def load_data(num_class):
path_to_data = 'data/cifar%d.pickle' % num_class
if os.path.exists(path_to_data):
print '%s found!' % path_to_data
duration = time.time()
dict = pickle.load(open(path_to_data, 'rb'))
duration = time.time() - duration
print 'Loading data takes %f second' % duration
return dict['images'], dict['labels']
print '%s not found. Prepareing Data...' % path_to_data
duration = time.time()
if num_class == 10:
images, labels = dataset.parse_data_cifar10()
else:
images, labels = dataset.parse_data_cifar100()
images, labels, _, _, _ = dataset.preprocess_image(images,
labels,
num_class=num_class)
duration = time.time() - duration
print 'Prepareing Data take %f second' % duration
pickle.dump({'images': images, 'labels': labels}, open(path_to_data, 'wb'))
return images, labels
def inference_on_single_image(image, image_size, model, device):
image = preprocess_image(image, image_size)
image = torch.from_numpy(image).type(torch.FloatTensor).cuda(device)
image = image.unsqueeze(0)
with torch.no_grad():
disparities = model(image)
disp = disparities[0]
return disp
def run_test(epoch, model, sess, times, len_accs):
logger.info('begin test epoch {}'.format(epoch))
st_time = time.time()
with open(args.testid_path) as f:###
ids = f.readline().strip().split(',')
#get test probe
probe_features = []
for id_ in ids:
#path = os.path.join(args.probe_dir, id_.zfill(3), 'slice233-feature.mat')
#if os.path.exists(path):
if False:
mat = sio.loadmat(path)
features = mat['features']
else:
path = os.path.join(args.probe_dir, id_.zfill(3), 'slice0.1.mat')
mat = sio.loadmat(path)
imgs = mat['slice']
imgs = dataset.preprocess_image(imgs)
feed = {model.input_img:imgs}
feed[K.learning_phase()] = 0
calc_obj = [model.output_feature]
calc_ans = sess.run(calc_obj, feed_dict=feed)
features = calc_ans[0]
#path = os.path.join(args.probe_dir, id_.zfill(3), 'slice233-100-feature.mat')
#sio.savemat(path, {'features':features}, do_compression=True)
probe_features.append(features)
logger.info('get features of probe end')
# get gallery feature
ixs = []
imgs = []
for id_ in ids:
path = os.path.join(args.gallery_dir, id_.zfill(3), 'slice0.1.mat')
mat = sio.loadmat(path)
len_ = len(mat['slice'])
id_ixs = []
id_imgs = []
for i in range(times):
ix = np.random.randint(len_)
img = mat['slice'][ix]
id_imgs.append(img)
id_ixs.append(ix)
ixs.append(id_ixs)
imgs.append(id_imgs)
ixs = list(zip(*ixs))
imgs = list(zip(*imgs))
gallery_features = []
for i in range(times):
img = imgs[i]
img = np.array(img)
img = dataset.preprocess_image(img)
feed = {model.input_img:img}
feed[K.learning_phase()] = 0
calc_obj = [model.output_feature]
calc_ans = sess.run(calc_obj, feed_dict=feed)
features = calc_ans[0]
gallery_features.append(features)
#logger.info('{}'.format(str(ixs)))
#calc acc
sum_imgs = 0
for id_features in probe_features:
sum_imgs += len(id_features)
sum_accs = np.zeros(len_accs)
p_args = []
for i in range(times):
p_args.append([ids,probe_features,gallery_features[i],len_accs])
pool = Pool(times)
res = pool.map(proc_calc, p_args)
pool.close()
pool.join()
for one_accs in res:
#print(one_accs/sum_imgs)
sum_accs += one_accs
avg_accs = sum_accs / (sum_imgs*times)
print(avg_accs)
logger.info('[epoch {}]{}'.format(epoch,avg_accs))
return avg_accs[0],int(time.time()-st_time)
image_files = os.listdir(image_dir)
else:
image_files = [image_dir.split('/')[-1]]
image_dir = '/'.join(image_dir.split('/')[:-1])
random.shuffle(image_files)
image_files = image_files[:10000]
model: Model = build_model()
model.load_weights(sys.argv[1])
correct, miss = 0, 0
X = []
for image_name in image_files:
fullpath = image_dir + '/' + image_name
img = preprocess_image(Image.open(fullpath))
X.append(np.expand_dims(img, 2))
X = np.asarray(X)
preds = model.predict(X)
Y_pred = preds.reshape((-1, char_count, len(char_set)))
Y_pred = np.argmax(Y_pred, axis=2)
Y = []
Y_map = np.zeros(shape=(len(char_set), len(char_set)), dtype='float32')
for i, y in enumerate(Y_pred):
label = ''
for ch_index in y:
label += char_set[ch_index]
if __name__ == '__main__':
oimages, labels = parse_data_cifar100()
ptions = parse_args()
if options.dataset == 'cifar10':
NUM_CLASS = 10
elif options.dataset == 'cifar100':
NUM_CLASS = 100
else:
print 'Unknown Dataset: %s' % options.dataset
exit(0)
print 'Prepareing Data'
duration = time.time()
if NUM_CLASS == 10:
images, labels = dataset.parse_data_cifar10()
else:
images, labels = dataset.parse_data_cifar100()
images, labels, _, _, _ = dataset.preprocess_image(images, labels, num_class=10)
duration = time.time() - duration
print 'Prepareing Data take %f second' % duration
model_dir = 'model/' + options.dataset
logdir = 'logs/' + options.dataset
sess = tf.Session()
net = model.DCTI(batch_size=options.batch_size, num_epoch=options.num_epoch, seed=options.seed, num_class=NUM_CLASS, learning_rate=options.learning_rate, \
regularization_rate=options.regularization_rate, model_dir=model_dir, logdir=logdir)
net.train(sess, images, labels)