def __init__(self, queue, sketch_dir, image_dir, triplet_path, mean, hard_ratio, batch_size, phase):
"""Setup the TripletSamplingDataLayer."""
super(TripletSamplingDataFetcher, self).__init__()
# mean = mean
self._queue = queue
self._phase = phase
self.sketch_transformer = Transformer(225, 1, mean, self._phase == "TRAIN")
self.sketch_dir = sketch_dir
self.anc_bm = MemoryBlockManager(sketch_dir)
self.pos_neg_bm = MemoryBlockManager(image_dir)
self.hard_ratio = hard_ratio
self.mini_batchsize = batch_size
self.load_triplets(triplet_path)
def __init__(self, queue, layer_params):
"""Setup the TripletSamplingDataLayer."""
super(TripletSamplingDataFetcher, self).__init__()
self._queue = queue
mean = layer_params['mean']
self._phase = layer_params['phase']
self.sketch_transformer = Transformer(225, 1, mean,
self._phase == "TRAIN")
self.anc_bm = MemoryBlockManager(layer_params['sketch_dir'])
self.pos_neg_bm = MemoryBlockManager(layer_params['image_dir'])
self.hard_ratio = layer_params['hard_ratio']
self.mini_batchsize = layer_params['batch_size']
self.load_triplets(layer_params['triplet_path'])
def __init__(self, queue, sketch_dir, mean, hard_ratio, batch_size):
"""Setup the StrokeSamplingDataLayer."""
super(SamplingDataFetcher, self).__init__()
# mean = mean
self._queue = queue
self.sketch_transformer = Transformer(225, 1, mean)
self.sketch_dir = sketch_dir
self.sketch_bm = MemoryBlockManager(sketch_dir)
self.hard_ratio = hard_ratio
self.mini_batchsize = batch_size
def __init__(self, queue, sketch_data, image_data, sketch_label, image_label, triplets, queue_paras, phase):
"""Setup the TripletSamplingDataLayer."""
super(TripletSamplingDataFetcher, self).__init__()
# mean = mean
self.im_size = queue_paras['im_size']
self.cp_size = queue_paras['cp_size']
self.chns = queue_paras['chns']
self.mean = queue_paras['mean']
self.batchsize = queue_paras['batch_size']
self.num_epoch = queue_paras['num_epochs']
self._queue = queue
self._phase = phase
self.sketch_transformer = Transformer(self.cp_size, self.chns, self.mean, self._phase == "TRAIN")
if sketch_label is None or image_label is None:
self.has_label = False
else:
self.has_label = True
self.anc_bm = MemoryBlockManager(sketch_data, sketch_label, self.has_label)
self.pos_neg_bm = MemoryBlockManager(image_data, image_label, self.has_label)
self.triplets = triplets
def extract_features(net_, config):
db = MemoryBlockManager(config.DB_PATH, False)
if config.verbose:
print 'extracting features of %s' % config.dataset
t = time()
num_samples = 10 * db.num_samples
feats = np.zeros((num_samples, get_feature_dims(net_, config.feat_layer)),
np.single)
idx = 0
while not db.eof():
batch_data, _ = db.pop_batch(config.batchsize) # batchsize 128
transformed = do_multiview_crop(batch_data, config.crop_dim)
transformed = transformed - config.mean_val
num, chns, rows, cols = transformed.shape
net_.blobs['data'].reshape(*(num, chns, rows, cols))
output = net_.forward(data=transformed.astype(np.float32, copy=False))
target = output[config.feat_layer]
feats[idx:idx + num, ::] = reshape_feature(target)
idx += num
if config.verbose:
print 'feature computation completed (%0.2f sec.)' % (time() - t)
return feats
