def main():
features = np.load('0_feat.npy')
labels = np.load('0_label.npy')
num_instances = 32
batch_size = 128
Batch = BatchGenerator(labels,
num_instances=num_instances,
batch_size=batch_size)
batch = Batch.batch()
inputs = Variable(torch.FloatTensor(features[batch, :])).cuda()
targets = Variable(torch.LongTensor(labels[batch])).cuda()
print(KmeanLoss(n_cluster=32)(inputs, targets))
def main():
features = np.load('0_feat.npy')
labels = np.load('0_label.npy')
centers, center_labels = cluster_(features, labels, n_clusters=3)
centers = Variable(torch.FloatTensor(centers).cuda(), requires_grad=True)
center_labels = Variable(torch.LongTensor(center_labels)).cuda()
cluster_counter = np.zeros([100, 3])
num_instances = 3
batch_size = 120
Batch = BatchGenerator(labels,
num_instances=num_instances,
batch_size=batch_size)
batch = Batch.batch()
# _mask = Variable(torch.ByteTensor(np.ones([num_class_dict[args.data], args.n_cluster])).cuda())
_mask = Variable(torch.ByteTensor(np.ones([100, 3])).cuda())
inputs = Variable(torch.FloatTensor(features[batch, :])).cuda()
targets = Variable(torch.LongTensor(labels[batch])).cuda()
# print(torch.mean(inputs))
mca = MCALoss(alpha=16,
centers=centers,
center_labels=center_labels,
cluster_counter=cluster_counter)
for i in range(2):
# loss, accuracy, dist_ap, dist_an =
# MCALoss(alpha=16, centers=centers, center_labels=center_labels)(inputs, targets)
loss, accuracy, dist_ap, dist_an = \
mca(inputs, targets, _mask)
# print(loss.data[0])
loss.backward()
# print(centers.grad.data)
centers.data -= centers.grad.data
centers.grad.data.zero_()
# print(centers.grad)
print(cluster_counter)