def train_model(index, mymodel, model_version):
model = mymodel
train_model = model.train
test_model = theano.function(model.X, model.f(model.X, is_train=True)[0])
PN_ratio = 1
ind_list = get_ind_list_zsz()
train_ind, validation_ind, test_ind = ind_list[index]
train_size, validation_size, test_size = train_ind.shape[
0], validation_ind.shape[0], test_ind.shape[0]
cam_ind = numpy.asarray([0, 1], dtype='int32')
batch_train_size = 5
iterations = 200001
rank_1 = 0
rank_5 = 0
for i in range(iterations):
person_inds = numpy.sort(
numpy.random.choice(train_ind, batch_train_size, replace=False))
[positive_pair_a, positive_pair_b, negative_pair_a, negative_pair_b] = \
setup.generate_ind_pairs(img_num_per_cam_person, person_inds, cam_ind)
pos_n = positive_pair_a.size
neg_n = negative_pair_a.size
neg_n = min(pos_n * PN_ratio, neg_n)
neg_ind = numpy.sort(
numpy.random.choice(negative_pair_a.size, neg_n, replace=False))
pair_ind_a = numpy.hstack((positive_pair_a, negative_pair_a[neg_ind]))
pair_ind_b = numpy.hstack((positive_pair_b, negative_pair_b[neg_ind]))
y = numpy.hstack((numpy.zeros(
(pos_n, ), dtype='int32'), numpy.ones((neg_n, ), dtype='int32')))
# dangerous! Do not touch any code above after the "for"!
cost = train_model(img_data[pair_ind_a, :, :, :],
img_data[pair_ind_b, :, :, :], y)
if i % 10 == 0:
print 'iteration: %d, cost: %f' % (i, cost)
if i % 500 == 499:
t_start = time.time()
v_rank = setup.testing(test_model, img_num_per_cam_person,
img_data, validation_ind, validation_size)
print(
'validation set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f'
% (v_rank[0] / validation_size * 100, v_rank[4] /
validation_size * 100, v_rank[9] / validation_size * 100))
t_end = time.time()
print "time consumed is " + str(t_end - t_start) + 'secs.'
t_start = time.time()
t_rank = setup.testing(test_model, img_num_per_cam_person,
img_data, test_ind, test_size)
print(
'test set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(t_rank[0] / test_size * 100, t_rank[4] / test_size * 100,
t_rank[9] / test_size * 100))
t_end = time.time()
print "time consumed is " + str(t_end - t_start) + 'secs.'
folder_out = './models/' + 'model_' + model_version + '_' + str(
index)
if not os.path.isdir(folder_out):
os.mkdir(folder_out)
if v_rank[0] / validation_size * 100 > rank_1 or v_rank[
4] / validation_size * 100 > rank_5:
rank_1 = v_rank[0] / validation_size * 100
rank_5 = v_rank[4] / validation_size * 100
rank_1_t = t_rank[0] / test_size * 100
rank_5_t = t_rank[4] / test_size * 100
model.save(folder_out + '/v_' + str(int(rank_1)) + '_' + str(int(rank_5)) \
+ '_t_' + str(int(rank_1_t)) + '_' + str(int(rank_5_t)) + '.mdl')
[positive_pair_a, positive_pair_b, negative_pair_a, negative_pair_b] = \
setup.generate_ind_pairs(img_num_per_cam_person, person_inds, cam_ind)
pos_n = positive_pair_a.size
neg_n = negative_pair_a.size
neg_n = min(pos_n * PN_ratio, neg_n)
neg_ind = numpy.sort(numpy.random.choice(negative_pair_a.size, neg_n, replace=False))
pair_ind_a = numpy.hstack((positive_pair_a, negative_pair_a[neg_ind]))
pair_ind_b = numpy.hstack((positive_pair_b, negative_pair_b[neg_ind]))
y = numpy.hstack((numpy.zeros((pos_n,), dtype='int32'), numpy.ones((neg_n,), dtype='int32')))
############ NO TOUCH ################
cost = train_model(img_data[pair_ind_a, :, :, :], img_data[pair_ind_b, :, :, :], y)
if i % 10 == 0:
print 'iteration: %d, cost: %f' % (i, cost)
if i % 1000 == 999:
## validation and testing
t_rank = setup.testing(test_model, img_num_per_cam_person, img_data, test_ind, test_size)
print('test set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(t_rank[0]/test_size * 100, t_rank[4]/test_size * 100, t_rank[9]/test_size * 100))
v_rank = setup.testing(test_model, img_num_per_cam_person, img_data, validation_ind, validation_size)
print('validation set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(v_rank[0]/validation_size * 100, v_rank[4]/validation_size * 100, v_rank[9]/validation_size * 100))
## save model
model.save_model(i)
rank_mat = 'rank_iter_%i' % i + '.mat'
scipy.io.savemat(rank_mat, {'v_rank': v_rank, 't_rank': t_rank})
def train_model(index, mymodel, model_version):
model = mymodel
train_model = model.train
test_model = theano.function(model.X, model.f(model.X, is_train = True)[0])
PN_ratio = 1
ind_list = get_ind_list_zsz()
train_ind, validation_ind, test_ind = ind_list[index]
train_size, validation_size ,test_size = train_ind.shape[0], validation_ind.shape[0], test_ind.shape[0]
cam_ind = numpy.asarray([0, 1], dtype='int32')
batch_train_size = 5
iterations = 200001
rank_1 = 0
rank_5 = 0
for i in range(iterations):
person_inds = numpy.sort(numpy.random.choice(train_ind, batch_train_size, replace=False))
[positive_pair_a, positive_pair_b, negative_pair_a, negative_pair_b] = \
setup.generate_ind_pairs(img_num_per_cam_person, person_inds, cam_ind)
pos_n = positive_pair_a.size
neg_n = negative_pair_a.size
neg_n = min(pos_n * PN_ratio, neg_n)
neg_ind = numpy.sort(numpy.random.choice(negative_pair_a.size, neg_n, replace=False))
pair_ind_a = numpy.hstack((positive_pair_a, negative_pair_a[neg_ind]))
pair_ind_b = numpy.hstack((positive_pair_b, negative_pair_b[neg_ind]))
y = numpy.hstack((numpy.zeros((pos_n,), dtype='int32'), numpy.ones((neg_n,), dtype='int32')))
# dangerous! Do not touch any code above after the "for"!
cost = train_model(img_data[pair_ind_a, :, :, :], img_data[pair_ind_b, :, :, :], y)
if i % 10 == 0:
print 'iteration: %d, cost: %f' % (i, cost)
if i % 500 == 499:
t_start = time.time()
v_rank = setup.testing(test_model, img_num_per_cam_person, img_data, validation_ind, validation_size)
print('validation set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(v_rank[0]/validation_size * 100, v_rank[4]/validation_size * 100, v_rank[9]/validation_size * 100))
t_end = time.time()
print "time consumed is " + str(t_end - t_start) + 'secs.'
t_start = time.time()
t_rank = setup.testing(test_model, img_num_per_cam_person, img_data, test_ind, test_size)
print('test set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(t_rank[0]/test_size * 100, t_rank[4]/test_size * 100, t_rank[9]/test_size * 100))
t_end = time.time()
print "time consumed is " + str(t_end - t_start) + 'secs.'
folder_out = './models/' + 'model_' + model_version + '_' + str(index)
if not os.path.isdir(folder_out):
os.mkdir(folder_out)
if v_rank[0]/validation_size *100 > rank_1 or v_rank[4]/validation_size *100 > rank_5:
rank_1 = v_rank[0]/validation_size *100
rank_5 = v_rank[4]/validation_size *100
rank_1_t = t_rank[0]/test_size*100
rank_5_t = t_rank[4]/test_size*100
model.save(folder_out + '/v_' + str(int(rank_1)) + '_' + str(int(rank_5)) \
+ '_t_' + str(int(rank_1_t)) + '_' + str(int(rank_5_t)) + '.mdl')
epochs = results.epochs
lr = results.learning_rate
gpu = results.boolean_t
print_every = 30
if gpu==True:
using_gpu = torch.cuda.is_available()
device = 'gpu'
print('GPU On');
else:
print('GPU Off');
device = 'cpu'
# Loading Dataset
image_trainloader, image_testloader, image_valloader, image_trainset = setup.loading_data(data_dir)
class_to_idx = image_trainset.class_to_idx
# Network Setup
model, input_size = setup.make_model(arch, hidden_units, lr)
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(), lr=lr)
# Training Model
setup.my_DLM(model, image_trainloader, image_valloader, epochs, print_every, criterion, optimizer, device)
# Testing Model
setup.testing(model, image_testloader)
# Saving Checkpoint
setup.save_checkpoints(model, arch, lr, epochs, input_size, hidden_units, class_to_idx, checkpoint_path)
neg_ind = numpy.sort(
numpy.random.choice(negative_pair_a.size, neg_n, replace=False))
pair_ind_a = numpy.hstack((positive_pair_a, negative_pair_a[neg_ind]))
pair_ind_b = numpy.hstack((positive_pair_b, negative_pair_b[neg_ind]))
y = numpy.hstack((numpy.zeros(
(pos_n, ), dtype='int32'), numpy.ones((neg_n, ), dtype='int32')))
############ NO TOUCH ################
cost = train_model(img_data[pair_ind_a, :, :, :],
img_data[pair_ind_b, :, :, :], y)
if i % 10 == 0:
print 'iteration: %d, cost: %f' % (i, cost)
if i % 1000 == 999:
## validation and testing
t_rank = setup.testing(test_model, img_num_per_cam_person, img_data,
test_ind, test_size)
print('test set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f' %
(t_rank[0] / test_size * 100, t_rank[4] / test_size * 100,
t_rank[9] / test_size * 100))
v_rank = setup.testing(test_model, img_num_per_cam_person, img_data,
validation_ind, validation_size)
print(
'validation set: rank 1: %4.2f, rank 5: %4.2f, rank 10: %4.2f'
% (v_rank[0] / validation_size * 100, v_rank[4] / validation_size *
100, v_rank[9] / validation_size * 100))
## save model
model.save_model(i)
rank_mat = 'rank_iter_%i' % i + '.mat'
scipy.io.savemat(rank_mat, {'v_rank': v_rank, 't_rank': t_rank})
