def train_trigger(self):
train, dev, test = self.t_train, self.t_dev, self.t_test
saver = tf.train.Saver()
maxlen = self.maxlen
print('--Training Trigger--')
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
devbest = 0
testbest = (0, 0, 0)
from tqdm import tqdm
for epoch in tqdm(range(constant.t_epoch)):
loss_list = []
for batch in get_batch(train, constant.t_batch_size, True):
loss, _ = sess.run([self.loss, self.train_op],
feed_dict=get_trigger_feeddict(
self, batch, self.stage, maxlen))
loss_list.append(loss)
print('epoch:{}'.format(str(epoch)))
print('loss:', np.mean(loss_list))
pred_labels = []
for batch in get_batch(dev, constant.t_batch_size, False):
pred_label = sess.run(self.pred_label,
feed_dict=get_trigger_feeddict(
self,
batch,
self.stage,
maxlen,
is_train=False))
pred_labels.extend(list(pred_label))
golds = list(dev[0][4])
dev_p, dev_r, dev_f = f_score(pred_labels, golds)
print("dev_Precision: {} dev_Recall:{} dev_F1:{}".format(
str(dev_p), str(dev_r), str(dev_f)))
pred_labels = []
for batch in get_batch(test, constant.t_batch_size, False):
pred_label = sess.run(self.pred_label,
feed_dict=get_trigger_feeddict(
self,
batch,
self.stage,
maxlen,
is_train=False))
pred_labels.extend(list(pred_label))
golds = list(test[0][4])
test_p, test_r, test_f = f_score(pred_labels, golds)
print("test_Precision: {} test_Recall:{} test_F1:{}\n".format(
str(test_p), str(test_r), str(test_f)))
if dev_f > devbest:
devbest = dev_f
testbest = (test_p, test_r, test_f)
saver.save(sess, "saved_models/trigger.ckpt")
test_p, test_r, test_f = testbest
print(
"test best Precision: {} test best Recall:{} test best F1:{}".
format(str(test_p), str(test_r), str(test_f)))
def train_argument(self):
print('--Training Argument--')
train, dev, test = self.a_train, self.a_dev, self.a_test
with tf.Session() as sess:
devbest = 0
testbest = (0, 0, 0)
sess.run(tf.global_variables_initializer())
for epoch in range(constant.a_epoch):
loss_list = []
for batch in get_batch(train,
constant.a_batch_size,
shuffle=True):
loss, _ = sess.run([self.loss, self.train_op],
feed_dict=get_argument_feeddict(
self, batch, True, "argument"))
loss_list.append(loss)
print('epoch:{}'.format(str(epoch)))
print('loss:', np.mean(loss_list))
pred_labels = []
for batch in get_batch(dev, constant.a_batch_size, False):
pred_event_types, feed_dict = get_argument_feeddict(
self, batch, False, "argument")
pred_label = sess.run(self.pred_label, feed_dict=feed_dict)
pred_labels.extend(
list(zip(list(pred_event_types), list(pred_label))))
golds = list(zip(list(dev[1]), list(dev[2])))
dev_p, dev_r, dev_f = f_score(pred_labels, golds,
self.classify)
print("dev_Precision: {} dev_Recall:{} dev_F1:{}".format(
str(dev_p), str(dev_r), str(dev_f)))
pred_labels = []
for batch in get_batch(test, constant.a_batch_size, False):
pred_event_types, feed_dict = get_argument_feeddict(
self, batch, False, "argument")
pred_label = sess.run(self.pred_label, feed_dict=feed_dict)
pred_labels.extend(
list(zip(list(pred_event_types), list(pred_label))))
golds = list(zip(list(test[1]), list(test[2])))
test_p, test_r, test_f = f_score(pred_labels, golds,
self.classify)
print("test_Precision: {} test_Recall:{} test_F1:{}\n".format(
str(test_p), str(test_r), str(test_f)))
if dev_f > devbest:
devbest = dev_f
testbest = (test_p, test_r, test_f)
test_p, test_r, test_f = testbest
print(
"test best Precision: {} test best Recall:{} test best F1:{}".
format(str(test_p), str(test_r), str(test_f)))
def predict_trigger(self, sess):
print('--Predict Trigger For Argument Stage--')
train, dev, test = self.a_train, self.a_dev, self.a_test
saver = tf.train.Saver()
saver.restore(sess, "saved_models/trigger.ckpt")
dev_pred_event_types = []
test_pred_event_types = []
for batch in get_batch(dev, constant.t_batch_size, False):
pred_label = list(
sess.run(self.pred_label,
feed_dict=get_argument_feeddict(
self, batch, False, 'trigger')))
dev_pred_event_types.extend(pred_label)
for batch in get_batch(test, constant.t_batch_size, False):
pred_label = list(
sess.run(self.pred_label,
feed_dict=get_argument_feeddict(
self, batch, False, 'trigger')))
test_pred_event_types.extend(pred_label)
return self.process_data_for_argument(
np.array(dev_pred_event_types, np.int32),
np.array(test_pred_event_types, np.int32))
def eval_trigger(self):
test = self.t_test
saver = tf.train.Saver()
maxlen = self.maxlen
from collections import defaultdict
import json
results = defaultdict(list)
print('--Eval Trigger--')
with tf.Session() as sess:
saver.restore(sess, "saved_models/trigger.ckpt")
pred_labels = []
for batch in get_batch(test, constant.t_batch_size, False):
pred_label = sess.run(self.pred_label,
feed_dict=get_trigger_feeddict(
self,
batch,
self.stage,
maxlen,
is_train=False))
pred_labels.extend(list(pred_label))
with open('{}/id_align.json'.format(constant.maven_path),
'r') as f:
ids = json.load(f)
with open('test_idxs.json', 'r') as f:
test_idxs = json.load(f)
test_idxs = {
test_idx: idx
for idx, test_idx in enumerate(test_idxs)
}
assert len(test_idxs) == len(pred_labels)
for idx in range(len(ids)):
id_ = ids[idx]
if idx in test_idxs:
label = pred_labels[test_idxs[idx]]
else:
label = 0
results[id_[0]].append({'id': id_[1], 'type_id': int(label)})
with open('results.jsonl', 'w') as f:
for key, val in results.items():
f.write(json.dumps({'id': key, 'predictions': val}) + '\n')
print("--Eval Finish--")
print("nn time: {:.3f} sec".format(time.time() - self._startTime))
# parameters
file_name = 'TNT_dataset.txt'
learning_rate = 0.001
epochs = 100
batch_size = 100
garm_cnt = 1
win_len = 100
win_num = 10
q = 1
w_d = 10000 # kHz
#get signal from dataset
xx0, yy0, sig_len, fi_num = func.get_batch(file_name, garm_cnt)
xx, yy, xx_ac, yy_ac = func.let_2_mass(xx0, yy0)
print(sig_len, fi_num)
out_y = len(yy[1])
s_1000 = int(sig_len / 1000)
print(out_y)
x = tf.placeholder(tf.float32, [None, sig_len])
x_shaped = tf.reshape(x, [-1, sig_len, 1])
y = tf.placeholder(tf.float32, [None, out_y])
print('begin')
import numpy as np
import sounddevice as sd
import tensorflow as tf
from scipy import signal
cl_num=10# количество классо при обучении
ch_num=4
sig=[]
y=[]
sig_len=[]
ch_num_list=[]
preps=['are','dia']
for k in preps:
sig_0,y_0,sig_len_0,ch_num_0=func.get_batch('train_dataset_'+k+'_test_2000.txt')
sig.append(sig_0)
y.append(y_0)
sig_len.append(sig_len_0)
ch_num_list.append(ch_num_0)
learning_rate=0.00015
cl=[]
cl_car=[]
for i in range(len(y)):
for j in y[i]:
c=i
cl.append([1 if j==c else 0 for j in range(cl_num)])
