def chinese_whispers_fast(feats, prefix, name, knn_method, knn, th_sim, iters,
**kwargs):
""" Chinese Whispers Clustering Algorithm
Paper: Chinese whispers: an efficient graph clustering algorithm
and its application to natural language processing problems.
This implementation follows the matrix operation as described in Figure.4
int the paper. We switch the `maxrow` and `D^{t-1} * A_G` to make it
easier for post-processing.
The current result is inferior to `chinese_whispers` as it lacks of the
random mechanism as the iterative algorithm. The paper introduce two
operations to tackle this issue, namely `random mutation` and `keep class`.
However, it is not very clear how to set this two hyper-parameters.
"""
assert len(feats) > 1
with Timer('create graph'):
knn_prefix = os.path.join(prefix, 'knns', name)
knns = build_knns(knn_prefix, feats, knn_method, knn)
dists, nbrs = knns2ordered_nbrs(knns, sort=True)
spmat = fast_knns2spmat(knns, knn, th_sim, use_sim=True)
A = build_symmetric_adj(spmat, self_loop=False)
node_num = len(feats)
edge_num = A.nnz
print('#nodes: {}, #edges: {}'.format(node_num, edge_num))
with Timer('whisper iteratively (iters={})'.format(iters)):
D = identity(node_num)
for _ in range(iters):
D = D * A # it is equal to D.dot(A)
D = _maxrow(D, node_num)
assert D.nnz == node_num
clusters = {}
assigned_clusters = D.tocoo().col
for (node, assigned_cluster) in enumerate(assigned_clusters):
if assigned_cluster not in clusters:
clusters[assigned_cluster] = []
clusters[assigned_cluster].append(node)
print('#cluster: {}'.format(len(clusters)))
labels = clusters2labels(clusters.values())
labels = list(labels.values())
return labels
def __init__(self, cfg):
feat_path = cfg['feat_path']
label_path = cfg.get('label_path', None)
knn_graph_path = cfg.get('knn_graph_path', None)
self.k = cfg['k']
self.feature_dim = cfg['feature_dim']
self.is_norm_feat = cfg.get('is_norm_feat', True)
self.th_sim = cfg.get('th_sim', 0.)
self.max_conn = cfg.get('max_conn', 1)
self.ignore_ratio = cfg.get('ignore_ratio', 0.8)
self.ignore_small_confs = cfg.get('ignore_small_confs', True)
self.use_candidate_set = cfg.get('use_candidate_set', True)
self.nproc = cfg.get('nproc', 1)
self.max_qsize = cfg.get('max_qsize', int(1e5))
with Timer('read meta and feature'):
if label_path is not None:
self.lb2idxs, self.idx2lb = read_meta(label_path)
self.inst_num = len(self.idx2lb)
self.gt_labels = intdict2ndarray(self.idx2lb)
self.ignore_label = False
else:
self.inst_num = -1
self.ignore_label = True
self.features = read_probs(feat_path, self.inst_num,
self.feature_dim)
if self.is_norm_feat:
self.features = l2norm(self.features)
if self.inst_num == -1:
self.inst_num = self.features.shape[0]
self.size = self.inst_num
assert self.size == self.features.shape[0]
print('feature shape: {}, k: {}, norm_feat: {}'.format(
self.features.shape, self.k, self.is_norm_feat))
with Timer('read knn graph'):
if knn_graph_path is not None:
knns = np.load(knn_graph_path)['data']
else:
prefix = osp.dirname(feat_path)
name = rm_suffix(osp.basename(feat_path))
# find root folder of `features`
prefix = osp.dirname(prefix)
knn_prefix = osp.join(prefix, 'knns', name)
knns = build_knns(knn_prefix, self.features, cfg.knn_method,
cfg.knn)
assert self.inst_num == len(knns), "{} vs {}".format(
self.inst_num, len(knns))
adj = fast_knns2spmat(knns, self.k, self.th_sim, use_sim=True)
# build symmetric adjacency matrix
adj = build_symmetric_adj(adj, self_loop=True)
self.adj = row_normalize(adj)
# convert knns to (dists, nbrs)
self.dists, self.nbrs = knns2ordered_nbrs(knns, sort=True)
if cfg.pred_confs != '':
print('read estimated confidence from {}'.format(
cfg.pred_confs))
self.confs = np.load(cfg.pred_confs)['pred_confs']
else:
print('use unsupervised density as confidence')
assert self.radius
from vegcn.confidence import density
self.confs = density(self.dists, radius=self.radius)
assert 0 <= self.ignore_ratio <= 1
if self.ignore_ratio == 1:
self.ignore_set = set(np.arange(len(self.confs)))
else:
num = int(len(self.confs) * self.ignore_ratio)
confs = self.confs
if not self.ignore_small_confs:
confs = -confs
self.ignore_set = set(np.argpartition(confs, num)[:num])
print(
'ignore_ratio: {}, ignore_small_confs: {}, use_candidate_set: {}'.
format(self.ignore_ratio, self.ignore_small_confs,
self.use_candidate_set))
print('#ignore_set: {} / {} = {:.3f}'.format(
len(self.ignore_set), self.inst_num,
1. * len(self.ignore_set) / self.inst_num))
with Timer('Prepare sub-graphs'):
# construct subgraphs with larger confidence
self.peaks = {i: [] for i in range(self.inst_num)}
self.dist2peak = {i: [] for i in range(self.inst_num)}
if self.nproc > 1:
# multi-process
import multiprocessing as mp
pool = mp.Pool(self.nproc)
results = []
num = int(self.inst_num / self.max_qsize) + 1
for i in tqdm(range(num)):
beg = int(i * self.max_qsize)
end = min(beg + self.max_qsize, self.inst_num)
lst = [j for j in range(beg, end)]
results.extend(
list(
tqdm(pool.map(self.get_subgraph, lst),
total=len(lst))))
pool.close()
pool.join()
else:
results = [
self.get_subgraph(i) for i in tqdm(range(self.inst_num))
]
self.adj_lst = []
self.feat_lst = []
self.lb_lst = []
self.subset_gt_labels = []
self.subset_idxs = []
self.subset_nbrs = []
self.subset_dists = []
for result in results:
if result is None:
continue
elif len(result) == 3:
i, nbr, dist = result
self.peaks[i].extend(nbr)
self.dist2peak[i].extend(dist)
continue
i, nbr, dist, feat, adj, lb = result
self.subset_idxs.append(i)
self.subset_nbrs.append(nbr)
self.subset_dists.append(dist)
self.feat_lst.append(feat)
self.adj_lst.append(adj)
if not self.ignore_label:
self.subset_gt_labels.append(self.idx2lb[i])
self.lb_lst.append(lb)
self.subset_gt_labels = np.array(self.subset_gt_labels)
self.size = len(self.feat_lst)
assert self.size == len(self.adj_lst)
if not self.ignore_label:
assert self.size == len(self.lb_lst)
def __init__(self, cfg):
feat_path = cfg['feat_path']
label_path = cfg.get('label_path', None)
knn_graph_path = cfg.get('knn_graph_path', None)
self.k = cfg['k']
self.feature_dim = cfg['feature_dim']
self.is_norm_feat = cfg.get('is_norm_feat', True)
self.save_decomposed_adj = cfg.get('save_decomposed_adj', False)
self.th_sim = cfg.get('th_sim', 0.)
self.max_conn = cfg.get('max_conn', 1)
self.conf_metric = cfg.get('conf_metric')
with Timer('read meta and feature'):
if label_path is not None:
self.lb2idxs, self.idx2lb = read_meta(label_path)
self.inst_num = len(self.idx2lb)
self.gt_labels = intdict2ndarray(self.idx2lb)
self.ignore_label = False
else:
self.inst_num = -1
self.ignore_label = True
self.features = read_probs(feat_path, self.inst_num,
self.feature_dim)
if self.is_norm_feat:
self.features = l2norm(self.features)
if self.inst_num == -1:
self.inst_num = self.features.shape[0]
self.size = 1 # take the entire graph as input
with Timer('read knn graph'):
if os.path.isfile(knn_graph_path):
knns = np.load(knn_graph_path)['data']
else:
if knn_graph_path is not None:
print('knn_graph_path does not exist: {}'.format(
knn_graph_path))
prefix = osp.dirname(feat_path)
name = rm_suffix(osp.basename(feat_path))
# find root folder of `features`
prefix = osp.dirname(prefix)
knn_prefix = osp.join(prefix, 'knns', name)
knns = build_knns(knn_prefix, self.features, cfg.knn_method,
cfg.knn)
adj = fast_knns2spmat(knns, self.k, self.th_sim, use_sim=True)
# build symmetric adjacency matrix
adj = build_symmetric_adj(adj, self_loop=True)
adj = row_normalize(adj)
if self.save_decomposed_adj:
adj = sparse_mx_to_indices_values(adj)
self.adj_indices, self.adj_values, self.adj_shape = adj
else:
self.adj = adj
# convert knns to (dists, nbrs)
self.dists, self.nbrs = knns2ordered_nbrs(knns)
print('feature shape: {}, k: {}, norm_feat: {}'.format(
self.features.shape, self.k, self.is_norm_feat))
if not self.ignore_label:
with Timer('Prepare ground-truth label'):
self.labels = confidence(feats=self.features,
dists=self.dists,
nbrs=self.nbrs,
metric=self.conf_metric,
idx2lb=self.idx2lb,
lb2idxs=self.lb2idxs)
if cfg.eval_interim:
_, self.peaks = confidence_to_peaks(
self.dists, self.nbrs, self.labels, self.max_conn)
def __init__(self, cfg):
feat_path = cfg['feat_path']
label_path = cfg.get('label_path', None)
knn_graph_path = cfg.get('knn_graph_path', None)
self.k = cfg['k']
self.feature_dim = cfg['feature_dim']
self.is_norm_feat = cfg.get('is_norm_feat', True)
self.save_decomposed_adj = cfg.get('save_decomposed_adj', False)
self.th_sim = cfg.get('th_sim', 0.)
self.max_conn = cfg.get('max_conn', 1)
self.conf_metric = cfg.get('conf_metric')
self.num_process = cfg.get('num_process',16)
with Timer('read meta and feature'):
if label_path is not None:
self.lb2idxs, self.idx2lb = read_meta(label_path)
self.inst_num = len(self.idx2lb)
self.gt_labels = intdict2ndarray(self.idx2lb)
self.ignore_label = False
else:
self.inst_num = -1
self.ignore_label = True
self.features = read_probs(feat_path, self.inst_num,
self.feature_dim)
if self.is_norm_feat:
self.features = l2norm(self.features)
if self.inst_num == -1:
self.inst_num = self.features.shape[0]
self.size = 1 # take the entire graph as input
with Timer('read knn graph'):
if os.path.isfile(knn_graph_path):
knns = np.load(knn_graph_path)['data'] # num_imgs*2*k
else:
if knn_graph_path is not None:
print('knn_graph_path does not exist: {}'.format(
knn_graph_path))
knn_prefix = os.path.join(cfg.prefix, 'knns', cfg.name)
# 通过faiss实现k近邻搜索，此处作者faiss_gpu版本实现可能有问题，但faiss大规模在cpu上跑还是慢
# 当然faiss有针内存和计算速度方面的优化，PQ,IVF等，可参考faiss
knns = build_knns(knn_prefix, self.features, cfg.knn_method,
cfg.knn,self.num_process)
# 依据k近邻搜索结果构建邻接矩阵
adj = fast_knns2spmat(knns, self.k, self.th_sim, use_sim=True)
# build symmetric adjacency matrix
adj = build_symmetric_adj(adj, self_loop=True)
adj = row_normalize(adj)
if self.save_decomposed_adj:
adj = sparse_mx_to_indices_values(adj)
self.adj_indices, self.adj_values, self.adj_shape = adj
else:
self.adj = adj
# convert knns to (dists, nbrs)
self.dists, self.nbrs = knns2ordered_nbrs(knns) # num_imgs*k
print('feature shape: {}, k: {}, norm_feat: {}'.format(
self.features.shape, self.k, self.is_norm_feat))
if not self.ignore_label:
with Timer('Prepare ground-truth label'):
self.labels = confidence(feats=self.features,
dists=self.dists,
nbrs=self.nbrs,
metric=self.conf_metric,
idx2lb=self.idx2lb,
lb2idxs=self.lb2idxs)
if cfg.eval_interim:
_, self.peaks = confidence_to_peaks(
self.dists, self.nbrs, self.labels, self.max_conn)
def train_gcn(model, cfg, logger):
# prepare dataset
for k, v in cfg.model['kwargs'].items():
setattr(cfg.train_data, k, v)
dataset = build_dataset(cfg.model['type'], cfg.train_data)
pre_features = torch.FloatTensor(dataset.features)
print('Have loaded the training data.')
inst_num = dataset.inst_num
feature_dim = dataset.feature_dim
lb2idxs = dataset.lb2idxs
center_fea = dataset.center_fea.astype('float32')
cls_num, dim = center_fea.shape
labels = torch.LongTensor(dataset.gt_labels)
HEAD1 = HEAD(nhid=512)
HEAD_test1 = HEAD_test(nhid=512)
#load parameters from the pretrained model
#model.load_state_dict(torch.load('./'))
#HEAD1.load_state_dict(torch.load('./'), False)
OPTIMIZER = optim.SGD([{'params': model.parameters(),'weight_decay':1e-5},
{'params': HEAD1.parameters(),'weight_decay':1e-5}], lr=0.01, momentum=0.9)
print('the learning rate is 0.01')
#model.load_state_dict(torch.load(''))
#HEAD1.load_state_dict(torch.load(''))
print("have load the pretrained model.")
cfg.cuda = True
model = model.cuda()
HEAD1 = HEAD1.cuda()
MODEL_ROOT = './src/train_model'
print('the model save path is', MODEL_ROOT)
#prepare the test data
target = "part1_test"
knn_path = "./data/knns/" + target + "/faiss_k_80.npz"
knns = np.load(knn_path, allow_pickle=True)['data']
inst_num = knns.shape[0]
k_num = knns.shape[2]
nbrs = knns[:, 0, :]
pair_a = []
pair_b = []
for i in range(inst_num):
pair_a.extend([i] * k_num)
pair_b.extend(nbrs[i])
for epoch in range(cfg.total_epochs):
if epoch == cfg.STAGES[0]: # adjust LR for each training stage after warm up, you can also choose to adjust LR manually (with slight modification) once plaueau observed
schedule_lr(OPTIMIZER)
if epoch == cfg.STAGES[1]:
schedule_lr(OPTIMIZER)
if epoch == cfg.STAGES[2]:
schedule_lr(OPTIMIZER)
model.train()
HEAD1.train()
index = faiss.IndexFlatIP(dim)
index.add(center_fea)
sims, cluster_id = index.search(center_fea, k=(cfg.cluster_num+200)) # search for the k-10 neighbor
#sims, cluster_id = index.search(center_fea, k=cfg.cluster_num) # search for the k-10 neighbor
print('Have selected the cluster ids.')
for batch in range(cls_num):
#for batch in range(20):
#0.select ids
sample_cluster_id = random.sample(list(cluster_id[batch]), cfg.cluster_num)
#sample_cluster_id = list(cluster_id[batch])
sample_id = []#the idx of the samples in this batch
for i in range(len(sample_cluster_id)):
sample_id.extend(random.sample(lb2idxs[sample_cluster_id[i]],int(len(lb2idxs[sample_cluster_id[i]])*0.9)))
#sample_id.extend(lb2idxs[sample_cluster_id[i]])
#sample_id.sort()
sample_num =len(sample_id)
#id = list(np.arange(0,sample_num,1))
#sample2sort = dict(zip(sample_id, id))
if (sample_num>100000)|(sample_num<100):
print('[too much samples] continue.')
continue
#1.create selected labels and images
batch_labels = labels[sample_id]
feature = pre_features[sample_id]
print(sample_num)
#2.create knn for this batch
with Timer('build knn:'):
knn_prefix = os.path.join("./data/rebuild_knn")
if not os.path.exists(knn_prefix):
os.makedirs(knn_prefix)
if os.path.exists(os.path.join(knn_prefix, 'faiss_k_80.npz')):
os.remove(os.path.join(knn_prefix, 'faiss_k_80.npz'))
if os.path.exists(os.path.join(knn_prefix, 'faiss_k_80.index')):
os.remove(os.path.join(knn_prefix, 'faiss_k_80.index'))
knns = build_knns(knn_prefix,
#l2norm(feature.clone().detach().cpu().numpy()),
l2norm(feature.numpy()),
"faiss",
80,
is_rebuild=True)
batch_adj = fast_knns2spmat(knns, 80, 0, use_sim=True)
batch_adj = build_symmetric_adj(batch_adj, self_loop=True)
batch_adj = row_normalize(batch_adj)
batch_adj = sparse_mx_to_torch_sparse_tensor(batch_adj, return_idx=False)
#3.put selected feature and labels to cuda
batch_labels = batch_labels.cuda()
feature = feature.cuda()
batch_adj = batch_adj.cuda()
train_data = [feature, batch_adj, batch_labels]
#x = model(train_data)
#4.train the model
#add
train_id_inst = batch_adj._indices().size()[1]
#print('train_id_inst:', train_id_inst)
#print('sample_num:', sample_num)
#train_id_inst = sample_num
rad_id = random.sample(range(0, train_id_inst), train_id_inst)+random.sample(range(0, train_id_inst), train_id_inst)
patch_num = 40
for i in range(patch_num*2):
id = rad_id[i * int(train_id_inst / patch_num):(i + 1) * int(train_id_inst / patch_num)]
x = model(train_data)
loss = HEAD1(x, train_data, id)
OPTIMIZER.zero_grad()
loss.backward()
OPTIMIZER.step()
print(datetime.datetime.now())
print('epoch:{}/{}, batch:{}/{}, batch2:{}/{},loss:{}'.format(epoch, cfg.total_epochs, batch, cls_num, i, patch_num*2, loss))
if (batch+1)%100==0:
if not os.path.exists(MODEL_ROOT):
os.makedirs(MODEL_ROOT)
print('save model in epoch:{} batch:{} to {}'.format(epoch, batch, MODEL_ROOT))
torch.save(model.state_dict(), os.path.join(MODEL_ROOT, "Backbone_Epoch_{}_batch_{}.pth".format(epoch + 1, batch)))
torch.save(HEAD1.state_dict(), os.path.join(MODEL_ROOT, "Head_Epoch_{}_batch_{}.pth".format(epoch + 1, batch)))
if (batch + 1) % 300 == 0:
avg_acc = perform_val(model, HEAD1, HEAD_test1, cfg, feature_dim, pair_a, pair_b)
print('the avg testing acc in epoch:{} batch:{} is :'.format(epoch,batch), avg_acc)
model.train()
HEAD1.train()
#5.test
avg_acc = perform_val(model, HEAD1, HEAD_test1, cfg, feature_dim, pair_a, pair_b)
print('the avg testing acc in epoch:{} batch:{} is :'.format(epoch,batch), avg_acc)
# 6.save model
if not os.path.exists(MODEL_ROOT):
os.makedirs(MODEL_ROOT)
print('save model in epoch:{} batch:{} to {}'.format(epoch, batch, MODEL_ROOT))
torch.save(model.state_dict(), os.path.join(MODEL_ROOT, "Backbone_Epoch_{}_batch_{}.pth".format(epoch + 1, batch)))
torch.save(HEAD1.state_dict(), os.path.join(MODEL_ROOT, "Head_Epoch_{}_batch_{}.pth".format(epoch + 1, batch)))
def __init__(self, cfg):
feat_path = cfg['feat_path']
label_path = cfg.get('label_path', None)
knn_graph_path = cfg.get('knn_graph_path', None)
self.k = cfg['k']
self.feature_dim = cfg['feature_dim']
self.is_norm_feat = cfg.get('is_norm_feat', True)
self.save_decomposed_adj = cfg.get('save_decomposed_adj', False)
self.th_sim = cfg.get('th_sim', 0.)
self.conf_metric = cfg.get('conf_metric')
with Timer('read meta and feature'):
if label_path is not None:
self.lb2idxs, self.idx2lb = read_meta(label_path)
self.inst_num = len(self.idx2lb)
self.cls_num = len(self.lb2idxs)
self.gt_labels = intdict2ndarray(self.idx2lb)
self.ignore_label = False
else:
self.inst_num = -1
self.ignore_label = True
self.features = read_probs(feat_path, self.inst_num,
self.feature_dim)
if self.is_norm_feat:
self.features = l2norm(self.features)
if self.inst_num == -1:
self.inst_num = self.features.shape[0]
self.size = 1 # take the entire graph as input
with Timer('Compute center feature'):
self.center_fea = np.zeros((self.cls_num, self.features.shape[1]))
for i in range(self.cls_num):
self.center_fea[i] = np.mean(self.features[self.lb2idxs[i]], 0)
self.center_fea = l2norm(self.center_fea)
with Timer('read knn graph'):
if os.path.isfile(knn_graph_path):
print("load knns from the knn_path")
self.knns = np.load(knn_graph_path)['data']
else:
if knn_graph_path is not None:
print('knn_graph_path does not exist: {}'.format(
knn_graph_path))
knn_prefix = os.path.join(cfg.prefix, 'knns', cfg.name)
self.knns = build_knns(knn_prefix, self.features,
cfg.knn_method, cfg.knn)
adj = fast_knns2spmat(self.knns, self.k, self.th_sim, use_sim=True)
# build symmetric adjacency matrix
adj = build_symmetric_adj(adj, self_loop=True)
#print('adj before norm')
#print(adj)
adj = row_normalize(adj)
if self.save_decomposed_adj:
adj = sparse_mx_to_indices_values(adj)
self.adj_indices, self.adj_values, self.adj_shape = adj
else:
self.adj = adj
# convert knns to (dists, nbrs)
self.dists, self.nbrs = knns2ordered_nbrs(self.knns)
print('feature shape: {}, k: {}, norm_feat: {}'.format(
self.features.shape, self.k, self.is_norm_feat))