def FW_within_catergory(data_, cid_, all_pid_, all_did_):
W_class_ = np.asarray([2, 6, 7, 8, 9, 10]) + offset_s2s_category_
FW_clustering_within_category_F_ = {
0: [8, 9, 11],
1: [20],
2: [1, 15],
3: [3, 21],
4: [10, 12],
5: [13, 14, 16, 17, 18, 19]
}
FW_ = {}
for ii_ in range(len(W_class_)):
# S2S
temp_W_category = W_class_[ii_]
temp_W_idx = np.where(cid_ == temp_W_category)[0]
temp_W_feat = data_[temp_W_idx, :]
# DF
temp_F_categories = FW_clustering_within_category_F_[ii_]
temp_F_idx = []
for jj_ in temp_F_categories:
temp_idx_ = np.where(cid_ == jj_)[0]
temp_F_idx = np.concatenate((temp_F_idx, temp_idx_)).astype(int)
temp_F_feat = data_[temp_F_idx, :]
temp_F_W_idx = np.concatenate((temp_F_idx, temp_W_idx)).astype(int)
temp_F_W_pid_ = all_pid_[temp_F_W_idx]
temp_F_W_did_ = all_did_[temp_F_W_idx]
feat_ = np.concatenate((temp_F_feat, temp_W_feat))
c, num_clust, req_c = FINCH(feat_,
initial_rank=None,
req_clust=None,
distance='cosine',
verbose=False) #verbose=True
FW_['{}'.format(W_class_[ii_])] = []
FW_['{}'.format(W_class_[ii_])].append({
'idx': temp_F_W_idx,
'c': c,
'nc': num_clust,
'pid_c': temp_F_W_pid_,
'did_c': temp_F_W_did_
})
return FW_
def W_within_catergory(data_, cid_):
W_clustering_within_category = np.asarray([0, 1, 3, 4, 5
]) + offset_s2s_category_
W_ = {}
for ii_ in W_clustering_within_category:
temp_idx = np.where(cid_ == ii_)[0]
feat_ = data_[temp_idx, :]
c, num_clust, req_c = FINCH(feat_,
initial_rank=None,
req_clust=None,
distance='cosine',
verbose=False) #verbose=True
W_['{}'.format(ii_)] = []
W_['{}'.format(ii_)].append({'idx': temp_idx, 'c': c, 'nc': num_clust})
return W_
t6 = time.time()
print("classification time = ", t6 - t5)
nu = 0
features_dict_incremental = OrderedDict()
class_to_process = []
p_max, i_max = torch.max(normalized_tensor, axis=1)
for k in range(normalized_tensor.shape[0]):
if i_max[k] == 0: # predicted unnkwon unknown
residual_dict[image_names[k]] = FV[k, :].numpy()
print(f"len(residual_dict) = {len(residual_dict)}")
if len(residual_dict) > 0:
image_names_residual, FVs_residual = zip(*residual_dict.items())
data = np.vstack(list(FVs_residual))
c_all, num_clust, req_c = FINCH(data)
modified_index_partition_selected = min(len(num_clust) - 1, index_partition_selected)
cluster_labels = c_all[:, modified_index_partition_selected]
number_of_clusters = num_clust[modified_index_partition_selected] # number of clusters after clustering.
for cluster_number in range(number_of_clusters): # number of clusters after clustering.
index = [iii for iii in range(len(cluster_labels)) if cluster_labels[iii] == cluster_number]
nu = nu + 1
class_number = int(nu + number_of_discovered_classes + number_of_known_classes)
features_dict_incremental[class_number] = torch.from_numpy(
np.array([FVs_residual[jjj] for jjj in index])
) # .double()#.cuda()
class_to_process.append(class_number)
if ((len(class_to_process) > 0) and (number_of_discovered_classes > 0)) or (len(class_to_process) > 2):
SCAIL.fit(features_dict_incremental)
FV_supervised, _ = cnn_model_supervised(x_supervised)
FV_moco_places, _ = cnn_model_moco_places(x_moco)
FV = torch.cat((FV_supervised, FV_moco_places), 1)
feature_storage[n : (n + FV.size(0)), 0] = y
feature_storage[n : (n + FV.size(0)), 1:] = FV.cpu()
image_list = image_list + list(A)
n = n + FV.size(0)
assert len(image_list) == n
data_ = feature_storage.detach().clone().cpu().numpy()
data_ = data_[:n, :]
feature = data_[:, 1:]
true_label = data_[:, 0]
c_all, num_clust, req_c = FINCH(feature)
cluster_labels = c_all[:, 1]
number_of_clusters = num_clust[1] # number of clusters after clustering.
storage_dict = dict()
for L, K, A in zip(true_label, cluster_labels, image_list):
if K in storage_dict:
storage_dict[K].append((L, A))
else:
storage_dict[K] = [(L, A)]
torch.save(storage_dict, f"/scratch/mjafarzadeh/clusters_all_b3_SP_{data_name}.pth")
del feature_storage, data_, storage_dict
t5 = time.time()
print(f"extracting {data_name} time = ", t5 - t4)