def split(self, train_ratio, valid_ratio):
"""Split all the samples into training, validation and testing sets
randomly
Args:
train_ratio: The ratio between training and all the samples
valid_ratio: The ratio between validation and all the samples
"""
m = self.X.shape[0]
m_train = int(m * train_ratio)
m_valid = int(m * valid_ratio)
p = numpy_rng.permutation(m)
self.train_ind = p[0:m_train]
self.valid_ind = p[m_train:m_train + m_valid]
self.test_ind = p[m_train + m_valid:]
self.train_x = self._create_shared(self.X[self.train_ind, :])
self.train_y = self._create_shared(self.Y[self.train_ind, :])
self.valid_x = self._create_shared(self.X[self.valid_ind, :])
self.valid_y = self._create_shared(self.Y[self.valid_ind, :])
self.test_x = self._create_shared(self.X[self.test_ind, :])
self.test_y = self._create_shared(self.Y[self.test_ind, :])
def split(self, train_ratio, valid_ratio):
"""Split all the samples into training, validation and testing sets
randomly
Args:
train_ratio: The ratio between training and all the samples
valid_ratio: The ratio between validation and all the samples
"""
m = self.X.shape[0]
m_train = int(m * train_ratio)
m_valid = int(m * valid_ratio)
p = numpy_rng.permutation(m)
self.train_ind = p[0 : m_train]
self.valid_ind = p[m_train : m_train+m_valid]
self.test_ind = p[m_train+m_valid : ]
self.train_x = self._create_shared(self.X[self.train_ind, :])
self.train_y = self._create_shared(self.Y[self.train_ind, :])
self.valid_x = self._create_shared(self.X[self.valid_ind, :])
self.valid_y = self._create_shared(self.Y[self.valid_ind, :])
self.test_x = self._create_shared(self.X[self.test_ind, :])
self.test_y = self._create_shared(self.Y[self.test_ind, :])
def sample(indices, pos_downsample=1.0, neg_pos_ratio=1.0):
"""Sample positive and negative data
Args:
indices: [(pid, vid)]
neg_pos_ratio: #neg / #pos
Returns:
The sampled data indices. (train, vaid, test) each is [(i, j, 0/1)].
"""
print "Sampling ..."
data = cachem.load('sample')
if data is not None: return data
import random
from reid.utils.math_utils import numpy_rng
n_imgs = len(indices)
def gensamples(pids):
samples = []
# Positive samples
for i in xrange(n_imgs):
if indices[i][0] not in pids: continue
j = i + 1
while j < n_imgs and indices[i][0] == indices[j][0]:
if numpy_rng.rand() < pos_downsample:
samples.append((i, j, True))
j += 1
# Negative samples
n = int(len(samples) * neg_pos_ratio)
for k in xrange(n):
while True:
i, j = numpy_rng.randint(0, n_imgs, 2)
if indices[i][0] in pids and indices[j][0] in pids and \
indices[i][0] != indices[j][0] and \
(i,j,False) not in samples and (j,i,False) not in samples:
samples.append((i, j, False))
break
random.shuffle(samples)
return samples
# Split by pid
m = indices[-1][0] + 1
m_train = int(m * 0.7)
m_valid = int(m * 0.2)
p = numpy_rng.permutation(m)
train_pids = p[0 : m_train]
valid_pids = p[m_train : m_train+m_valid]
test_pids = p[m_train+m_valid : ]
train = gensamples(train_pids)
valid = gensamples(valid_pids)
test = gensamples(test_pids)
return (train, valid, test, train_pids, valid_pids, test_pids)
def sample(indices, pos_downsample=1.0, neg_pos_ratio=1.0):
"""Sample positive and negative data
Args:
indices: [(pid, vid)]
neg_pos_ratio: #neg / #pos
Returns:
The sampled data indices. (train, vaid, test) each is [(i, j, 0/1)].
"""
print "Sampling ..."
data = cachem.load('sample')
if data is not None: return data
import random
from reid.utils.math_utils import numpy_rng
n_imgs = len(indices)
def gensamples(pids):
samples = []
# Positive samples
for i in xrange(n_imgs):
if indices[i][0] not in pids: continue
j = i + 1
while j < n_imgs and indices[i][0] == indices[j][0]:
if numpy_rng.rand() < pos_downsample:
samples.append((i, j, True))
j += 1
# Negative samples
n = int(len(samples) * neg_pos_ratio)
for k in xrange(n):
while True:
i, j = numpy_rng.randint(0, n_imgs, 2)
if indices[i][0] in pids and indices[j][0] in pids and \
indices[i][0] != indices[j][0] and \
(i,j,False) not in samples and (j,i,False) not in samples:
samples.append((i, j, False))
break
random.shuffle(samples)
return samples
# Split by pid
m = indices[-1][0] + 1
m_train = int(m * 0.7)
m_valid = int(m * 0.2)
p = numpy_rng.permutation(m)
train_pids = p[0:m_train]
valid_pids = p[m_train:m_train + m_valid]
test_pids = p[m_train + m_valid:]
train = gensamples(train_pids)
valid = gensamples(valid_pids)
test = gensamples(test_pids)
return (train, valid, test, train_pids, valid_pids, test_pids)
