patches = []
labeles = []
for path in paths:
patches.append(np.load(path))
labeles.append(int(os.path.basename(path).split('.npy')[0].split('_')[1]))
all_patches = []
counts = []
for case_index, cur_patches in enumerate(patches):
# print np.shape(cur_patches)
if len(paths) == 0:
np.save(os.path.join(save_dir, str(case_index) + '_' + str(labeles[case_index])), cur_patches)
all_patches.extend(cur_patches)
counts.append(len(cur_patches))
print 'all patches shape are ', np.shape(all_patches)
predicted_labels = kmeans_model.predict(all_patches)
start = 0
strs = convert2str(predicted_labels, counts)
if vectorizer is None:
vectorizer = TfidfVectorizer(analyzer='char', min_df=1, ngram_range=(1, 3), use_idf=False, stop_words=None)
vectorizer = vectorizer.fit(strs)
crs_matrix = vectorizer.transform(strs).toarray()
return np.asarray(crs_matrix, np.float32), labeles, vectorizer
def generate_representor(data_dir, dictionary_path, subclass, phase_name):
dictionary = load_vocabulary(dictionary_path)
shape_vocabulary = np.shape(dictionary)
vocabulary_size = shape_vocabulary[0]
representers = []
patches, coding_labeles, labeles = extract_patches_multidir(
data_dir,
subclasses=[subclass],
return_flag=True,
phase_name=phase_name)
all_patches = []
counts = []
for case_index, cur_patches in enumerate(patches):
print np.shape(cur_patches)
all_patches.extend(cur_patches)
counts.append(len(cur_patches))
all_distance_arr = cal_distance(all_patches, dictionary)
start = 0
for case_index, count in enumerate(counts):
distance_arr = all_distance_arr[start:start + count]
cur_case_representor = np.zeros([1, vocabulary_size])
for i in range(len(distance_arr)):
min_index = np.argmin(distance_arr[i])
cur_case_representor[0, min_index] += 1
representers.append(cur_case_representor.squeeze())
start += count
return representers, labeles
def generate_representor_version2(data_dir, dictionary_path, subclass):
kmeans_model = joblib.load(dictionary_path)
shape_vocabulary = np.shape(kmeans_model.cluster_centers_)
vocabulary_size = shape_vocabulary[0]
representers = []
patches, coding_labeles, labeles = extract_patches_multidir(data_dir, subclasses=[subclass], return_flag=True)
all_patches = []
counts = []
for case_index, cur_patches in enumerate(patches):
print np.shape(cur_patches)
all_patches.extend(cur_patches)
counts.append(len(cur_patches))
predicted_labels = kmeans_model.predict(all_patches)
start = 0
for case_index, count in enumerate(counts):
cur_predicted_label = predicted_labels[start: start + count]
representer = np.histogram(cur_predicted_label, bins=vocabulary_size, normed=True)[0]
representers.append(np.array(representer).squeeze())
start += count
return representers, labeles
def generate_representor(data_dir, dictionary_path, subclass):
dictionary = load_vocabulary(dictionary_path)
shape_vocabulary = np.shape(dictionary)
vocabulary_size = shape_vocabulary[0]
representers = []
patches, coding_labeles, labeles = extract_patches_multidir(
data_dir, subclasses=[subclass], return_flag=True)
all_patches = []
counts = []
for case_index, cur_patches in enumerate(patches):
print np.shape(cur_patches)
all_patches.extend(cur_patches)
counts.append(len(cur_patches))
all_distance_arr = cal_distance(all_patches, dictionary)
start = 0
for case_index, count in enumerate(counts):
distance_arr = all_distance_arr[start:start + count]
cur_case_representor = np.zeros([1, vocabulary_size])
for i in range(len(distance_arr)):
min_index = np.argmin(distance_arr[i])
cur_case_representor[0, min_index] += 1
representers.append(cur_case_representor.squeeze())
start += count
# patches_coding_labeles = {}
# for patch_index, cur_patch in enumerate(cur_patches):
# cur_coding_label = coding_labeles[case_index][patch_index]
# if cur_coding_label not in patches_coding_labeles.keys():
# patches_coding_labeles[cur_coding_label] = []
# patches_coding_labeles[cur_coding_label].append(cur_patch)
# for key in patches_coding_labeles.keys():
# cur_patches_coding_label = patches_coding_labeles[key]
# cur_vocabulary = vocabulary_dict[key]
# distance_arr = cal_distance(cur_patches_coding_label, cur_vocabulary)
# for i in range(len(distance_arr)):
# min_index = np.argmin(distance_arr[i])
# cur_case_representor[int(key), min_index] += 1
# representers.append(cur_case_representor.flatten())
return representers, labeles