def GetFeature(self, dataset):
if os.path.exists(self.SESSION_SAVE_PATH + '.meta'):
with tf.Session(config=self._config) as sess:
self.InitVariables()
self._saver = tf.train.Saver()
self._saver.restore(sess, self.SESSION_SAVE_PATH)
PrintWithTime("Restored model from " + self.SESSION_SAVE_PATH)
database = Object()
dim = self.NetPQ.X.get_shape().as_list()[1]
Nb = dataset.DataNum
database_feature = np.zeros([Nb, dim], dtype=np.float32)
database.label = np.zeros([Nb, self._classNum], dtype=np.int16)
codes = np.zeros([Nb, self._stackLevel], np.int32)
total_db = (Nb // self._batchSize) + 1
with trange(total_db, ncols=50) as t:
for i in t:
idx = np.arange(start=i * self._batchSize,
stop=np.minimum(
Nb, (i + 1) * self._batchSize),
step=1)
inp, label = dataset.Get(idx)
num = inp.shape[0]
database.label[i *
self._batchSize:(i * self._batchSize +
num)] = label
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
out, hardCode = sess.run(
[self.NetPQ.X, self.NetPQ.HardCode],
{self.Input: inp})
hardCode = sess.run(self.NetPQ.HardCode,
{self.Input: inp})
codes[i * self._batchSize:(i * self._batchSize) +
num] = np.array(hardCode, np.int32).T[:num]
database_feature[i *
self._batchSize:(i *
self._batchSize) +
num] = out[:num]
database.output = database_feature
# [N, M]
database.codes = codes
codebook = sess.run(self.NetPQ.Codebook)
scale = sess.run(self.NetPQ.CodebookScale)
return database, codebook, scale
def GetRetrievalMat(self, queryX, queryY, dataset):
self.R = self._recallatR if self._recallatR > 0 else dataset.DataNum
if os.path.exists(SESSION_SAVE_PATH + '.meta'):
with tf.Session(config=self._config) as sess:
self.InitVariables()
self._saver = tf.train.Saver()
self._saver.restore(sess, SESSION_SAVE_PATH)
PrintWithTime("Restored model from " + SESSION_SAVE_PATH)
query = Object()
database = Object()
query.label = queryY
Nq = queryX.shape[0]
dim = self.NetPQ.X.get_shape().as_list()[1]
query_feature = np.zeros([Nq, dim], np.float16)
for i in range((Nq // self._batchSize) + 1):
inp = queryX[i * self._batchSize:(i + 1) * self._batchSize]
num = inp.shape[0]
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
out = sess.run(self.NetPQ.X, {self.Input: inp})
query_feature[i * self._batchSize:(i * self._batchSize) +
num] = out[:num]
query.output = query_feature
Nb = dataset.DataNum
database.label = np.zeros([Nb, self._classNum], dtype=np.int16)
codes = np.zeros([Nb, self._stackLevel], np.int32)
total_db = (Nb // self._batchSize) + 1
for i in range(total_db):
idx = np.arange(start=i * self._batchSize,
stop=np.minimum(Nb,
(i + 1) * self._batchSize),
step=1)
inp, label = dataset.Get(idx)
print(inp.shape, label.shape)
num = inp.shape[0]
database.label[i * self._batchSize:(i * self._batchSize +
num)] = label
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
hardCode = sess.run(self.NetPQ.HardCode, {self.Input: inp})
codes[i * self._batchSize:(i * self._batchSize) +
num] = np.array(hardCode, np.int32).T[:num]
ProgressBar((i + 1) / total_db)
# [N, M]
database.codes = codes
codebook = sess.run(self.NetPQ.Codebook)
# np.save('database_codes_DSQ', codes)
db = mAP.Quantize_AQ(database.codes, codebook, 4).T
del dataset
id_all = np.zeros([query.output.shape[0], self.R], np.int)
retrieval_mat = np.zeros([query.output.shape[0], self.R], np.bool)
for j in range(query.output.shape[0] // 50 + 1):
q = query.output[j * 50:(j + 1) * 50]
d = -np.dot(q, db)
ids = np.argsort(d, 1)
for i in range(d.shape[0]):
label = query.label[j * 50 + i, :]
label[label == 0] = -1
idx = ids[i, :]
imatch = np.sum(database.label[idx[0:self.R], :] == label,
1) > 0
id_all[j * 50 + i] = idx[:self.R]
retrieval_mat[j * 50 + i] = imatch[:self.R]
np.save('retrievalMat_' + self.DatasetName, retrieval_mat)
np.save('ids', id_all)
return retrieval_mat, id_all
def Evaluate(self, queryX, queryY, dataset):
print(self._recallatR if self._recallatR > 0 else 'all')
if os.path.exists(SESSION_SAVE_PATH + '.meta'):
with tf.Session(config=self._config) as sess:
self.InitVariables()
self._saver = tf.train.Saver()
self._saver.restore(sess, SESSION_SAVE_PATH)
PrintWithTime("Restored model from " + SESSION_SAVE_PATH)
query = Object()
database = Object()
query.label = queryY
Nq = queryX.shape[0]
dim = self.NetPQ.X.get_shape().as_list()[1]
query.output = np.zeros([Nq, dim], np.float16)
for i in range((Nq // self._batchSize) + 1):
inp = queryX[i * self._batchSize:(i + 1) * self._batchSize]
num = inp.shape[0]
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
out = sess.run(self.NetPQ.X, {self.Input: inp})
query.output[i * self._batchSize:(i * self._batchSize) +
num] = out[:num]
Nb = dataset.DataNum
database_feature = np.zeros([Nb, dim], dtype=np.float16)
database.label = np.zeros([Nb, self._classNum], dtype=np.int16)
database.codes = np.zeros([Nb, self._stackLevel], np.int32)
start = time.time()
print('Encoding database')
total_db = (Nb // self._batchSize) + 1
for i in range(total_db):
idx = np.arange(start=i * self._batchSize,
stop=np.minimum(Nb,
(i + 1) * self._batchSize),
step=1)
inp, label = dataset.Get(idx)
print(inp.shape, label.shape)
num = inp.shape[0]
database.label[i * self._batchSize:(i * self._batchSize +
num)] = label
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
hardCode = sess.run(self.NetPQ.HardCode, {self.Input: inp})
database.codes[i * self._batchSize:(i * self._batchSize) +
num] = np.array(hardCode, np.int32).T[:num]
database_feature[i *
self._batchSize:(i * self._batchSize) +
num] = out[:num]
ProgressBar((i + 1) / total_db)
end = time.time()
print('Encoding Complete')
print('Time:', end - start)
print('Average time for single sample:')
print((end - start) / Nb)
database.output = database_feature
del dataset
codebook = sess.run(self.NetPQ.Codebook)
res = mAP(
codebook, self._recallatR if self._recallatR > 0 else
database.codes.shape[0], database)
return res.AQD_mAP(query)
def GetFeature(self, queryX, queryY, dataset):
if os.path.exists(self.SESSION_SAVE_PATH + '.meta'):
with tf.Session(config=self._config) as sess:
self.InitVariables()
self._saver = tf.train.Saver()
self._saver.restore(sess, self.SESSION_SAVE_PATH)
PrintWithTime("Restored model from " + self.SESSION_SAVE_PATH)
query = Object()
database = Object()
query.label = queryY
Nq = queryX.shape[0]
dim = self.NetPQ.X.get_shape().as_list()[1]
query_feature = np.zeros([Nq, dim], np.float32)
for i in range((Nq // self._batchSize) + 1):
inp = queryX[i * self._batchSize:(i + 1) * self._batchSize]
num = inp.shape[0]
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
out = sess.run(self.NetPQ.X, {self.Input: inp})
query_feature[i * self._batchSize:(i * self._batchSize) +
num] = out[:num]
query.output = query_feature
Nb = dataset.DataNum
database_feature = np.zeros([Nb, dim], dtype=np.float32)
database.label = np.zeros([Nb, self._classNum], dtype=np.int16)
codes = np.zeros([Nb, self._stackLevel], np.int32)
start = time.time()
print('Encoding database')
total_db = (Nb // self._batchSize) + 1
with trange(total_db, ncols=50) as t:
for i in t:
idx = np.arange(start=i * self._batchSize,
stop=np.minimum(
Nb, (i + 1) * self._batchSize),
step=1)
inp, label = dataset.Get(idx)
num = inp.shape[0]
database.label[i *
self._batchSize:(i * self._batchSize +
num)] = label
if inp.shape[0] != self._batchSize:
placeholder = np.zeros([
self._batchSize - inp.shape[0], inp.shape[1],
inp.shape[2], inp.shape[3]
])
inp = np.concatenate((inp, placeholder))
out, hardCode = sess.run(
[self.NetPQ.X, self.NetPQ.HardCode],
{self.Input: inp})
hardCode = sess.run(self.NetPQ.HardCode,
{self.Input: inp})
codes[i * self._batchSize:(i * self._batchSize) +
num] = np.array(hardCode, np.int32).T[:num]
database_feature[i *
self._batchSize:(i *
self._batchSize) +
num] = out[:num]
end = time.time()
print('Encoding Complete')
print('Time:', end - start)
print('Average time for single sample:')
print((end - start) / Nb)
database.output = database_feature
scale = sess.run(self.NetPQ.CodebookScale)
# [N, M]
database.codes = codes
return query, database, scale
