def main():
# logging configuration
logging.basicConfig(level=logging.INFO,
format="[%(asctime)s]: %(levelname)s: %(message)s")
# command line paser
opt = parse.parse_arg()
# GPU
opt.cuda = opt.gpuid >= 0
if opt.gpuid >= 0:
torch.cuda.set_device(opt.gpuid)
else:
logging.info("WARNING: RUN WITHOUT GPU")
# prepare dataset
opt.dataset = 'mnist'
db = dataset.prepare_db(opt)
#add wavelets:
opt.wavelets = True
opt.cutoff = 20
# initalize neural decision forest
NDF = model.prepare_model(opt)
# prepare optimizer
optim, sche = optimizer.prepare_optim(NDF, opt)
# train the neural decision forest5
best_acc = trainer.train(NDF, optim, sche, db, opt)
logging.info('The best evaluation accuracy is %f' % best_acc)
# For now only GPU version is supported
torch.cuda.set_device(opt.gpuid)
# please place the downloaded pre-trained models in the following directory
if opt.dataset == 'mnist':
model_path = "../pre-trained/mnist_depth_9_tree_1_acc_0.993.pth"
elif opt.dataset == 'cifar10':
model_path = "../pre-trained/cifar10_depth_9_tree_1_ResNet50_acc_0.9341.pth"
else:
raise NotImplementedError
# load model
model = torch.load(model_path).cuda()
# prepare dataset
db = prepare_db(opt)
# use only the evaluation subset. use db['train'] for fetching the training subset
dataset = db['eval']
# ==================================================================================
# compute saliency maps for different inputs for one splitting node
# pick a tree index and splitting node index
# tree_idx = 0
# node_idx = 0 # 0 - 510 for the 511 splitting nodes in a tree of depth 9
# get saliency maps for a specified node for different input tensors
# utils.get_node_saliency_map(dataset, model, tree_idx, node_idx, name=opt.dataset)
# ==================================================================================
# get the computational paths for the some random inputs
sample, paths, class_pred = utils.get_paths(dataset,
model,