def train(sess,graph,config):
batch_size=config["batch_size"]
learning_rate=config["learning_rate"]
if config["validation_dataset"] is None:
_, train_data,valid_data,info = load_and_split_data(config,filename=config["dataset"],valid_data_rate=config["validation_data_rate"])
else:
print("[INFO] training")
train_data, info = load_data(config, filename=config["dataset"])
print("[INFO] validation")
valid_data, valid_info = load_data(config, filename=config["validation_dataset"])
info["graph_node_num"] = max(info["graph_node_num"], valid_info["graph_node_num"])
info["graph_num"] = info["graph_num"] + valid_info["graph_num"]
model = CoreModel(sess,config,info)
model.build(importlib.import_module(config["model.py"]))
if config["profile"]:
vars_to_train = tf.trainable_variables()
print(vars_to_train)
writer = tf.summary.FileWriter('logs', sess.graph)
# Training
start_t = time.time()
model.fit(train_data,valid_data)
train_time = time.time() - start_t
print("traing time:{0}".format(train_time) + "[sec]")
if valid_data.num>0:
# Validation
start_t = time.time()
validation_cost,validation_metrics,prediction_data=model.pred_and_eval(valid_data)
infer_time = time.time() - start_t
print("final cost =",validation_cost)
print("accuracy =",validation_metrics["accuracy"])
print("validation time:{0}".format(infer_time) + "[sec]")
# Saving
if config["save_info_valid"] is not None:
result={}
result["validation_cost"]=validation_cost
result["validation_accuracy"]=validation_metrics
result["train_time"]=train_time
result["infer_time"]=infer_time
save_path=config["save_info_valid"]
print("[SAVE] ",save_path)
fp=open(save_path,"w")
json.dump(result,fp, indent=4, cls=NumPyArangeEncoder)
if config["export_model"]:
try:
print("[SAVE]",config["export_model"])
graph_def = graph_util.convert_variables_to_constants(sess, graph.as_graph_def(), ['output'])
tf.train.write_graph(graph_def, '.', config["export_model"], as_text=False)
except:
print('[ERROR] output has been not found')
if config["save_result_valid"] is not None:
filename=config["save_result_valid"]
save_prediction(filename,prediction_data)
if config["make_plot"]:
plot_cost(config,valid_data,model)
plot_auc(config,valid_data.labels,np.array(prediction_data))
def reconstruct(sess, config):
batch_size = config["batch_size"]
model = importlib.import_module(config["model.py"])
dataset_filename = config["dataset"]
if "dataset_test" in config:
dataset_filename = config["dataset_test"]
all_data, info = load_data(config, filename=dataset_filename)
graph_index_list = []
for i in range(all_data.num):
graph_index_list.append([i, i])
info.graph_index_list = graph_index_list
info.pos_weight = get_pos_weight(all_data)
info.norm = get_norm(all_data)
print("pos_weight=", info.pos_weight)
print("norm=", info.pos_weight)
model = CoreModel(sess,
config,
info,
construct_feed_callback=construct_feed)
model.build(importlib.import_module(config["model.py"]), is_train=False)
vars_to_train = tf.trainable_variables()
for v in vars_to_train:
print(v)
# initialize session
saver = tf.train.Saver()
#sess.run(tf.global_variables_initializer())
print("[load]", config["load_model"])
saver.restore(sess, config["load_model"])
start_t = time.time()
cost, acc, pred_data = model.pred_and_eval(all_data)
recons_data = pred_data
"""
recons_data=[]
for i in range(3):
print(i)
cost,acc,pred_data=model.pred_and_eval(all_data)
recons_data.append(pred_data)
"""
if "reconstruction_test" in config:
filename = config["reconstruction_test"]
os.makedirs(os.path.dirname(filename), exist_ok=True)
print("[SAVE]", filename)
joblib.dump(recons_data, filename)
def infer(sess, graph, config):
batch_size = config["batch_size"]
model = importlib.import_module(config["model.py"])
dataset_filename = config["dataset"]
if "dataset_test" in config:
dataset_filename = config["dataset_test"]
all_data, info = load_data(config, filename=dataset_filename)
model = CoreModel(sess, config, info)
model.build(importlib.import_module(config["model.py"]), is_train=False)
# Initialize session
saver = tf.train.Saver()
#sess.run(tf.global_variables_initializer())
print("[LOAD]", config["load_model"])
saver.restore(sess, config["load_model"])
# Validation
start_t = time.time()
test_cost, test_metrics, prediction_data = model.pred_and_eval(all_data)
infer_time = time.time() - start_t
print("final cost =", test_cost)
print("accuracy =", test_metrics["accuracy"])
print("infer time:{0}".format(infer_time) + "[sec]")
if config["save_info_test"] is not None:
result = {}
result["test_cost"] = test_cost
result["test_accuracy"] = test_metrics
result["infer_time"] = infer_time
save_path = config["save_info_test"]
print("[SAVE] ", save_path)
fp = open(save_path, "w")
json.dump(result, fp, indent=4, cls=NumPyArangeEncoder)
if config["prediction_data"] is None:
print("[ERROR] prediction_data is required")
quit()
obj = {}
obj["prediction_data"] = prediction_data
obj["labels"] = all_data.labels
os.makedirs(os.path.dirname(config["prediction_data"]), exist_ok=True)
joblib.dump(obj, config["prediction_data"])
def visualize(sess, config, ig_targets):
from tensorflow.python import debug as tf_debug
from gcn_modules.visualization import cal_feature_IG
# 入力は１分子づつ
batch_size = 1
# 入力データから、全データの情報, 学習用データの情報, 検証用データの情報, および
# グラフに関する情報を順に取得する
all_data, info = load_data(config,
filename=config["dataset"],
prohibit_shuffle=True)
#all_data.labels = tf.one_hot(tf.cast(tf.squeeze(all_data.labels), tf.int32), depth=2)
model = importlib.import_module(config["model.py"])
placeholders = model.build_placeholders(info,
config,
batch_size=batch_size)
try:
_model, prediction, _, _, _ = model.build_model(
placeholders,
info,
config,
batch_size=batch_size,
feed_embedded_layer=True)
except:
_model, prediction, _, _, _ = model.build_model(placeholders,
info,
config,
batch_size=batch_size)
#--- セッションの初期化
saver = tf.train.Saver()
print("[LOAD]", config["load_model"])
saver.restore(sess, config["load_model"])
#--- integrated gradientsの計算
cal_feature_IG(sess,
all_data,
placeholders,
info,
prediction,
ig_targets,
logger=tf.logging,
model=_model)
def train(sess, config):
batch_size = config["batch_size"]
learning_rate = config["learning_rate"]
model = importlib.import_module(config["model.py"])
all_data, info = load_data(config, filename=config["dataset"])
placeholders = model.build_placeholders(
info, batch_size=batch_size, adj_channel_num=info.adj_channel_num)
out, prediction, cost, cost_sum, metrics = model.build_model(
placeholders,
info,
batch_size=batch_size,
adj_channel_num=info.adj_channel_num,
embedding_dim=config["embedding_dim"])
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cost)
#train_step = tf.train.MomentumOptimizer(learning_rate,0.01).minimize(cost)
# Initialize session
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
# Train model
all_list_num = len(info.graph_index_list)
print("#graph_index list = ", all_list_num)
print("#data = ", all_data.num)
data_idx = list(range(len(info.graph_index_list)))
n = int(all_list_num * 0.8)
train_idx = data_idx[:n]
train_num = len(train_idx)
valid_idx = data_idx[n:]
valid_num = len(valid_idx)
eary_stopping = EarlyStopping(config)
start_t = time.time()
for epoch in range(config["epoch"]): #[range(FLAGS.epochs):
np.random.shuffle(train_idx)
# training
itr_num = int(np.ceil(train_num / batch_size))
training_cost = 0
training_correct_count = 0
for itr in range(itr_num):
offset_b = itr * batch_size
batch_idx = train_idx[offset_b:offset_b + batch_size]
feed_dict = construct_feed(batch_idx,
placeholders,
all_data,
info.graph_index_list,
batch_size=batch_size,
dropout_rate=0.5)
# running parameter update with tensorflow
out_prediction = sess.run([prediction], feed_dict=feed_dict)
#print(out_prediction)
_, out_cost_sum, out_metrics = sess.run(
[train_step, cost_sum, metrics], feed_dict=feed_dict)
training_cost += out_cost_sum
training_correct_count += out_metrics["correct_count"]
#print(out_metrics["correct_count"])
#print(batch_size)
training_cost /= train_num
training_accuracy = training_correct_count / train_num
# validation
itr_num = int(np.ceil(valid_num / batch_size))
validation_cost = 0
validation_correct_count = 0
for itr in range(itr_num):
offset_b = itr * batch_size
batch_idx = valid_idx[offset_b:offset_b + batch_size]
feed_dict = construct_feed(batch_idx,
placeholders,
all_data,
info.graph_index_list,
batch_size=batch_size)
out_cost_sum, out_metrics = sess.run([cost_sum, metrics],
feed_dict=feed_dict)
validation_cost += out_cost_sum
validation_correct_count += out_metrics["correct_count"]
validation_cost /= valid_num
validation_accuracy = validation_correct_count / valid_num
# check point
save_path = None
if (epoch) % config["save_interval"] == 0:
# save
save_path = config["save_model_path"] + "/model.%05d.ckpt" % (
epoch)
saver.save(sess, save_path)
# early stopping and printing information
if eary_stopping.evaluate_validation(
validation_cost, {
"epoch": epoch,
"validation_accuracy": validation_accuracy,
"validation_cost": validation_cost,
"training_accuracy": training_accuracy,
"training_cost": training_cost,
"save_path": save_path
}):
break
train_time = time.time() - start_t
print("traing time:{0}".format(train_time) + "[sec]")
# saving last model
#save_path = config["save_model_path"]+"/model.last.ckpt"
if "save_model" in config and config["save_model"] is not None:
save_path = config["save_model"]
print("[SAVE] ", save_path)
saver.save(sess, save_path)
# validation
start_t = time.time()
data_idx = list(range(all_data.num))
itr_num = int(np.ceil(all_data.num / batch_size))
validation_cost = 0
validation_correct_count = 0
prediction_data = []
for itr in range(itr_num):
offset_b = itr * batch_size
batch_idx = data_idx[offset_b:offset_b + batch_size]
feed_dict = construct_feed(batch_idx,
placeholders,
all_data,
batch_size=batch_size)
out_cost_sum, out_metrics, out_prediction = sess.run(
[cost_sum, metrics, prediction], feed_dict=feed_dict)
validation_cost += out_cost_sum
validation_correct_count += out_metrics["correct_count"]
prediction_data.append(out_prediction)
validation_cost /= all_data.num
validation_accuracy = validation_correct_count / all_data.num
print("final cost =", validation_cost)
print("accuracy =", validation_accuracy)
train_time = time.time() - start_t
print("infer time:{0}".format(train_time) + "[sec]")
if "save_result_train" in config:
filename = config["save_result_train"]
save_prediction(filename, prediction_data)
def train_cv(sess, graph, config):
from sklearn.model_selection import KFold
from gcn_modules.make_plots import make_auc_plot, make_cost_acc_plot
import sklearn
from sklearn.metrics import roc_curve, auc, accuracy_score, precision_recall_fscore_support
from scipy import interp
batch_size = config["batch_size"]
learning_rate = config["learning_rate"]
all_data, info = load_data(config, filename=config["dataset"])
model = CoreModel(sess, config, info)
model.build(importlib.import_module(config["model.py"]))
# Training
kf = KFold(n_splits=config["k-fold_num"], shuffle=True, random_state=123)
kf_count = 1
fold_data_list = []
if all_data["labels"] is not None:
split_base = all_data["labels"]
else:
split_base = all_data["label_list"][0]
score_metrics = []
if config["task"] == "regression":
metric_name = "mse"
else:
metric_name = "accuracy"
for train_valid_list, test_list in kf.split(split_base):
print("starting fold:{0}".format(kf_count))
train_valid_data, test_data = split_data(
all_data,
indices_for_train_data=train_valid_list,
indices_for_valid_data=test_list)
train_data, valid_data = split_data(
train_valid_data, valid_data_rate=config["validation_data_rate"])
# Training
start_t = time.time()
model.fit(train_data, valid_data, k_fold_num=kf_count)
train_time = time.time() - start_t
print("traing time:{0}".format(train_time) + "[sec]")
# Test
start_t = time.time()
test_cost, test_metrics, prediction_data = model.pred_and_eval(
test_data)
infer_time = time.time() - start_t
print("final cost =", test_cost)
print("%s =%f" % (metric_name, test_metrics[metric_name]))
score_metrics.append(test_metrics[metric_name])
print("infer time:{0}".format(infer_time) + "[sec]")
if config["export_model"]:
try:
name, ext = os.path.splitext(config["export_model"])
filename = name + "." + str(kf_count) + ext
print("[SAVE]", filename)
graph_def = graph_util.convert_variables_to_constants(
sess, graph.as_graph_def(), ['output'])
tf.train.write_graph(graph_def, '.', filename, as_text=False)
except:
print('[ERROR] output has been not found')
# save fold data
fold_data = dotdict({})
fold_data.prediction_data = prediction_data
if all_data["labels"] is not None:
fold_data.test_labels = test_data.labels
else:
fold_data.test_labels = test_data.label_list
fold_data.test_data_idx = test_list
if config["task"] == "regression":
fold_data.training_mse = [
el["training_mse"] for el in model.training_metrics_list
]
fold_data.validation_mse = [
el["validation_mse"] for el in model.validation_metrics_list
]
else:
fold_data.training_acc = [
el["training_accuracy"] for el in model.training_metrics_list
]
fold_data.validation_acc = [
el["validation_accuracy"]
for el in model.validation_metrics_list
]
fold_data.test_acc = test_metrics[metric_name]
fold_data.training_cost = model.training_cost_list
fold_data.validation_cost = model.validation_cost_list
fold_data.test_cost = test_cost
fold_data.train_time = train_time
fold_data.infer_time = infer_time
fold_data_list.append(fold_data)
kf_count += 1
print("cv %s(mean) =%f" % (metric_name, np.mean(score_metrics)))
print("cv %s(std.) =%f" % (metric_name, np.std(score_metrics)))
if "save_info_cv" in config and config["save_info_cv"] is not None:
save_path = config["save_info_cv"]
print("[SAVE] ", save_path)
_, ext = os.path.splitext(save_path)
if ext == ".json":
json.dump(fold_data_list, fp, indent=4, cls=NumPyArangeEncoder)
else:
joblib.dump(fold_data_list, save_path, compress=True)
#
if "save_result_cv" in config and config["save_result_cv"] is not None:
result_cv = []
for j, fold_data in enumerate(fold_data_list):
pred_score = np.array(fold_data.prediction_data)
if len(pred_score.shape) == 3: # multi-label-multi-task
# #data x # task x #class
# => this program supports only 2 labels
pred_score = pred_score[:, :, 1]
true_label = np.array(fold_data.test_labels)
# #data x # task x #class
if len(pred_score.shape) == 1:
pred_score = pred_score[:, np.newaxis]
if len(true_label.shape) == 1:
true_label = true_label[:, np.newaxis]
v = []
for i in range(info.label_dim):
el = {}
if config["task"] == "regression":
el["r2"] = sklearn.metrics.r2_score(
true_label[:, i], pred_score[:, i])
el["mse"] = sklearn.metrics.mean_squared_error(
true_label[:, i], pred_score[:, i])
else:
pred = np.zeros(pred_score.shape)
pred[pred_score > 0.5] = 1
fpr, tpr, _ = roc_curve(true_label[:, i],
pred_score[:, i],
pos_label=1)
roc_auc = auc(fpr, tpr)
acc = accuracy_score(true_label[:, i], pred[:, i])
scores = precision_recall_fscore_support(true_label[:, i],
pred[:, i],
average='binary')
el["auc"] = roc_auc
el["acc"] = acc
el["pre"] = scores[0]
el["rec"] = scores[1]
el["f"] = scores[2]
el["sup"] = scores[3]
v.append(el)
result_cv.append(v)
save_path = config["save_result_cv"]
print("[SAVE] ", save_path)
fp = open(save_path, "w")
json.dump(result_cv, fp, indent=4, cls=NumPyArangeEncoder)
#
for i, fold_data in enumerate(fold_data_list):
prefix = "fold" + str(i) + "_"
result_path = config["plot_path"]
os.makedirs(result_path, exist_ok=True)
if config["make_plot"]:
if config["task"] == "regression":
# plot cost
make_cost_acc_plot(fold_data.training_cost,
fold_data.validation_cost,
fold_data.training_mse,
fold_data.validation_mse,
result_path + prefix)
pred_score = np.array(fold_data.prediction_data)
plot_r2(config,
fold_data.test_labels,
pred_score,
prefix=prefix)
else:
# plot cost
make_cost_acc_plot(fold_data.training_cost,
fold_data.validation_cost,
fold_data.training_acc,
fold_data.validation_acc,
result_path + prefix)
# plot AUC
pred_score = np.array(fold_data.prediction_data)
plot_auc(config,
fold_data.test_labels,
pred_score,
prefix=prefix)