def __init__(self, dimension, dir_path):
super(SubtitleIndex, self).__init__(dimension, dir_path)
self.sentence_feature = None
self.sentence_model = TF_Sentence(
osp.join(
THIS_DIR,
'../../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73'
))
def __init__(self, db_path, image_model=None, sentence_model=None, audio_model=None):
'''
:param db_path: the path of sqlite database
:param image_model: get only one image feature to represent the scene
:param sentence_model: get the subtitle feature in this scene shot
:param audio_model: get one audio feature in this scene shot
'''
self.database = db_path
# for store the processed feature into .pkl file
self.sample_list = list()
if image_model is None:
self.image_model = scene_visual('resnet50', '../weights/places365/{}.pth',
'../weights/places365/categories.txt',
'cuda:0')
if sentence_model is None:
self.sentence_model = TF_Sentence('../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73')
if audio_model is None:
vgg_graph = tf.Graph()
with vgg_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile('../weights/audioset/vggish_fr.pb', 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
self.audio_model = AudiosetFeatureExtractor(vgg_graph, '../weights/audioset/vggish_pca_params.npz')
# detect the shot bound and get the split frame and millisecond
self.shot_detector = Shot_Detector()
# create a simple hysia database or connect to the exist database
if not osp.isfile(self.database):
statement = (
'CREATE TABLE %s (tv_name TEXT, start_time TEXT, end_time TEXT, scene_name TEXT, image_path TEXT, image_feature BLOB, subtitle TEXT, subtitle_feature BLOB, audio_path TEXT, audio_feature BLOB, object_json_path TEXT, face_json_path TEXT, insert_product_path TEXT);')
tables = ['video']
statements = [statement % table for table in tables]
self.db = SqliteDatabase(self.database, statements)
else:
self.db = SqliteDatabase(self.database)
class SubtitleIndex(BasicIndex):
'''
For senentence index
'''
def __init__(self, dimension, dir_path):
super(SubtitleIndex, self).__init__(dimension, dir_path)
self.sentence_feature = None
self.sentence_model = TF_Sentence(
osp.join(
THIS_DIR,
'../../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73'
))
def index(self, tv_name=None):
self.database = []
if tv_name is not None:
tv_file = '{}/{}_index.pkl'.format(self.dir_path, tv_name)
else:
tv_file = glob.glob('{}/*index.pkl'.format(self.dir_path))
for i in tv_file:
with open(i, 'rb') as f:
data = pickle.load(f)
self.database.extend(data)
print('Finish loading {} samples'.format(len(self.database)))
self.sentence_feature = list()
if osp.isfile('{}/sentence.pkl'.format(self.dir_path)):
print('The sentence has been encoded')
with open(self.dir_path + '/sentence.pkl', 'rb') as f:
self.sentence_feature = pickle.load(f)
else:
print('Start to encode sentence')
for i in self.database:
sentence_feature = self.sentence_model.encode(i['SUBTITLE'])
self.sentence_feature.append(sentence_feature)
self.sentence_feature = np.array(self.sentence_feature)
with open(self.dir_path + '/sentence.pkl', 'wb') as f:
pickle.dump(self.sentence_feature, f)
# TypeError: ndarrays must be of numpy.float32, and not float64.
self.gpu_index_flat.add(self.sentence_feature.astype(np.float32))
print('Finish loading {} sentence features'.format(
self.gpu_index_flat.ntotal))
scene_model = scene_visual('resnet50', '../weights/places365/{}.pth',
'../weights/places365/categories.txt', 'cuda:0')
temp = cv2.imread('test_sofa2.jpg')
q_tensor = Image.fromarray(cv2.cvtColor(temp, cv2.COLOR_BGR2RGB))
q_vec = scene_model.extract_vec(q_tensor, True)
print(q_vec.shape)
gpu_machine = BasicIndex.init_size(VIDEO_DATA_PATH)
gpu_machine.index()
tic = time.time()
results = gpu_machine.search(q_vec, 10)
toc = time.time()
print("It takes " + str((toc - tic)) + " ms")
sentence_model = TF_Sentence(
'../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73')
product_description_vector = sentence_model.encode(
'Nike sports shirts is very good.')
sentence_machine = SubtitleIndex(512, VIDEO_DATA_PATH)
sentence_machine.index()
tic = time.time()
results = sentence_machine.search(
np.array(product_description_vector).astype(np.float32), 30)
toc = time.time()
print("It takes " + str((toc - tic)) + " ms")
print(results)
# show results
for i in results:
class BasicSearch(object):
'''
This is the product-content match module.
'''
def __init__(self, database):
'''
:param database: the directory path
'''
try:
self.database = database
self.image_machine = BasicIndex.init_size(self.database)
self.sentence_machine = SubtitleIndex(512, self.database)
self.image_machine.index()
print('Finish loading image GPU indexer')
self.sentence_machine.index()
print('Finish loading sentence GPU indexer')
except ValueError:
print("Please specify the pkl directory!")
# TODO load scene/imagenet/faster-rcnn model to extract query image feature.
self.scene_model = scene_visual(
'resnet50', os.path.join(THIS_DIR,
'../../weights/places365/{}.pth'),
os.path.join(THIS_DIR, '../../weights/places365/categories.txt'),
'cuda:0')
self.sentence_model = TF_Sentence(
os.path.join(
THIS_DIR,
'../../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73'
))
def search(self,
image_query,
subtitle_query=None,
audio_query=None,
face_query=None,
topK=30):
'''
:param image_query:
:param subtitle_query:
:param audio_query:
:param face_query:
:return:
'''
# search subtitle first.
if (image_query is not None) and (subtitle_query is not None):
product_description_vector = self.sentence_model.encode(
subtitle_query)
tic = time.time()
_, idx, _ = self.sentence_machine.search(
np.array(product_description_vector).astype(np.float32), topK)
toc = time.time()
print("Searching subtitle takes " + str((toc - tic)) + " ms")
temp = cv2.imread(image_query)
q_tensor = Image.fromarray(cv2.cvtColor(temp, cv2.COLOR_BGR2RGB))
q_vec = self.scene_model.extract_vec(q_tensor, True)
print(q_vec.shape)
tic = time.time()
self.image_machine.re_index(idx)
results, _, _ = self.image_machine.search(q_vec, topK, 2)
toc = time.time()
print("Searching image takes " + str((toc - tic)) + " ms")
elif subtitle_query is not None:
product_description_vector = self.sentence_model.encode(
subtitle_query)
tic = time.time()
results, _, _ = self.sentence_machine.search(
np.array(product_description_vector).astype(np.float32), topK)
toc = time.time()
print("Searching subtitle takes " + str((toc - tic)) + " ms")
elif image_query is not None:
temp = cv2.imread(image_query)
q_tensor = Image.fromarray(cv2.cvtColor(temp, cv2.COLOR_BGR2RGB))
q_vec = self.scene_model.extract_vec(q_tensor, True)
print(q_vec.shape)
tic = time.time()
results, _, _ = self.image_machine.search(q_vec, topK)
toc = time.time()
print("Searching image takes " + str((toc - tic)) + " ms")
else:
results = []
return results
class DatabasePklSearch(BasicSearch):
'''
This is for pkl search and then connect to database to do more operations.
'''
def __init__(self, database):
'''
:param database: the path contains many .pkl file
'''
try:
self.database = database
self.image_machine = BasicIndex(2048, self.database)
print('Finish loading scene-image GPU indexer')
self.sentence_machine = BasicIndex(512, self.database)
print('Finish loading scene-sentence GPU indexer')
except ValueError:
print("Please specify the pkl directory!")
# TODO load scene/imagenet/faster-rcnn model to extract query image feature.
self.scene_model = scene_visual(
'resnet50', os.path.join(THIS_DIR,
'../../weights/places365/{}.pth'),
os.path.join(THIS_DIR, '../../weights/places365/categories.txt'),
'cuda:0')
self.sentence_model = TF_Sentence(
os.path.join(
THIS_DIR,
'../../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73'
))
def search(self,
image_query,
subtitle_query=None,
audio_query=None,
face_query=None,
topK=30,
tv_name=None):
'''
:param image_query:
:param subtitle_query:
:param audio_query:
:param face_query:
:return:
'''
if tv_name is not None:
self.image_machine.index(tv_name=tv_name)
self.sentence_machine.index(feature='SUBTITLE_FEATURE',
tv_name=tv_name)
else:
self.image_machine.index()
self.sentence_machine.index(feature='SUBTITLE_FEATURE')
# search subtitle first.
if (image_query is not None) and (subtitle_query is not None):
product_description_vector = self.sentence_model.encode(
subtitle_query)
tic = time.time()
_, idx, _ = self.sentence_machine.search(
np.array(product_description_vector).astype(np.float32), topK)
toc = time.time()
print("Searching subtitle takes " + str((toc - tic)) + " ms")
temp = cv2.imread(image_query)
q_tensor = Image.fromarray(cv2.cvtColor(temp, cv2.COLOR_BGR2RGB))
q_vec = self.scene_model.extract_vec(q_tensor, True)
# print(q_vec.shape)
tic = time.time()
self.image_machine.re_index(idx)
results, _, _ = self.image_machine.search(q_vec, topK, 2)
toc = time.time()
print("Searching image takes " + str((toc - tic)) + " ms")
elif subtitle_query is not None:
product_description_vector = self.sentence_model.encode(
subtitle_query)
tic = time.time()
results, _, _ = self.sentence_machine.search(
np.array(product_description_vector).astype(np.float32), topK)
toc = time.time()
print("Searching subtitle takes " + str((toc - tic)) + " ms")
elif image_query is not None:
temp = cv2.imread(image_query)
q_tensor = Image.fromarray(cv2.cvtColor(temp, cv2.COLOR_BGR2RGB))
q_vec = self.scene_model.extract_vec(q_tensor, True)
# print(q_vec.shape)
tic = time.time()
results, _, _ = self.image_machine.search(q_vec, topK)
toc = time.time()
print("Searching image takes " + str((toc - tic)) + " ms")
else:
results = []
# TODO better reset design.
self.sentence_machine.reset_GPU()
print("Reset sentence GPU indexer")
self.image_machine.reset_GPU()
print("Reset product image GPU indexer")
return results
def add_pkl(self):
'''
Every time we finish processing one video, we should add a pkl file
:return:
'''
pass
if __name__ == '__main__':
word = "Elephant"
sentence = "I am a sentence for which I would like to get its embedding."
paragraph = (
"Universal Sentence Encoder embeddings also support short paragraphs. "
"There is no hard limit on how long the paragraph is. Roughly, the longer "
"the more 'diluted' the embedding will be.")
messages = [word, sentence, paragraph]
sentences = [
# Smartphones
"smartphone",
"I bought a new phone yesterday.",
# Clothes
"Nike produces a new sports shirt.",
"Your new sports shirt looks great.",
# Food and health
"Strawberry ice-cream",
"Eating strawberries is healthy"
]
m = TF_Sentence(
'../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73')
for i in messages:
m.encode(i)
m.run_and_plot(sentences)
# Author: Zhou Shengsheng
# Date: 15-02-19
from __future__ import print_function
import rpc
import os
import sys
import cloudpickle
import pickle
import base64
from models.nlp.sentence import TF_Sentence
IMPORT_ERROR_RETURN_CODE = 3
tf_sentence = TF_Sentence('/hysia-deps/weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73')
print('tf_sentence:', tf_sentence)
def predict(serialized_input_bytes):
global tf_sentence
sentence = pickle.loads(serialized_input_bytes)
sentence_embedding = tf_sentence.encode(sentence)
serialized_pred = pickle.dumps(sentence_embedding)
base64_pred = base64.b64encode(serialized_pred).decode()
return [base64_pred]
def load_predict_func(file_path):
if sys.version_info < (3, 0):
with open(file_path, 'r') as serialized_func_file:
return cloudpickle.load(serialized_func_file)
class VideoPostDB(object):
'''
This class is for video post_processing.
We insert some key information of processed videos to a small database and index it for inserting product.
'''
def __init__(self, db_path, image_model=None, sentence_model=None, audio_model=None):
'''
:param db_path: the path of sqlite database
:param image_model: get only one image feature to represent the scene
:param sentence_model: get the subtitle feature in this scene shot
:param audio_model: get one audio feature in this scene shot
'''
self.database = db_path
# for store the processed feature into .pkl file
self.sample_list = list()
if image_model is None:
self.image_model = scene_visual('resnet50', '../weights/places365/{}.pth',
'../weights/places365/categories.txt',
'cuda:0')
if sentence_model is None:
self.sentence_model = TF_Sentence('../weights/sentence/96e8f1d3d4d90ce86b2db128249eb8143a91db73')
if audio_model is None:
vgg_graph = tf.Graph()
with vgg_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile('../weights/audioset/vggish_fr.pb', 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
self.audio_model = AudiosetFeatureExtractor(vgg_graph, '../weights/audioset/vggish_pca_params.npz')
# detect the shot bound and get the split frame and millisecond
self.shot_detector = Shot_Detector()
# create a simple hysia database or connect to the exist database
if not osp.isfile(self.database):
statement = (
'CREATE TABLE %s (tv_name TEXT, start_time TEXT, end_time TEXT, scene_name TEXT, image_path TEXT, image_feature BLOB, subtitle TEXT, subtitle_feature BLOB, audio_path TEXT, audio_feature BLOB, object_json_path TEXT, face_json_path TEXT, insert_product_path TEXT);')
tables = ['video']
statements = [statement % table for table in tables]
self.db = SqliteDatabase(self.database, statements)
else:
self.db = SqliteDatabase(self.database)
def process(self, video_path):
'''
:param video_path: a video path which can also help to find the subtitle and audio
:param features_path: store the features into a pickle file
:return:
'''
# read video and process video, subtitle and audio information
# save features into the database
# TODO This part can be optimized during video content analysis stage in the 0.2 version
video_name = video_path.split('/')[-1]
splits_frame, splits_ms = self.shot_detector.detect(video_path)
# print(splits_ms)
# get middle frame as the key frame in this shot
# package the split time
middle_frame = list()
middle_time = {}
for i in range(len(splits_ms) - 1):
temp = math.floor((splits_frame[i] + splits_frame[i + 1]) / 2.0)
middle_frame.append(temp)
middle_time[temp] = [splits_ms[i], splits_ms[i + 1]]
# print(middle_frame)
# print(middle_time)
vid = imageio.get_reader(video_path, 'ffmpeg')
frame_cnt = vid.get_meta_data()["nframes"]
# get key frame image features, subtitle features and audio features
try:
with tqdm.tqdm(total=frame_cnt, unit="frames") as pbar:
for id, img in enumerate(vid):
if id in middle_frame:
# TODO Try to classify image_io, cv.read, PIL
# get image feature
try:
scene_name, scene_feature = self.__get_scene_feature(img)
except:
# print('Can not extract image feature and assume it is none')
scene_feature = 'unknown_feature'
scene_name = 'unknown_scene'
try:
subtitle, subtitle_feature = self.__get_subtitle_feature(middle_time[id][0], middle_time[id][1],
video_path)
except:
# print('Can not extract subtitle feature and assume it is none')
subtitle_feature = 'unknown_feature'
subtitle = 'unknown_subtitle'
try:
audio_path, audio_feature = self.__get_audio_feature(middle_time[id][0], middle_time[id][1],
video_path)
except:
# print('Can not extract audio feature and assume it is none')
audio_feature = 'unknown_feature'
audio_path = 'unknown_audio_path'
# TODO get object json file and face json file
sql = "INSERT INTO video (tv_name, start_time, end_time, scene_name, image_feature, subtitle, subtitle_feature, audio_path, audio_feature) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)"
sample = (
video_name, middle_time[id][0], middle_time[id][1], scene_name, scene_feature, subtitle,
subtitle_feature, audio_path, audio_feature)
self.db.add(sql, sample)
# pack
self.__pack_samples(sample)
pbar.update(1)
except:
print('Throw the last frames')
self.__insert_index(video_path)
def __get_scene_feature(self, img):
img_pil = Image.fromarray(img)
scene_feature = self.image_model.extract_vec(img_pil, True)
scene_name = self.image_model.detect(img_pil, True)
# only return top 1 scene name
return scene_name['scene'][0], scene_feature
def __get_subtitle_feature(self, start_time, end_time, video_path):
srt_name = video_path.split("/")[-1].split(".")[0] + ".srt"
srt_path = osp.join(osp.dirname(osp.abspath(video_path)), 'subtitles', srt_name)
sentences = extract_srt(start_time, end_time, srt_path)
if len(sentences) == 0:
sentences_feature = 'unknown_feature'
sentences = 'unknown_subtitle'
else:
# TODO TEXT support what data types (BLOB only support numpy)
sentences = " ".join(sentences)
sentences_feature = self.sentence_model.encode(sentences)
return sentences, np.array(sentences_feature)
def __get_audio_feature(self, start_time, end_time, video_path):
audio_name = video_path.split("/")[-1].split(".")[0] + ".wav"
audio_path = osp.join(osp.dirname(osp.abspath(video_path)), 'audios', audio_name)
# command = "ffmpeg -i %s -ab 160k -ac 2 -ar 44100 -vn %s" % (video_path, audio_path)
# subprocess.call(command, shell=True)
audio_feature = self.audio_model.extract(audio_path, start_time, end_time)[0]
return audio_path, audio_feature
def __get_object_json(self):
pass
def __get_frame_path(self):
pass
def __get_face_json(self):
pass
def __pack_samples(self, sample):
sample_dict = {'TV_NAME': sample[0],
'START_TIME': sample[1],
'END_TIME': sample[2],
'SCENE': sample[3],
'FEATURE': sample[4],
'SUBTITLE': sample[5],
'SUBTITLE_FEATURE': sample[6],
'AUDIO': sample[7],
'AUDIO_FEATURE': sample[8]
}
self.sample_list.append(sample_dict)
def __insert_index(self, video_path):
pkl_name = video_path.split("/")[-1].split(".")[0] + "_index.pkl"
pkl_path = osp.join(osp.dirname(osp.abspath(video_path)), 'multi_features', pkl_name)
with open(pkl_path, 'wb') as f:
pickle.dump(self.sample_list, f)