def plot_sift_upto():
for i in [4000]:
glove_data = utils.load_sift_data('train')
#means = glove_data.mean(0)
#glove_data -= means
#glove_data = glove_data / (glove_data**2).sum(-1, keepdim=True).sqrt()
glove_q = utils.load_sift_data('query')[:300]
#glove_q -= means
utils.plot_dist_hist_upto(glove_data, glove_q, i, 'sift')
#plt.clf()
for i in [4000]:
glove_c_data = utils.load_sift_c_data('train')
glove_c_q = utils.load_sift_c_data('query')[:300]
pdb.set_trace()
utils.plot_dist_hist_upto(glove_c_data, glove_c_q, i, 'sift_c')
from kahip.kmkahip import run_kmkahip
if __name__ == '__main__':
opt = utils.parse_args()
#adjust the number of parts and the height of the hierarchy
n_cluster_l = [2]
height_l = [1]
# load dataset
if opt.glove:
dataset = utils.load_glove_data('train').to(utils.device)
queryset = utils.load_glove_data('query').to(utils.device)
neighbors = utils.load_glove_data('answers').to(utils.device)
elif opt.sift:
dataset = utils.load_sift_data('train').to(utils.device)
queryset = utils.load_sift_data('query').to(utils.device)
neighbors = utils.load_sift_data('answers').to(utils.device)
else:
dataset = utils.load_data('train').to(utils.device)
queryset = utils.load_data('query').to(utils.device)
neighbors = utils.load_data('answers').to(utils.device)
#specify which action to take at each level, actions can be km, kahip, train, or svm. Lower keys indicate closer to leaf.
#Note that if 'kahip' is included, evaluation must be on training rather than test set, since partitioning was performed on training, but not test, set.
#e.g.: opt.level2action = {0:'km', 1:'train', 3:'train'}
opt.level2action = {0: 'train'}
for n_cluster in n_cluster_l:
print('n_cluster {}'.format(n_cluster))
opt.n_clusters = n_cluster
def compute_degrees_distr():
dataset = utils.load_sift_data('train').to(utils.device)
#dataset = utils.load_glove_data('train').to(utils.device)
distr = utils.compute_degree_distr(dataset, 10)
print(distr[:30])
pdb.set_trace()