def show_cmc(model, X, indices, samples):
print "Computing cmc ..."
result = cachem.load('cmc')
if result is not None: return result
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
from reid.utils import cmc
x = T.matrix()
y, thr = model.get_output(x)
thr.append(y)
comp = CompLayer()
y, thr = comp.get_output(thr)
f = theano.function(inputs=[x], outputs=y)
train_pids = samples[3]
gX, gY, pX, pY = [], [], [], []
for i, (pid, vid) in enumerate(indices):
if pid in train_pids: continue
if vid == 0:
gX.append(i)
gY.append(pid)
else:
pX.append(i)
pY.append(pid)
def compute_distance(i, j):
y = f(numpy.hstack(
(X[gX[i]:gX[i] + 1, :], X[pX[j]:pX[j] + 1, :]))).ravel()
return -y[-1]
return cmc.count_lazy(compute_distance, gY, pY, 100, 1)
def show_cmc(model, X, indices, samples):
print "Computing cmc ..."
result = cachem.load('cmc')
if result is not None: return result
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
from reid.utils import cmc
x = T.matrix()
y, thr = model.get_output(x)
thr.append(y)
comp = CompLayer()
y, thr = comp.get_output(thr)
f = theano.function(inputs=[x], outputs=y)
train_pids = samples[3]
gX, gY, pX, pY = [], [], [], []
for i, (pid, vid) in enumerate(indices):
if pid in train_pids: continue
if vid == 0:
gX.append(i)
gY.append(pid)
else:
pX.append(i)
pY.append(pid)
def compute_distance(i, j):
y = f(numpy.hstack((X[gX[i]:gX[i]+1, :], X[pX[j]:pX[j]+1, :]))).ravel()
return -y[-1]
return cmc.count_lazy(compute_distance, gY, pY, 100, 1)
def compute_result(model, batch_dir, images, samples):
"""Compute output for dataset
Args:
model: Deep model
dataset: Dataset X = X1_X2, Y = A1_A2_(0/1)
images: [img]
samples: (train, vaid, test) each is [(i, j, 0/1)]
Returns:
(train, valid, test) where each is
(img_list_1, img_list_2 output_matrix, target_matrix)
"""
print "Computing result ..."
result = cachem.load('result')
if result is not None: return result
import glob
import cPickle
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
x = T.matrix()
y, thr = model.get_output(x)
thr.append(y)
comp = CompLayer()
y, thr = comp.get_output(thr)
f = theano.function(inputs=[x], outputs=y)
def compute_output(X):
outputs = [f(X[i:i + 1, :]).ravel() for i in xrange(X.shape[0])]
return numpy.asarray(outputs)
train_files = glob.glob(os.path.join(batch_dir, 'train_*.pkl'))
valid_files = glob.glob(os.path.join(batch_dir, 'valid_*.pkl'))
test_files = glob.glob(os.path.join(batch_dir, 'test_*.pkl'))
train_files.sort()
valid_files.sort()
test_files.sort()
# Setup parameters
n_train_batches = len(train_files)
n_valid_batches = len(valid_files)
n_test_batches = len(test_files)
for i in xrange(n_train_batches):
with open(train_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
train = ([images[i] for i, __, __ in samples[0]],
[images[j] for __, j, __ in samples[0]], cum_Y, cum_T)
for i in xrange(n_valid_batches):
with open(valid_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
valid = ([images[i] for i, __, __ in samples[1]],
[images[j] for __, j, __ in samples[1]], cum_Y, cum_T)
for i in xrange(n_test_batches):
with open(test_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
test = ([images[i] for i, __, __ in samples[2]],
[images[j] for __, j, __ in samples[2]], cum_Y, cum_T)
return (train, valid, test)
def compute_result(model, batch_dir, images, samples):
"""Compute output for dataset
Args:
model: Deep model
dataset: Dataset X = X1_X2, Y = A1_A2_(0/1)
images: [img]
samples: (train, vaid, test) each is [(i, j, 0/1)]
Returns:
(train, valid, test) where each is
(img_list_1, img_list_2 output_matrix, target_matrix)
"""
print "Computing result ..."
result = cachem.load('result')
if result is not None: return result
import glob
import cPickle
import theano
import theano.tensor as T
from reid.models.layers import CompLayer
x = T.matrix()
y, thr = model.get_output(x)
thr.append(y)
comp = CompLayer()
y, thr = comp.get_output(thr)
f = theano.function(inputs=[x], outputs=y)
def compute_output(X):
outputs = [f(X[i:i+1, :]).ravel() for i in xrange(X.shape[0])]
return numpy.asarray(outputs)
train_files = glob.glob(os.path.join(batch_dir, 'train_*.pkl'))
valid_files = glob.glob(os.path.join(batch_dir, 'valid_*.pkl'))
test_files = glob.glob(os.path.join(batch_dir, 'test_*.pkl'))
train_files.sort()
valid_files.sort()
test_files.sort()
# Setup parameters
n_train_batches = len(train_files)
n_valid_batches = len(valid_files)
n_test_batches = len(test_files)
for i in xrange(n_train_batches):
with open(train_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
train = ([images[i] for i, __, __ in samples[0]],
[images[j] for __, j, __ in samples[0]],
cum_Y, cum_T)
for i in xrange(n_valid_batches):
with open(valid_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
valid = ([images[i] for i, __, __ in samples[1]],
[images[j] for __, j, __ in samples[1]],
cum_Y, cum_T)
for i in xrange(n_test_batches):
with open(test_files[i], 'rb') as fid:
X, T = cPickle.load(fid)
Y = compute_output(X)
if i == 0:
cum_T = T
cum_Y = Y
else:
cum_T = numpy.vstack((cum_T, T))
cum_Y = numpy.vstack((cum_Y, Y))
test = ([images[i] for i, __, __ in samples[2]],
[images[j] for __, j, __ in samples[2]],
cum_Y, cum_T)
return (train, valid, test)
