2, 1,
0)) # the reference model has channels in BGR order instead of RGB
if net_name == "AlexNet":
self.net.blobs['data'].reshape(1, 3, 227, 227)
else:
self.net.blobs['data'].reshape(1, 3, 224, 224)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("net", help="Net, Choose from: AlexNet, VGG, VGG_SOS")
parser.add_argument("figure_name", help="Please specify MAIN file name")
parser.add_argument("PATH_TO_DATA",
help="Please specify the path to the images")
parser.add_argument("annotations", help="Annotated frames")
parser.add_argument("--PATH_TO_DATA_2",
help="Please specify the path to second set of images",
default=None)
parser.add_argument("--annotations_2",
help="Annotated frames for second set of images",
default=None)
args = parser.parse_args()
fe = CNNFeatureExtractor(args.net)
layers = constants.NET_PARAMS[args.net][2]
X = fe.forward_pass(args.PATH_TO_DATA,
args.annotations,
list_of_layers=layers,
sampling_rate=1)
utils.plot_all_layers(X, args.net, args.figure_name, list_of_layers=layers)
hypercolumns_layers = ['conv2','conv3', 'conv4'] X, label_map, frm_map = fe.forward_pass(args.PATH_TO_DATA, args.annotations, hypercolumns_layers) X_hc = utils.make_hypercolumns_vector(hypercolumns_layers, X) utils.plot_hypercolumns(X_hc, args.net, label_map, frm_map, args.figure_name, hypercolumns_layers) elif args.vlad: layers = ['conv4', 'conv3'] k_values = [1, 5, 10] pc_values = [5, 100, 200] X, label_map, frm_map = fe.forward_pass(args.PATH_TO_DATA, args.annotations, list_of_layers = layers, sampling_rate = 1) utils.vlad_experiment(X, k_values, pc_values, label_map, frm_map, args.figure_name, list_of_layers = layers) elif args.batch_size != -1 and not args.LCD: layers = ['conv3', 'conv4', 'conv5', 'pool5'] X, label_map, frm_map = fe.forward_pass(args.PATH_TO_DATA, args.annotations,list_of_layers = layers, sampling_rate = 1, batch_size = int(args.batch_size)) utils.plot_all_layers(X, args.net, label_map, frm_map, args.figure_name, list_of_layers = layers) elif args.LCD: if args.batch_size == -1: print "ERROR: Please provide both batch size and LCD" sys.exit() layers = ['conv5_1', 'conv5_2', 'conv5_3' ] X, label_map, frm_map = fe.forward_pass(args.PATH_TO_DATA, args.annotations, list_of_layers = layers, sampling_rate = 1, batch_size = int(args.batch_size), LCD = True) utils.plot_all_layers(X, args.net, label_map, frm_map, args.figure_name, list_of_layers = layers) else: print "Plotting Individual frames" X, label_map, frm_map = fe.forward_pass(args.PATH_TO_DATA, args.annotations, list_of_layers = layers, sampling_rate = 3) data = [X, label_map, frm_map] pickle.dump(data, open(args.figure_name + ".p", "wb")) utils.plot_all_layers(X, args.net, label_map, frm_map, args.figure_name, list_of_layers = layers)
def init_caffe(self, net_name):
caffe.set_mode_gpu()
net_params = constants.NET_PARAMS[net_name]
self.net = caffe.Net(net_params[0], net_params[1], caffe.TEST)
self.transformer = caffe.io.Transformer({'data': self.net.blobs['data'].data.shape})
self.transformer.set_transpose('data', (2,0,1))
# self.transformer.set_mean('data', np.load(constants.CAFFE_ROOT + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1)) # mean pixel
self.transformer.set_raw_scale('data', 255) # the reference model operates on images in [0,255] range instead of [0,1]
self.transformer.set_channel_swap('data', (2,1,0)) # the reference model has channels in BGR order instead of RGB
if net_name == "AlexNet":
self.net.blobs['data'].reshape(1,3,227,227)
else:
self.net.blobs['data'].reshape(1,3,224,224)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("net", help = "Net, Choose from: AlexNet, VGG, VGG_SOS")
parser.add_argument("figure_name", help = "Please specify MAIN file name")
parser.add_argument("PATH_TO_DATA", help="Please specify the path to the images")
parser.add_argument("annotations", help = "Annotated frames")
parser.add_argument("--PATH_TO_DATA_2", help="Please specify the path to second set of images", default = None)
parser.add_argument("--annotations_2", help = "Annotated frames for second set of images", default = None)
args = parser.parse_args()
fe = CNNFeatureExtractor(args.net)
layers = constants.NET_PARAMS[args.net][2]
X = fe.forward_pass(args.PATH_TO_DATA, args.annotations, list_of_layers = layers, sampling_rate = 1)
utils.plot_all_layers(X, args.net, args.figure_name, list_of_layers = layers)