def __init__(self, model_path, dataset):
self.net_ = None
self.dataset_ = None
self.image_feats_ = None
self.config_ = load_model_config(dataset)
self.load_dataset(dataset)
self.set_cnn_model(model_path, self.config_.deploy_file)
self.load_cached_features()
def __init__(self, model_path, dataset):
self.net_ = None
self.dataset_ = None
self.image_feats_ = None
self.config_ = load_model_config(dataset)
self.load_dataset(dataset)
self.set_cnn_model(model_path, self.config_.deploy_file)
self.load_cached_features()
def cache_features(net_, dataset):
config = load_model_config(dataset)
feat_path = config.FEAT_PATH
if os.path.exists(feat_path):
return loadmat(feat_path)['feats']
# make dir
feat_rt = os.path.split(feat_path)[0]
if not os.path.exists(feat_rt):
os.makedirs(feat_rt)
# extract feature
print 'caching features'
feats = extract_features(net_, config)
savemat(feat_path, {'feats': feats})
return feats
def cache_features(net_, dataset):
config = load_model_config(dataset)
feat_path = config.FEAT_PATH
if os.path.exists(feat_path):
return loadmat(feat_path)['feats']
# make dir
feat_rt = os.path.split(feat_path)[0]
if not os.path.exists(feat_rt):
os.makedirs(feat_rt)
# extract feature
print 'caching features'
feats = extract_features(net_, config)
savemat(feat_path, {'feats': feats})
return feats
def sketch_retrieval(net_, im, image_feats):
config = load_model_config()
im = imresize(im, config.input_dim)
im = im.astype(np.single) - config.mean_val
transformed = do_multiview_crop(im, config.crop_dim)
num, chns, rows, cols = transformed.shape
# reshape data layer
net_.blobs['data'].reshape(*(num, chns, rows, cols))
# forward
output = net_.forward(data=transformed.astype(np.float32, copy=False))
target = reshape_feature(output[config.feat_layer].copy())
# get retrieval results
multiview_dists = compute_view_specific_distance(target, image_feats)
ave_dist = multiview_dists.mean(axis=0)
return ave_dist.flatten().argsort() # return ranklist
def sketch_retrieval(net_, im, image_feats):
config = load_model_config()
im = imresize(im, config.input_dim)
im = im.astype(np.single) - config.mean_val
transformed = do_multiview_crop(im, config.crop_dim)
num, chns, rows, cols = transformed.shape
# reshape data layer
net_.blobs['data'].reshape(*(num, chns, rows, cols))
# forward
output = net_.forward(data=transformed.astype(np.float32, copy=False))
target = reshape_feature(output[config.feat_layer].copy())
# get retrieval results
multiview_dists = compute_view_specific_distance(target, image_feats)
ave_dist = multiview_dists.mean(axis=0)
return ave_dist.flatten().argsort() # return ranklist
