-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
215 lines (166 loc) · 7.73 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import time
import os
import datetime
import tensorflow as tf
import numpy as np
import scipy.sparse as sp
import scanpy as sc
from sklearn import metrics
from sklearn.metrics import roc_auc_score
from sklearn.metrics import average_precision_score
from sklearn.metrics import silhouette_score
from sklearn.metrics import mean_squared_error
from optimizer import OptimizerSCI
from input_data import load_data
from network import GraphSCI
from preprocessing import *
os.environ['CUDA_VISIBLE_DEVICES'] = "4"
import warnings
warnings.filterwarnings('ignore')
now = datetime.datetime.now()
now = now.strftime("%Y-%m-%d %H:%M:%S")
np.random.seed(42)
tf.set_random_seed(42)
os.environ['PYTHONHASHSEED'] = '0'
flags = tf.flags
FLAGS = flags.FLAGS
# args for training
flags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate.')
flags.DEFINE_integer('epochs', 100, 'Number of epochs to train.')
flags.DEFINE_integer('hidden1', 16, 'Number of units in hidden layer 1.')
flags.DEFINE_integer('hidden2', 32, 'Number of units in hidden layer 2.')
flags.DEFINE_integer('batch_size', 50, 'Number of batch size for training.')
flags.DEFINE_float('weight_decay', 0.01, 'Weight for L2 loss on embedding matrix.')
flags.DEFINE_float('dropout', 0.2, 'Dropout rate (1 - keep probability).')
# args for path
flags.DEFINE_string('output', '../output/', 'The direction for output files')
# args for single-cell datasets
flags.DEFINE_string('adata', '../data/splatter_data/counts_simulated_dataset1_1500x2500_dropout0.17.h5ad',
'input file for adata.')
flags.DEFINE_string('adj', '../data/splatter_data/adj/true_counts_dataset1_1500x2500_dropout0.17_adj.npz', 'input adjacency.')
# args for data-preprocessing
flags.DEFINE_boolean('normalize_per_cell', True, 'If true, library size normalization is performed using \
the `sc.pp.normalize_per_cell` function in Scanpy and saved into adata \
object.')
flags.DEFINE_boolean('scale', True, 'If true, the input of the autoencoder is centered using \
`sc.pp.scale` function of Scanpy.')
flags.DEFINE_boolean('log1p', True, 'If true, the input of the autoencoder is log transformed with a \
pseudocount of one using `sc.pp.log1p` function of Scanpy.')
flags.DEFINE_boolean('use_raw_as_output', True, 'If true, the ground-truth of express data is adata.raw.X')
# make dirs
if FLAGS.output is not None:
os.makedirs(FLAGS.output, exist_ok=True)
output_dir = os.path.join(FLAGS.output, now)
model_path = os.path.join(output_dir, 'checkpoint')
prediction_path = os.path.join(output_dir, 'prediction')
log_path = os.path.join(output_dir, 'log')
create_dir_if_not_exists(model_path)
create_dir_if_not_exists(prediction_path)
create_dir_if_not_exists(log_path)
adj, adata = load_data()
adj_orig = adj
adj_orig = adj_orig - sp.dia_matrix((adj_orig.diagonal()[np.newaxis, :], [0]), shape=adj_orig.shape)
adj_orig.eliminate_zeros()
adj_train, train_edges, val_edges, val_edges_false, test_edges, test_edges_false = mask_test_edges(adj)
features, features_orig, size_factors, val_features, val_features_idx, test_features, test_features_idx = mask_test_express(adata)
adj = adj_train
adj_norm = preprocess_graph(adj)
# Define placeholders
placeholders = {
'features': tf.placeholder(tf.float32),
'adj': tf.sparse_placeholder(tf.float32),
'adj_orig': tf.sparse_placeholder(tf.float32),
'features_orig': tf.placeholder(tf.float32),
'dropout': tf.placeholder_with_default(0., shape=()),
'size_factors': tf.placeholder(tf.float32),
'is_training': tf.placeholder_with_default(True, shape=())
}
num_features = features.shape[1]
num_nodes = features.shape[0]
model = GraphSCI(placeholders, num_features, num_nodes)
pos_weight_adj = float(adj.shape[0] * adj.shape[0] - adj.sum()) / adj.sum()
norm_adj = adj.shape[0] * adj.shape[0] / float((adj.shape[0] * adj.shape[0] - adj.sum()) * 2)
global_step = tf.Variable(0, trainable=False)
# Optimizer
with tf.name_scope('optimizer'):
opt = OptimizerSCI(preds=(tf.reshape(model.z_adj, [-1]), tf.reshape(model.z_express, [-1])),
labels=(tf.reshape(tf.sparse_tensor_to_dense(placeholders['adj_orig'], validate_indices=False), [-1]),
tf.reshape(placeholders['features_orig'], [-1])),
model=model,
num_nodes=num_nodes,
num_features=num_features,
pos_weight_adj=pos_weight_adj,
norm_adj=norm_adj,
global_step=global_step)
adj_label = adj_train + sp.eye(adj_train.shape[0])
adj_label = sparse_to_tuple(adj_label)
# Initialize session
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
saver = tf.train.Saver(var_list=tf.global_variables())
sess.run(tf.global_variables_initializer())
def get_roc_score(edges_pos, edges_neg):
def sigmoid(x):
x = np.clip(x, -500, 500)
return 1.0 / (1 + np.exp(-x))
# Predict on test set of edges
feed_dict.update({placeholders['is_training']: False})
adj_rec = sess.run(model.z_adj, feed_dict=feed_dict).reshape([num_nodes, num_nodes])
preds = []
pos = []
for e in edges_pos:
preds.append(sigmoid(adj_rec[e[0], e[1]]))
pos.append(adj_orig[e[0], e[1]])
preds_neg = []
neg = []
for e in edges_neg:
preds_neg.append(sigmoid(adj_rec[e[0], e[1]]))
neg.append(adj_orig[e[0], e[1]])
preds_all = np.hstack([preds, preds_neg])
labels_all = np.hstack([np.ones(len(preds)), np.zeros(len(preds))])
roc_score = roc_auc_score(labels_all, preds_all)
ap_score = average_precision_score(labels_all, preds_all)
return roc_score, ap_score
def get_mse_score(features, features_idx, epoch):
# Predict on test set of features
feed_dict.update({placeholders['is_training']: False})
features_rec = sess.run(model.z_express, feed_dict=feed_dict).reshape([num_features, num_nodes])
adata_copy = adata.copy()
adata_copy.X = features_rec
output_file = os.path.join(prediction_path, 'graphsci_tf_simulated_counts_%d.h5ad' % epoch)
adata_copy.write(output_file)
preds = []
for idx in features_idx:
preds.append(features_rec[idx[0], idx[1]])
mse_score = mean_squared_error(np.array(features), np.array(preds))
return mse_score
# Train model
log_file = os.path.join(log_path, 'log.txt')
fp = open(log_file, 'w')
learning_rates = []
costs = []
zinb_losses = []
adj_losses = []
for epoch in range(FLAGS.epochs):
# Construct feed dictionary
feed_dict = construct_feed_dict(adj_norm, adj_label, features, features_orig, size_factors, placeholders, is_training=True)
feed_dict.update({placeholders['dropout']: FLAGS.dropout})
feed_dict.update({global_step: epoch})
# Run single weight update
outs = sess.run([opt.opt_op, opt.cost, opt.cost_adj, opt.cost_express, opt.kl_express, opt.learning_rate], feed_dict=feed_dict)
# Compute average loss
cost = outs[1]
cost_adj = outs[2]
cout_express = outs[3]
kl = outs[4]
learning_rates.append(outs[5])
costs.append(cost)
zinb_losses.append(cout_express)
adj_losses.append(cost_adj)
mse_score = get_mse_score(val_features, val_features_idx, epoch+1)
roc_score, ap_score = get_roc_score(val_edges, val_edges_false)
log_str = "[Epoch%d] train_loss %.6f adj_loss %.6f express_loss %.6f mse_score %.6f roc_score %.6f kl %.6f" \
% (epoch + 1, cost, cost_adj, cout_express, mse_score, roc_score, kl)
make_log(log_str, fp)
fp.close()
print("Optimization Finished!")