def infer_2(network):
valid_dict = infer.build_valid_dict()
feature_matrix, row_index_to_id = build_train_feature_matrix_mean(network)
feature_matrix = feature_matrix.T
total = 0
hit = 0
print(infer.VALID_NPZ_PATH)
for npz in glob.glob(infer.VALID_NPZ_PATH):
head, tail = os.path.split(npz)
mp4_filename, npz_ext = os.path.splitext(tail)
print(mp4_filename)
features = np.load(npz)
features = features.f.arr_0
features = network.predict(features)
assert (len(features.shape) == 2)
sim = np.dot(features, feature_matrix)
sim = np.sum(sim, axis=0)
index = np.argmax(sim)
predict_id = row_index_to_id[index][0]
ground_truth = valid_dict[mp4_filename]
if predict_id == ground_truth:
hit += 1
total += 1
print('hit: {0}, total: {1}, precision: {2}'.format(
hit, total, hit / total))
def infer_2():
valid_dict = infer.build_valid_dict()
feature_matrix, row_index_to_id, feature_group = build_train_feature_matrix_mean(
)
feature_matrix = feature_matrix.T
total = 0
hit = 0
total_2 = 0
hit_2 = 0
misclassified = {}
for npz in glob.glob(infer.VALID_NPZ_PATH):
head, tail = os.path.split(npz)
mp4_filename, npz_ext = os.path.splitext(tail)
features = np.load(npz)
features = features.f.arr_0
assert (len(features.shape) == 2)
sim = np.dot(features, feature_matrix)
sim = np.sum(sim, axis=0)
index = np.argmax(sim)
predict_id = row_index_to_id[index][0]
ground_truth = valid_dict[mp4_filename]
if predict_id == ground_truth:
hit += 1
else:
misclassified[mp4_filename] = (ground_truth, predict_id)
total += 1
#
sim_2 = []
for feature in feature_group:
class_sim = np.mean(np.dot(features, feature.T), axis=0)
sim_2.append(np.max(class_sim))
index = np.argmax(np.array(sim_2))
predict_id = row_index_to_id[index][0]
ground_truth = valid_dict[mp4_filename]
if predict_id == ground_truth:
hit_2 += 1
#else:
# misclassified[mp4_filename] = (ground_truth, predict_id)
total_2 += 1
print('hit: {0}, total: {1}, precision: {2}'.format(
hit, total, hit / total))
print('hit_2: {0}, total_2: {1}, precision: {2}'.format(
hit_2, total_2, hit_2 / total_2))
def calc_mean_average_precision(network):
feature_matrix, row_index_to_id = build_train_feature_matrix_mean(network)
print(len(row_index_to_id))
feature_matrix = feature_matrix.T
row_to_filename = []
sim_list = []
for npz in glob.glob(infer.VALID_NPZ_PATH):
head, tail = os.path.split(npz)
mp4_filename, npz_ext = os.path.splitext(tail)
row_to_filename.append(mp4_filename)
features = np.load(npz)
features = features.f.arr_0
features = network.predict(features)
# sim[i] == class_i cos similarity
sim = np.mean(np.dot(features, feature_matrix), axis=0)
sim_list.append([sim])
# sim_matrix[i, j] == file_i cos similarity to class_j
sim_matrix = np.concatenate(tuple(sim_list), axis=0)
print("sim matrix shape: ", sim_matrix.shape)
print("sim matrix max:{0}, min:{1}".format(sim_matrix.max(),
sim_matrix.min()))
# calc class_i average precision
filename_to_id = infer.build_valid_dict()
class_to_ap = {}
valid_id_to_filenames = build_valid_id_to_filenames()
for i in range(sim_matrix.shape[1]):
class_id, _ = row_index_to_id[i]
gt_count = len(valid_id_to_filenames[class_id])
if gt_count > 100:
gt_count = 100
confidence = sim_matrix[:, i]
def index_to_class_id(index):
filename = row_to_filename[index]
id = filename_to_id[filename]
return id
ap = calc_ap(class_id, confidence, gt_count, index_to_class_id)
class_to_ap[class_id] = ap
# print(class_to_ap[class_id])
mean_ap = 0.0
for ap in class_to_ap.values():
mean_ap += ap
mean_ap = mean_ap / len(class_to_ap)
return mean_ap
def calc_map(row_to_filename, row_index_to_id, sim_matrix):
filename_to_id = infer.build_valid_dict()
class_to_ap = {}
valid_id_to_filenames = build_valid_id_to_filenames()
for i in range(sim_matrix.shape[1]):
class_id, _ = row_index_to_id[i]
gt_count = len(valid_id_to_filenames[class_id])
if gt_count > 100:
gt_count = 100
confidence = sim_matrix[:, i]
def index_to_class_id(index):
filename = row_to_filename[index]
id = filename_to_id[filename]
return id
ap = calc_ap(class_id, confidence, gt_count, index_to_class_id)
class_to_ap[class_id] = ap
# print(class_to_ap[class_id])
mean_ap = 0.0
for ap in class_to_ap.values():
mean_ap += ap
mean_ap = mean_ap / len(class_to_ap)
return mean_ap