def test():
np.random.seed(17)
X = np.load('data/cifar_gist.npy')
Y = np.load('data/cifar_label.npy')
traindata = X[:5000]
trainlabel = Y[:5000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# train model
dksh = Sparse_DKSH(64, 1000, 10, RBF)
tic = time.clock()
dksh.train(traindata, trainlabel)
toc = time.clock()
print 'time:', toc-tic
H_test = dksh.queryhash(testdata)
H_base = dksh.queryhash(basedata)
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']
def test():
#np.random.seed(47)
X = np.load('data/cifar_gist.npy')
Y = np.load('data/cifar_label.npy')
traindata = X[:59000]
trainlabel = Y[:59000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# train model
sdh = SDH(32, 300, 10, RBF)
tic = time.clock()
sdh.train(traindata, trainlabel)
toc = time.clock()
print 'time:', toc-tic
H_test = sdh.queryhash(testdata)
H_base = sdh.queryhash(basedata)
idx = np.argsort(sdh.trainlabel).squeeze()
bb = sdh.B[idx[0:59000:6000]]
classham = (32-np.dot(bb,bb.T))/2
print np.sum(classham)
print classham
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']
def test():
np.random.seed(47)
X = np.load('data/cifar_gist.npy')
Y = np.load('data/cifar_label.npy')
traindata = X[:5000]
trainlabel = Y[:5000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# train model
alg = SHLE(32, 10, lambda : RBFClassifier(X[:300]))
alg.train(traindata, trainlabel)
H_test = alg.queryhash(testdata)
H_base = alg.queryhash(basedata)
'''
idx = np.argsort(alg.trainlabel).squeeze()
bb = alg.H[idx[0:5000:530]]
classham = (32-np.dot(bb,bb.T))/2
print np.sum(classham)
print classham
'''
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']
def test():
X = np.load('data/cifar_gist.npy')
Y = np.load('data/cifar_label.npy')
idx = np.arange(60000, dtype=np.int32)
np.random.shuffle(idx)
X = X[idx]
Y = Y[idx]
traindata = X[:5000]
trainlabel = Y[:5000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
# train model
ksh = KSH(48, 300, 10, RBF)
tic = time.clock()
ksh.train(traindata, trainlabel)
# ksh = ksh2.KSH(5000, 300, 12, traindata, trainlabel, RBF)
toc = time.clock()
print 'time:', toc-tic
H_base = ksh.basehash(basedata)
H_test = ksh.queryhash(testdata)
# evaluate
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']
def test(list_algo_name, list_bits, loader):
seeds = [7, 17]#, 37, 47, 67, 97, 107, 127, 137, 157]
for algo_name in list_algo_name:
for nbit in list_bits:
print '======execute {} at bit {}======'.format(algo_name, nbit)
print '====total process round: {}====='.format(len(seeds))
li_results = []
for sd in seeds:
print '\nround #{}...'.format(len(li_results)+1)
traindata, trainlabel, basedata, baselabel, testdata, testlabel = loader.split(sd)
alg = hash_factory(algo_name, nbit, 21, 300)
tic = time.clock()
alg.train(traindata, trainlabel)
toc = time.clock()
print 'time:', toc-tic
H_test = alg.queryhash(testdata)
H_base = alg.queryhash(basedata)
# make labels
#gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
gnd_truth = (np.dot(testlabel, baselabel.T) >= 1).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, len(baselabel), len(baselabel), trn_time=toc-tic)
li_results.append(res)
eva_checkpoint(algo_name, nbit, li_results)
def test(n_bit):
np.random.seed(17)
X = np.load('cifar10_data/cifar10_gist.npy')
Y = np.load('cifar10_data/cifar10_label.npy')
traindata = X[:59000]
trainlabel = Y[:59000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# train model
dish = DISH_K(n_bit, 1000, 10, RBF)
tic = time.time()
dish.train(traindata, trainlabel)
toc = time.time()
print 'time:', toc - tic
H_test = dish.queryhash(testdata)
H_base = dish.basehash(basedata)
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']
def test(list_algo_name, list_bits):
root = '../../ILSVRC2012_caffe/npy_data/'
baselabel = np.load(root+'trainlabel.npy')
testlabel = np.load(root+'testlabel.npy')
testdata = np.load(root+'testdata_fc6_norm.npy')
basicdata = np.load(root+'traindata_fc6_norm_1.npy')
seeds = [7]#, 17, 37, 47, 67, 97, 107, 127, 137, 157]
for algo_name in list_algo_name:
for nbit in list_bits:
print '======execute {} at bit {}======'.format(algo_name, nbit)
print '====total process round: {}====='.format(len(seeds))
li_results = []
for sd in seeds:
print '\nround #{}...'.format(len(li_results)+1)
# load data
np.random.seed(sd)
idx = np.arange(100000, dtype=np.int32)
np.random.shuffle(idx)
traindata = basicdata[idx]
trainlabel = baselabel[idx]
alg = hash_factory(algo_name, nbit, 1000, 1000)
tic = time.clock()
alg.train(traindata, trainlabel)
toc = time.clock()
print 'time:', toc-tic
print 'hash testing data...'
H_test = alg.queryhash(testdata)
print 'hash training data...'
H_base = np.zeros((baselabel.shape[0], H_test.shape[1]), dtype=np.uint8)
pt = 0
for i in xrange(13):
print 'hashing', i, '...'
X = np.load('traindata_fc6_norm_{}.npy'.format(i+1))
n = X.shape[0]
H_base[pt:pt+n] = alg.basehash(X)
pt += n
print 'testing...'
batch_results = []
for i in xrange(50):
# make labels
print 'testing batch', i, '...'
gnd_truth = np.array([y == baselabel for y in testlabel[i*1000:(i+1)*1000]]).astype(np.int8)
batch_results.append(hash_evaluation(H_test, H_base, gnd_truth, 0, topN=5000, trn_time=toc-tic))
eva_checkpoint(algo_name, nbit, batch_results)
li_results.append(batch_eva_ensem(batch_results))
eva_checkpoint(algo_name, nbit, li_results)
def test_cnn(n_bit):
np.random.seed(17)
if use_vgg:
X = np.load('cifar10_data/cifar10_256_data.npy')
else:
X = np.load('cifar10_data/cifar10_data.npy')
Y = np.load('cifar10_data/cifar10_label.npy')
is_train = np.load('cifar10_data/cifar10_istrain.npy')
train_idx = np.argwhere(is_train==1)[:,0]
nottrain_idx = np.argwhere(is_train==0)[:,0]
np.random.shuffle(nottrain_idx)
test_idx = nottrain_idx[-1000:]
base_idx = np.setdiff1d(np.arange(60000, dtype=np.int32), test_idx)
traindata = X[train_idx]
trainlabel = Y[train_idx]
testdata = X[test_idx]
testlabel = Y[test_idx]
basedata = X[base_idx]
baselabel = Y[base_idx]
del X
# train model
dksh = DISH_D(n_bit, 10)
tic = time.time()
dksh.train(traindata, trainlabel)
toc = time.time()
print 'time:', toc-tic
H_test = dksh.queryhash(testdata)
H_base = dksh.queryhash(basedata)
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, len(H_base), topN=len(H_base), trn_time=toc-tic)
print 'MAP:', res['map'], 'Pre2:', res['pre2']
np.random.seed(97)
X = np.load('data/cifar_gist.npy')
Y = np.load('data/cifar_label.npy')
traindata = X[:59000]
trainlabel = Y[:59000]
basedata = X[:59000]
baselabel = Y[:59000]
testdata = X[59000:]
testlabel = Y[59000:]
# train model
dksh = DGH(48, 300)
tic = time.clock()
dksh.train(traindata)
toc = time.clock()
print 'time:', toc-tic
H_test = dksh.queryhash(testdata)
H_base = dksh.queryhash(basedata)
idx = np.argsort(baselabel)
print H_base[idx[0:59000:3000]]
# make labels
gnd_truth = np.array([y == baselabel for y in testlabel]).astype(np.int8)
print 'testing...'
res = hash_evaluation(H_test, H_base, gnd_truth, 59000)
print 'MAP:', res['map']