def evaluate(self, timestamp, batch_idx, run_kmeans=False):
data_y, label_y, _ = self.y_sampler.load_all()
#data_y, label_y = self.y_sampler.tst_data, self.y_sampler.tst_label
N = data_y.shape[0]
data_x_, data_x_onehot_ = self.predict_x(data_y)
np.savez('{}/data_at_{}.npz'.format(self.save_dir, batch_idx + 1),
data_x_, data_x_onehot_, label_y)
label_infer = np.argmax(data_x_onehot_, axis=1)
purity = metric.compute_purity(label_infer, label_y)
nmi = normalized_mutual_info_score(label_y, label_infer)
ari = adjusted_rand_score(label_y, label_infer)
#self.cluster_heatmap(batch_idx, label_infer, label_y)
print('RTM: Purity = {}, NMI = {}, ARI = {}'.format(purity, nmi, ari))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write('%.4f\t%.4f\t%.4f\t%d\n' % (purity, nmi, ari, batch_idx))
f.close()
#k-means
if run_kmeans:
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_y)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y)
nmi = normalized_mutual_info_score(label_y, label_kmeans)
ari = adjusted_rand_score(label_y, label_kmeans)
print('K-means: Purity = {}, NMI = {}, ARI = {}'.format(
purity, nmi, ari))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write('%.4f\t%.4f\t%.4f\n' % (purity, nmi, ari))
f.close()
def evaluate(self, timestamp, epoch, run_kmeans=False):
data_y1, label_y1, data_y2, label_y2 = self.y_sampler.load_all()
data_x1_, data_x_onehot1_ = self.predict_x1(data_y1)
data_x2_, data_x_onehot2_ = self.predict_x2(data_y2)
np.savez('{}/data_at_{}.npz'.format(self.save_dir, epoch + 1),
data_x1_, data_x_onehot1_, label_y1, data_x2_,
data_x_onehot2_, label_y2)
#scRNA-seq
label_infer1 = np.argmax(data_x_onehot1_, axis=1)
purity1 = metric.compute_purity(label_infer1, label_y1)
nmi1 = normalized_mutual_info_score(label_y1, label_infer1)
ari1 = adjusted_rand_score(label_y1, label_infer1)
self.cluster_heatmap(epoch, label_infer1, label_y1, 'scRNA')
print('CoupleRTM scRNA-seq: NMI = {}, ARI = {}, Purity = {}'.format(
nmi1, ari1, purity1))
#scATAC-seq
label_infer2 = np.argmax(data_x_onehot2_, axis=1)
purity2 = metric.compute_purity(label_infer2, label_y2)
nmi2 = normalized_mutual_info_score(label_y2, label_infer2)
ari2 = adjusted_rand_score(label_y2, label_infer2)
self.cluster_heatmap(epoch, label_infer2, label_y2, 'scATAC')
print('CoupleRTM scATAC-seq: NMI = {}, ARI = {}, Purity = {}'.format(
nmi2, ari2, purity2))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write('%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%d\n' %
(nmi1, ari1, purity1, nmi2, ari2, purity2, epoch))
f.close()
#k-means
if run_kmeans:
#scRNA-seq
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_y1)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y1)
nmi = normalized_mutual_info_score(label_y1, label_kmeans)
ari = adjusted_rand_score(label_y1, label_kmeans)
print('K-means scRNA-seq: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write(
'K-means scRNA-seq: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
#scATAC-seq
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_y2)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y2)
nmi = normalized_mutual_info_score(label_y2, label_kmeans)
ari = adjusted_rand_score(label_y2, label_kmeans)
print(
'K-means scATAC-seq: NMI = {}, ARI = {}, Purity = {}\n'.format(
nmi, ari, purity))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write(
'K-means scATAC-seq: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
f.close()
def evaluate(self, timestamp, epoch, run_kmeans=False):
data_y, label_y = self.y_sampler.load_all()
N = data_y.shape[0]
data_x_, data_x_onehot_ = self.predict_x(data_y)
np.savez('{}/data_at_{}.npz'.format(self.save_dir, epoch + 1), data_x_,
data_x_onehot_, label_y)
label_infer = np.argmax(data_x_onehot_, axis=1)
purity = metric.compute_purity(label_infer, label_y)
nmi = normalized_mutual_info_score(label_y, label_infer)
ari = adjusted_rand_score(label_y, label_infer)
self.cluster_heatmap(epoch, label_infer, label_y)
print('RTM: NMI = {}, ARI = {}, Purity = {}'.format(nmi, ari, purity))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write('RTM\t%.4f\t%.4f\t%.4f\t%d\n' % (nmi, ari, purity, epoch))
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_x_)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y)
nmi = normalized_mutual_info_score(label_y, label_kmeans)
ari = adjusted_rand_score(label_y, label_kmeans)
print('Latent-kmeans: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
f.write('Latent-kmeans\t%.4f\t%.4f\t%.4f\t%d\n' %
(nmi, ari, purity, epoch))
#k-means
if run_kmeans:
from sklearn.decomposition import PCA
pca = PCA(n_components=10)
pca.fit(data_y)
data_pca_y = pca.fit_transform(data_y)
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_pca_y)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y)
nmi = normalized_mutual_info_score(label_y, label_kmeans)
ari = adjusted_rand_score(label_y, label_kmeans)
print('PCA + K-means: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
f.write('PCA+Kmeans%.4f\t%.4f\t%.4f\n' % (nmi, ari, purity))
km = KMeans(n_clusters=nb_classes, random_state=0).fit(data_y)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, label_y)
nmi = normalized_mutual_info_score(label_y, label_kmeans)
ari = adjusted_rand_score(label_y, label_kmeans)
print('K-means: NMI = {}, ARI = {}, Purity = {}'.format(
nmi, ari, purity))
f.write('Kmeans%.4f\t%.4f\t%.4f\n' % (nmi, ari, purity))
f.close()
def correlation(self, X, Y, heatmap=False):
nb_classes = len(set(Y))
print nb_classes
km = KMeans(n_clusters=nb_classes, random_state=0).fit(X)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, Y)
nmi = normalized_mutual_info_score(Y, label_kmeans)
ari = adjusted_rand_score(Y, label_kmeans)
homogeneity = homogeneity_score(Y, label_kmeans)
ami = adjusted_mutual_info_score(Y, label_kmeans)
print('NMI = {}, ARI = {}, Purity = {},AMI = {}, Homogeneity = {}'.
format(nmi, ari, purity, ami, homogeneity))
if heatmap:
x_ticks = [''] * len(Y)
y_ticks = [''] * len(Y)
idx = []
for i in range(nb_classes):
sub_idx = [j for j, item in enumerate(Y) if item == i]
idx += [j for j, item in enumerate(Y) if item == i]
x_ticks[len(idx) - 1] = str(i)
assert len(idx) == len(Y)
X = X[idx, :]
Y = Y[idx]
#similarity_mat = pairwise_distances(X,metric='cosine')
similarity_mat = cosine_similarity(X)
#sns.heatmap(similarity_mat,cmap='Blues')
fig, ax = plt.subplots()
#ax.set_yticks(range(len(y_ticks)))
ax.set_yticklabels(y_ticks)
ax.set_xticks(range(len(x_ticks)))
ax.set_xticklabels(x_ticks)
im = ax.imshow(similarity_mat, cmap='Blues')
plt.colorbar(im)
plt.savefig('heatmap_%s_dim%d.png' % (self.name, X.shape[1]),
dpi=600)
def _eval_cluster(self, latent_rep, labels_true, timestamp, val):
map_labels = {0: 0, 1: 1, 2: 2, 4: 3, 6: 4, 7: 5, 8: 6, 9: 7}
labels_true = np.array([map_labels[i] for i in labels_true])
km = KMeans(n_clusters=max(self.num_classes,
len(np.unique(labels_true))),
random_state=0).fit(latent_rep)
labels_pred = km.labels_
purity = metric.compute_purity(labels_pred, labels_true)
ari = adjusted_rand_score(labels_true, labels_pred)
nmi = normalized_mutual_info_score(labels_true, labels_pred)
if val:
data_split = 'Validation'
else:
data_split = 'Test'
#data_split = 'All'
print('Data = {}, Model = {}, sampler = {}, z_dim = {}, beta_reg = {}'.
format(self.data, self.model, self.sampler, self.z_dim,
self.beta_reg))
print(' #Points = {}, K = {}, Purity = {}, NMI = {}, ARI = {}, '.
format(latent_rep.shape[0], self.num_classes, purity, nmi, ari))
with open('logs/Res_{}_{}.txt'.format(self.data, self.model),
'a+') as f:
f.write(
'{}, {} : K = {}, z_dim = {}, beta_reg = {}, sampler = {}, Purity = {}, NMI = {}, ARI = {}\n'
.format(timestamp, data_split, self.num_classes, self.z_dim,
self.beta_reg, self.sampler, purity, nmi, ari))
f.flush()
def evaluate(self,timestamp,batch_idx):
if has_label:
data_y, label_y = self.y_sampler.load_all()
else:
data_y = self.y_sampler.load_all()
data_x_, data_x_onehot_ = self.predict_x(data_y)
label_infer = np.argmax(data_x_onehot_, axis=1)
print([list(label_infer).count(item) for item in np.unique(label_infer)])
print([list(label_y).count(item) for item in np.unique(label_y)])
# label_kmeans = KMeans(n_clusters=4, n_init = 200).fit_predict(data_y)
# ari = adjusted_rand_score(label_y, label_kmeans)
# nmi = normalized_mutual_info_score(label_y, label_kmeans)
# print('kmeans',ari,nmi,self.nb_classes)
# sys.exit()
if has_label:
purity = metric.compute_purity(label_infer, label_y)
nmi = normalized_mutual_info_score(label_y, label_infer)
ari = adjusted_rand_score(label_y, label_infer)
h**o = homogeneity_score(label_y,label_infer)
print('scDEC: NMI = {}, ARI = {}, Homogeneity = {}'.format(nmi,ari,h**o))
if is_train:
f = open('%s/log.txt'%self.save_dir,'a+')
f.write('NMI = {}\tARI = {}\tHomogeneity = {}\t batch_idx = {}\n'.format(nmi,ari,h**o,batch_idx))
f.close()
np.savez('{}/data_at_{}.npz'.format(self.save_dir, batch_idx+1),data_x_,data_x_onehot_,label_y)
else:
np.savez('results/{}/data_pre.npz'.format(self.data),data_x_,data_x_onehot_,label_y)
else:
if is_train:
np.savez('{}/data_at_{}.npz'.format(self.save_dir, batch_idx+1),data_x_,data_x_onehot_)
else:
np.savez('results/{}/data_pre.npz'.format(self.data),data_x_,data_x_onehot_)
def analyze_cistopic():
ratio = 0.7
df = pd.read_csv('cisTopic/results/%s' % path,
sep=' ',
header=0,
index_col=[0])
data = df.values
data = data * 1.0 / np.sum(data, axis=0)
data = data.T
#mouse atlas
labels = [
item for item in open('datasets/scATAC/Mouse_atlas/label_%s.txt' %
str(ratio)).readlines()
]
#forebrain
#labels = loadmat('datasets/scATAC/scATAC-seq_data_for_liuqiao/REN/original/cell_labels.mat')['cell_labels']
uniq_labels = list(np.unique(labels))
Y = np.array([uniq_labels.index(item) for item in labels])
#GMvsHL,
#name='GMvsHL'
#Y = np.load('datasets/scATAC/%s_label.npy'%name).astype('int64')
print(data.shape)
assert len(Y) == data.shape[0]
km = KMeans(n_clusters=len(uniq_labels), random_state=0).fit(data)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, Y)
nmi = normalized_mutual_info_score(Y, label_kmeans)
ari = adjusted_rand_score(Y, label_kmeans)
print(nmi, ari, purity)
def cluster_eval(labels_true, labels_infer):
purity = metric.compute_purity(labels_infer, labels_true)
nmi = normalized_mutual_info_score(labels_true, labels_infer)
ari = adjusted_rand_score(labels_true, labels_infer)
homogeneity = homogeneity_score(labels_true, labels_infer)
ami = adjusted_mutual_info_score(labels_true, labels_infer)
#print('NMI = {}, ARI = {}, Purity = {},AMI = {}, Homogeneity = {}'.format(nmi,ari,purity,ami,homogeneity))
return nmi, ari, homogeneity
def correlation(self, X, Y, heatmap=False):
nb_classes = len(set(Y))
print(nb_classes)
km = KMeans(n_clusters=nb_classes, random_state=0).fit(X)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, Y)
nmi = normalized_mutual_info_score(Y, label_kmeans)
ari = adjusted_rand_score(Y, label_kmeans)
homogeneity = homogeneity_score(Y, label_kmeans)
ami = adjusted_mutual_info_score(Y, label_kmeans)
print('NMI = {}, ARI = {}, Purity = {},AMI = {}, Homogeneity = {}'.
format(nmi, ari, purity, ami, homogeneity))
def eval_cluster(pathname1, labels_pred, labels_true, no_of_spk, timestamp, z_dim, sampler, beta_cycle_label, beta_cycle_gen):
purity = metric.compute_purity(labels_pred, labels_true)
ari = adjusted_rand_score(labels_true, labels_pred)
nmi = normalized_mutual_info_score(labels_true, labels_pred)
print(' #Points = {}, K = {}, Purity = {}, NMI = {}, ARI = {},'.format(labels_pred.shape[0], no_of_spk, purity, nmi, ari))
with open(pathname1 + '/Result.txt', 'a+') as f:
f.write('{}, K = {}, z_dim = {}, beta_label = {}, beta_gen = {}, sampler = {}, Purity = {}, NMI = {}, ARI = {}\n'
.format(timestamp, no_of_spk, z_dim, beta_cycle_label, beta_cycle_gen,
sampler, purity, nmi, ari))
f.flush()
def cluster_eval(labels_true, latent_feat):
assert len(labels_true) == latent_feat.shape[0]
n_clusters = len(np.unique(labels_true))
km = KMeans(n_clusters=n_clusters, n_init=20,
random_state=0).fit(latent_feat)
label_kmeans = km.labels_
confusion_mat(labels_true, label_kmeans)
purity = metric.compute_purity(label_kmeans, labels_true)
nmi = normalized_mutual_info_score(labels_true, label_kmeans)
ari = adjusted_rand_score(labels_true, label_kmeans)
homogeneity = homogeneity_score(labels_true, label_kmeans)
ami = adjusted_mutual_info_score(labels_true, label_kmeans)
print('NMI = {}, ARI = {}, Purity = {},AMI = {}, Homogeneity = {}'.format(
nmi, ari, purity, ami, homogeneity))
def evaluate(self, timestamp, batch_idx, run_kmeans=False):
data_y, label_y = self.y_sampler.load_all()
N = data_y.shape[0]
data_x_, data_x_onehot_ = self.predict_x(data_y)
np.savez('{}/data_at_{}.npz'.format(self.save_dir, batch_idx + 1),
data_x_, data_x_onehot_, label_y)
label_infer = np.argmax(data_x_onehot_, axis=1)
purity = metric.compute_purity(label_infer, label_y)
nmi = normalized_mutual_info_score(label_y, label_infer)
ari = adjusted_rand_score(label_y, label_infer)
#print('scDEC: NMI = {}, ARI = {}, Purity = {}'.format(nmi,ari,purity))
f = open('%s/log.txt' % self.save_dir, 'a+')
f.write('NMI = {}\tARI = {}\tPurity = {}\t batch_idx = {}\n'.format(
nmi, ari, purity, batch_idx))
f.close()
def analyze_SCALE():
ratio = 0.1
data = np.loadtxt('SCALE/%s/feature.txt' % path,
delimiter='\t',
usecols=range(1, 1 + 10))
labels = [
item.strip()
for item in open('datasets/scATAC/Mouse_atlas/label_%s.txt' %
str(ratio)).readlines()
]
uniq_labels = list(np.unique(labels))
Y = np.array([uniq_labels.index(item) for item in labels])
print(data.shape)
assert len(Y) == data.shape[0]
km = KMeans(n_clusters=len(uniq_labels), random_state=0).fit(data)
label_kmeans = km.labels_
purity = metric.compute_purity(label_kmeans, Y)
nmi = normalized_mutual_info_score(Y, label_kmeans)
ari = adjusted_rand_score(Y, label_kmeans)
print(nmi, ari, purity)
plot_embedding(data,
np.array(labels),
save='SCALE/embedding_tsne_cont_%s.png' % path)