def main():
data_loader = InputHelper()
data_loader.create_dictionary(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.data_dir + '/')
FLAGS.vocab_size = data_loader.vocab_size
FLAGS.n_classes = data_loader.n_classes
model = BiRNN(FLAGS.rnn_size, FLAGS.layer_size, FLAGS.vocab_size,
FLAGS.batch_size, FLAGS.sequence_length, FLAGS.n_classes,
FLAGS.grad_clip)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state(FLAGS.save_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
while True:
# x = raw_input('请输入一个地址:\n')
x = ''
x = [data_loader.transform_raw(x, FLAGS.sequence_length)]
labels = model.inference(sess, data_loader.labels, x)
print(labels)
def main():
data_loader = InputHelper(log=log)
data_loader.load_embedding(FLAGS.embedding_file, FLAGS.embedding_size)
data_loader.load_label_dictionary(FLAGS.label_dic)
x, y, x_w_p, x_s_p = data_loader.load_valid(FLAGS.valid_file,
FLAGS.interaction_rounds,
FLAGS.sequence_length)
FLAGS.embeddings = data_loader.embeddings
FLAGS.vocab_size = len(data_loader.word2idx)
FLAGS.n_classes = len(data_loader.label_dictionary)
model = BiRNN(embedding_size=FLAGS.embedding_size,
rnn_size=FLAGS.rnn_size,
layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size,
attn_size=FLAGS.attn_size,
sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes,
interaction_rounds=FLAGS.interaction_rounds,
batch_size=FLAGS.batch_size,
embeddings=FLAGS.embeddings,
grad_clip=FLAGS.grad_clip,
learning_rate=FLAGS.learning_rate)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state(FLAGS.save_dir)
model_path = FLAGS.save_dir + '/model.ckpt-45'
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, model_path)
labels = model.inference(sess, y, x, x_w_p, x_s_p)
corrcet_num = 0
for i in range(len(labels)):
if labels[i] == y[i]:
corrcet_num += 1
print('eval_acc = {:.3f}'.format(corrcet_num * 1.0 / len(labels)))
data_loader.output_result(labels, FLAGS.valid_file, FLAGS.result_file)
def main():
data_loader = InputHelper()
data_loader.create_dictionary(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.data_dir + '/')
FLAGS.vocab_size = data_loader.vocab_size
FLAGS.n_classes = data_loader.n_classes
wl = load_wl()
# Define specified Model
model = BiRNN(embedding_size=FLAGS.embedding_size,
rnn_size=FLAGS.rnn_size,
layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size,
attn_size=FLAGS.attn_size,
sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes,
grad_clip=FLAGS.grad_clip,
learning_rate=FLAGS.learning_rate)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state(FLAGS.save_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
now = 0
for file in open('./data/total_txt_img_cat.list'):
word_list = {}
arr = np.zeros([len(wl), 50])
lab = file.split('\t')[2]
for line in open('./data/text_seqs/' + file.split()[0] + '.xml'):
seq = line.split('\t')[0]
x, w = data_loader.transform_raw(seq, FLAGS.sequence_length)
_, out_features = model.inference(sess, data_loader.labels,
[x])
for i, j in enumerate(w):
punc = '[,.!\'%*+-/=><]'
j = re.sub(punc, '', j)
if j in word_list:
word_list[j] += out_features[i]
else:
word_list[j] = out_features[i]
count = 0
for w in word_list:
if w in wl:
arr[wl[w]] = word_list[w]
count += 1
print('now:', now, 'count:', count, 'shape:', arr.shape)
s = str(now)
while len(s) < 4:
s = '0' + s
np.save('./text_lstm/text_' + s + '_' + lab.strip() + '.npy', arr)
now += 1
def train():
data_loader = InputHelper()
data_loader.create_dictionary(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.data_dir + '/')
data_loader.create_batches(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.batch_size, FLAGS.sequence_length)
FLAGS.vocab_size = data_loader.vocab_size
FLAGS.n_classes = data_loader.n_classes
test_data_loader = InputHelper()
test_data_loader.load_dictionary(FLAGS.data_dir + '/dictionary')
test_data_loader.create_batches(FLAGS.data_dir + '/' + FLAGS.test_file,
1000, FLAGS.sequence_length)
model = BiRNN(FLAGS.rnn_size, FLAGS.layer_size, FLAGS.vocab_size,
FLAGS.batch_size, FLAGS.sequence_length, FLAGS.n_classes,
FLAGS.grad_clip)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
for e in xrange(FLAGS.num_epochs):
data_loader.reset_batch()
sess.run(
tf.assign(model.lr,
FLAGS.learning_rate * (FLAGS.decay_rate**e)))
for b in xrange(data_loader.num_batches):
start = time.time()
x, y = data_loader.next_batch()
feed = {
model.input_data: x,
model.targets: y,
model.output_keep_prob: FLAGS.dropout_keep_prob
}
train_loss, _ = sess.run([model.cost, model.train_op],
feed_dict=feed)
end = time.time()
print("{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}" \
.format(e * data_loader.num_batches + b,
FLAGS.num_epochs * data_loader.num_batches,
e, train_loss, end - start))
test_data_loader.reset_batch()
for i in xrange(test_data_loader.num_batches):
test_x, test_y = test_data_loader.next_batch()
feed = {
model.input_data: test_x,
model.targets: test_y,
model.output_keep_prob: 1.0
}
accuracy = sess.run(model.accuracy, feed_dict=feed)
print 'accuracy:{0}'.format(accuracy)
checkpoint_path = os.path.join(FLAGS.save_dir, 'model.ckpt')
saver.save(sess,
checkpoint_path,
global_step=e * data_loader.num_batches)
print 'model saved to {}'.format(checkpoint_path)
def train():
data_loader = InputHelper()
data_loader.create_dictionary(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.data_dir + '/')
data_loader.create_batches(FLAGS.data_dir + '/' + FLAGS.train_file,
FLAGS.batch_size, FLAGS.sequence_length)
FLAGS.vocab_size = data_loader.vocab_size
FLAGS.n_classes = data_loader.n_classes
FLAGS.num_batches = data_loader.num_batches
test_data_loader = InputHelper()
test_data_loader.load_dictionary(FLAGS.data_dir + '/dictionary')
test_data_loader.create_batches(FLAGS.data_dir + '/' + FLAGS.test_file,
100, FLAGS.sequence_length)
if FLAGS.pre_trained_vec:
embeddings = np.load(FLAGS.pre_trained_vec)
print(embeddings.shape)
FLAGS.vocab_size = embeddings.shape[0]
FLAGS.embedding_size = embeddings.shape[1]
if FLAGS.init_from is not None:
assert os.path.isdir(FLAGS.init_from), '{} must be a directory'.format(
FLAGS.init_from)
ckpt = tf.train.get_checkpoint_state(FLAGS.init_from)
assert ckpt, 'No checkpoint found'
assert ckpt.model_checkpoint_path, 'No model path found in checkpoint'
# Define specified Model
model = BiRNN(embedding_size=FLAGS.embedding_size,
rnn_size=FLAGS.rnn_size,
layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size,
attn_size=FLAGS.attn_size,
sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes,
grad_clip=FLAGS.grad_clip,
learning_rate=FLAGS.learning_rate)
# define value for tensorboard
tf.summary.scalar('train_loss', model.cost)
tf.summary.scalar('accuracy', model.accuracy)
merged = tf.summary.merge_all()
# 调整GPU内存分配方案
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
train_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
# using pre trained embeddings
if FLAGS.pre_trained_vec:
sess.run(model.embedding.assign(embeddings))
del embeddings
# restore model
if FLAGS.init_from is not None:
saver.restore(sess, ckpt.model_checkpoint_path)
total_steps = FLAGS.num_epochs * FLAGS.num_batches
for e in range(FLAGS.num_epochs):
data_loader.reset_batch()
for b in range(FLAGS.num_batches):
start = time.time()
x, y = data_loader.next_batch()
feed = {
model.input_data: x,
model.targets: y,
model.output_keep_prob: FLAGS.dropout_keep_prob
}
train_loss, summary, _ = sess.run(
[model.cost, merged, model.train_op], feed_dict=feed)
end = time.time()
global_step = e * FLAGS.num_batches + b
print(
'{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}'
.format(global_step, total_steps, e, train_loss,
end - start))
if global_step % 20 == 0:
train_writer.add_summary(summary,
e * FLAGS.num_batches + b)
if global_step % FLAGS.save_steps == 0:
checkpoint_path = os.path.join(FLAGS.save_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=global_step)
print('model saved to {}'.format(checkpoint_path))
test_data_loader.reset_batch()
test_accuracy = []
for i in range(test_data_loader.num_batches):
test_x, test_y = test_data_loader.next_batch()
feed = {
model.input_data: test_x,
model.targets: test_y,
model.output_keep_prob: 1.0
}
accuracy = sess.run(model.accuracy, feed_dict=feed)
test_accuracy.append(accuracy)
print('test accuracy:{0}'.format(np.average(test_accuracy)))
# -*- coding: utf-8 -*- import os import utils from config import Config from model import BiRNN os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' conf = Config() wav_files, text_labels = utils.get_wavs_lables() words_size, words, word_num_map = utils.create_dict(text_labels) bi_rnn = BiRNN(wav_files, text_labels, words_size, words, word_num_map) bi_rnn.build_train()
train = tf.data.Dataset.from_tensor_slices(train).shuffle(BUFFER_SIZE)
train = train.batch(BATCH_SIZE, drop_remainder=True)
# Load embeddings matrix
embedding_matrix = load_embeddings(embedding_path=EMBEDDING_PATH,
tokenizer=lang,
vocab_size=vocab_size,
embedding_dim=EMBEDDING_DIM,
unk_token=UNK_TOKEN,
start_token=START_TOKEN,
end_token=END_TOKEN)
##------------------------------ Setup training ------------------------------##
birnn = BiRNN(UNITS, PROJECTION_UNITS, MAX_SEQ_LEN, vocab_size, EMBEDDING_DIM,
embedding_matrix)
optimizer = tf.keras.optimizers.Adam()
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
ckpt = tf.train.Checkpoint(birnn=birnn)
ckpt_manager = tf.train.CheckpointManager(ckpt, CHECKPOINT_PATH, max_to_keep=2)
train_summary_writer = tf.summary.create_file_writer(LOG_DIR)
# if a checkpoint exists, restore the latest checkpoint.
if ckpt_manager.latest_checkpoint:
ckpt.restore(ckpt_manager.latest_checkpoint)
print('Latest checkpoint restored')
##------------------------------ Training ------------------------------##
FLAGS.vocab_size = embeddings.shape[0]
FLAGS.embedding_size = embeddings.shape[1]
# 获取模型
if FLAGS.init_from is not None:
assert os.path.isdir(FLAGS.init_from), '{} must be a directory'.format(
FLAGS.init_from)
ckpt = tf.train.get_checkpoint_state(FLAGS.init_from)
assert ckpt, 'No checkpoint found'
assert ckpt.model_checkpoint_path, 'No model path found in checkpoint'
# Define specified Model
model = BiRNN(embedding_size=FLAGS.embedding_size,
rnn_size=FLAGS.rnn_size,
layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size,
attn_size=FLAGS.attn_size,
sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes,
grad_clip=FLAGS.grad_clip,
learning_rate=FLAGS.learning_rate)
with tf.Session() as sess:
train_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables(), max_to_keep=1)
# using pre trained embeddings
if FLAGS.pre_trained_vec:
sess.run(model.embedding.assign(embeddings))
del embeddings
def main(_):
'''
test_data_loader = InputHelper()
test_data_loader.load_dictionary(FLAGS.data_dir+'/dictionary')
test_data_loader.create_batches(FLAGS.data_dir+'/'+FLAGS.test_file, 100, FLAGS.sequence_length)
'''
if FLAGS.pre_trained_vec:
embeddings = np.load(FLAGS.pre_trained_vec)
print(embeddings.shape)
FLAGS.vocab_size = embeddings.shape[0]
FLAGS.embedding_size = embeddings.shape[1]
if FLAGS.init_from is not None:
assert os.path.isdir(FLAGS.init_from), '{} must be a directory'.format(FLAGS.init_from)
ckpt = tf.train.get_checkpoint_state(FLAGS.init_from)
assert ckpt,'No checkpoint found'
assert ckpt.model_checkpoint_path,'No model path found in checkpoint'
# Define specified Model
model = BiRNN(embedding_size=FLAGS.embedding_size, rnn_size=FLAGS.rnn_size, layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size, attn_size=FLAGS.attn_size, sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes, grad_clip=FLAGS.grad_clip, learning_rate=FLAGS.learning_rate)
# define value for tensorboard
tf.summary.scalar('train_loss', model.cost)
tf.summary.scalar('accuracy', model.accuracy)
merged = tf.summary.merge_all()
# 调整GPU内存分配方案
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
train_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
# using pre trained embeddings
if FLAGS.pre_trained_vec:
sess.run(model.embedding.assign(embeddings))
del embeddings
# restore model
if FLAGS.init_from is not None:
saver.restore(sess, ckpt.model_checkpoint_path)
count=patchlength
for e in xrange(FLAGS.num_batches):
total_loss=0
start = time.time()
print('start')
count,inputs,pads,answers = list_tags(count,FLAGS.batch_size)
if count>=len(resp):
count=patchlength
continue
feed = {model.input_data:inputs, model.targets:answers, model.output_keep_prob:FLAGS.dropout_keep_prob,model.pad:pads}
train_loss, summary, _ = sess.run([model.cost, merged, model.train_op], feed_dict=feed)
end = time.time()
print('{}/{} , train_loss = {:.3f}, time/batch = {:.3f}'.format(global_step.eval(), FLAGS.num_batches, train_loss, end - start))
total_loss+=train_loss
if global_step.eval() % 20 == 0:
train_writer.add_summary(summary, e)
if acc>max_acc:
max_acc=acc
checkpoint_path = os.path.join(FLAGS.save_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=global_step)
print 'model saved to {}'.format(checkpoint_path)
print ' loss:',total_loss/FLAGS.num_batches
'''
# -*- coding: utf-8 -*-
import os
import utils
from config import Config
from model import BiRNN
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
print('# Config')
conf = Config()
print('# Load waves and labels')
wav_files, text_labels = utils.get_wavs_lables()
print('# Create dictionary')
words_size, words, word_num_map = utils.create_dict(text_labels)
print('# Build Model')
bi_rnn = BiRNN(wav_files, text_labels, words_size, words, word_num_map)
print('# Begin to test')
wav_file = ['前进.wav']
wav_file = ['后退.wav']
# wav_file = ['左转.wav']
# wav_file = ['右转.wav']
text_label = ['前进 后退 左转 右转']
# print(type(wav_file))
bi_rnn.build_target_wav_file_test(wav_file, text_label)
# -*- coding: utf-8 -*- import os import utils from config import Config from model import BiRNN os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' conf = Config() wav_files, text_labels = utils.get_wavs_lables() words_size, words, word_num_map = utils.create_dict(text_labels) bi_rnn = BiRNN(wav_files, text_labels, words_size, words, word_num_map) bi_rnn.build_test() # wav_files = ['E:\\wav\\train\\A2\\A2_11.wav'] # txt_labels = ['北京 丰台区 农民 自己 花钱 筹办 万 佛 延寿 寺 迎春 庙会 吸引 了 区内 六十 支 秧歌队 参赛'] # bi_rnn = BiRNN(wav_files, text_labels, words_size, words, word_num_map) # bi_rnn.build_target_wav_file_test(wav_files, txt_labels)
def m2c_generator(max_num_sample):
'''
m2c Generator
Input : a testing sample index
Output : Chord Label (n, 16)
Monophony Melody Label (n, 2)
BPM float
Average Elasped Time for one sample : 0.16 sec
'''
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
cpu_device = torch.device('cpu')
# Load Data
chord_dic = pd.read_pickle(CONFIG_ALL['data']['chord_dic'])
# prepare features
all_files = find_files(CONFIG_ALL['data']['test_dir'], '*.mid')
input_dic = []
for i_file in all_files:
_ = midi_feature(i_file, sampling_fac=2)
_ = np.reshape(_, (1, _.shape[0], _.shape[1]))
input_dic.append({'midi': i_file, 'm_embed': _})
print 'Total Number of files : ', len(input_dic)
# training
model = BiRNN(CONFIG_ALL['model']['input_size'],
CONFIG_ALL['model']['lstm_hidden_size'],
CONFIG_ALL['model']['fc_hidden_size'],
CONFIG_ALL['model']['num_layers'],
CONFIG_ALL['model']['num_classes_cf'],
CONFIG_ALL['model']['num_classes_c'], device).to(device)
# Load Model
path = os.path.join(CONFIG_ALL['model']['log_dir'],
CONFIG_ALL['model']['exp_name'], 'models/',
CONFIG_ALL['model']['eval_model'])
model.load_state_dict(torch.load(path))
# Test the model
with torch.no_grad():
while True:
test_idx = yield
if test_idx >= max_num_sample or test_idx < 0:
print "Invalid sample index"
continue
m_embedding = input_dic[test_idx]['m_embed']
out_cf, out_c = model(
torch.tensor(m_embedding, dtype=torch.float).to(device))
out_c = out_c.data.cpu().numpy()
_, pred_cf = torch.max(out_cf.data, 1)
pred_cf = pred_cf.data.cpu().numpy()
i_out_tn1 = -1
i_out_tn2 = -1
i_out_tn3 = -1
i_out_t = -1
predicted = []
c_threshold = 0.825
f_threshold = 0.35
#ochord_threshold = 1.0
for idx, i_out in enumerate(out_c):
# Seventh chord
#T_chord_label = [0, 1, 2, 3, 4, 5, 102, 103, 104]
#D_chord_label = [77, 78, 79, 55, 56, 57]
#R_chord_label = [132]
# Triad Chord
T_chord_label = [0, 1, 37]
D_chord_label = [20, 28]
R_chord_label = [48]
O_chord_label = [
i for i in range(0, 48) if not (i in T_chord_label) or (
i in D_chord_label) or (i in R_chord_label)
]
# Bean Search for repeated note
if pred_cf[idx] == 0:
L = np.argsort(
-np.asarray([i_out[i] for i in T_chord_label]))
if i_out_tn1 == T_chord_label[
L[0]] and i_out_tn2 == T_chord_label[L[0]]:
i_out_t = T_chord_label[L[1]]
else:
i_out_t = T_chord_label[L[0]]
elif pred_cf[idx] == 1:
i_out_t = D_chord_label[np.argmax(
[i_out[i] for i in D_chord_label])]
elif pred_cf[idx] == 3:
L = np.argsort(
-np.asarray([i_out[i] for i in O_chord_label]))
if i_out_tn1 == O_chord_label[
L[0]] and i_out_tn2 == O_chord_label[L[0]]:
i_out_t = O_chord_label[L[1]]
else:
i_out_t = O_chord_label[L[0]]
else:
i_out_t = 48
predicted.append(i_out_t)
i_out_tn2 = i_out_tn1
i_out_tn1 = i_out_t
i_out_last = i_out
# Write file to midi
midi_original = pretty_midi.PrettyMIDI(input_dic[test_idx]['midi'])
midi_chord = pro_chordlabel_to_midi(
predicted,
chord_dic,
inv_beat_resolution=CONFIG_ALL['data']['chord_resolution'],
constant_tempo=midi_original.get_tempo_changes()[1])
midi_chord.instruments[0].name = "Predicted_w_func"
midi_original.instruments.append(midi_chord.instruments[0])
out_path = os.path.join('eval_test/', str(test_idx) + '.mid')
ensure_dir(out_path)
midi_original.write(out_path)
print "Write Files to : ", out_path
out_mc = midi_to_list(midi_original, predicted)
yield {
'melody': out_mc['melody'],
'chord': out_mc['chord'],
'BPM': float(midi_original.get_tempo_changes()[1])
}
from torch import nn as nn
from model import BiRNN
from data import vocab
import torchtext.vocab as Vocab
import os
from data import DATA_ROOT
from data import train_iter, test_iter
import time
# 判定是否能用GPU
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 设置词嵌入维度、隐藏层神经元数量、隐藏层数量
embed_size, num_hiddens, num_layers = 300, 100, 2
net = BiRNN(vocab, embed_size, num_hiddens, num_layers)
# 加载维基百科预训练词向量(使用fasttext),cache为保存目录
fasttext_vocab = Vocab.FastText(cache=os.path.join(DATA_ROOT, "fasttext"))
def load_pretrained_embedding(words, pretrained_vocab):
"""从预训练好的vocab中提取出words对应的词向量"""
# 初始化为0
embed = torch.zeros(len(words), pretrained_vocab.vectors[0].shape[0])
oov_count = 0 # out of vocabulary
for i, word in enumerate(words):
try:
idx = pretrained_vocab.stoi[word]
embed[i, :] = pretrained_vocab.vectors[idx]
except KeyError:
oov_count += 1
print("索引化后的训练数据:", len(train_features))
print("索引化后的测试数据:", len(test_features))
print("训练集标签:", len(labels))
from model import generator, BiRNN, train, evaluate_accuracy
from torch import nn, optim
import torch
from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(np.array(train_features),
np.array(labels),
test_size=0.2)
x_train = torch.tensor(x_train, dtype=torch.long)
y_train = torch.tensor(y_train)
x_test = torch.tensor(x_test, dtype=torch.long)
y_test = torch.tensor(y_test)
test_features = torch.tensor(np.array(test_features), dtype=torch.long)
embeddings_matrix = torch.tensor(embeddings_matrix)
net = BiRNN(word_index, 100, 100, 2)
net.embedding.weight.data.copy_(embeddings_matrix)
net.embedding.weight.requires_grad = False
loss = nn.CrossEntropyLoss()
optimizer = optim.Adam(filter(lambda p: p.requires_grad, net.parameters()),
lr=0.0001)
num_epochs = 10
train_iter = generator(x_train, y_train, 10, train=True)
test_iter = generator(x_test, y_test, 10, train=False)
train(train_iter, test_iter, net, loss, optimizer, num_epochs)
def train():
#train data load
data_loader = InputHelper(log=log)
data_loader.load_embedding(FLAGS.embedding_file,FLAGS.embedding_size)
train_data = data_loader.load_data(FLAGS.data_dir+'/'+FLAGS.train_file, FLAGS.data_dir+'/',FLAGS.interaction_rounds,FLAGS.sequence_length)
x_batch,y_batch,train_interaction_point, train_word_point = data_loader.generate_batches(train_data,FLAGS.batch_size,FLAGS.interaction_rounds)
FLAGS.vocab_size = len(data_loader.word2idx)
FLAGS.n_classes = len(data_loader.label_dictionary)
print FLAGS.n_classes
FLAGS.num_batches = data_loader.num_batches
FLAGS.embeddings = data_loader.embeddings
# test data load
test_data_loader = InputHelper(log=log)
test_data_loader.load_info(embeddings=FLAGS.embeddings,word2idx=data_loader.word2idx,idx2word=data_loader.idx2word,
label_dictionary=data_loader.label_dictionary)
test_data = test_data_loader.load_data(FLAGS.data_dir + '/' + FLAGS.test_file, FLAGS.data_dir + '/',
FLAGS.interaction_rounds, FLAGS.sequence_length)
test_x_batch, test_y_batch, test_interaction_point,test_word_point = test_data_loader.generate_batches(test_data, FLAGS.batch_size, FLAGS.interaction_rounds)
# Define specified Model
model = BiRNN(embedding_size=FLAGS.embedding_size, rnn_size=FLAGS.rnn_size, layer_size=FLAGS.layer_size,
vocab_size=FLAGS.vocab_size, attn_size=FLAGS.attn_size, sequence_length=FLAGS.sequence_length,
n_classes=FLAGS.n_classes, interaction_rounds=FLAGS.interaction_rounds, batch_size=FLAGS.batch_size,
embeddings=FLAGS.embeddings,grad_clip=FLAGS.grad_clip, learning_rate=FLAGS.learning_rate)
# define value for tensorboard
tf.summary.scalar('train_loss', model.cost)
tf.summary.scalar('accuracy', model.accuracy)
merged = tf.summary.merge_all()
# 调整GPU内存分配方案
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
train_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(max_to_keep=1000)
total_steps = FLAGS.num_epochs * FLAGS.num_batches
for e in xrange(FLAGS.num_epochs):
data_loader.reset_batch()
e_avg_loss = []
t_acc = []
start = time.time()
num_tt = []
# w=open('temp/pre'+str(e)+'.txt','w')
for b in xrange(FLAGS.num_batches):
x, y, z,m = data_loader.next_batch(x_batch,y_batch,train_interaction_point,train_word_point)
feed = {model.input_data:x, model.targets:y, model.output_keep_prob:FLAGS.dropout_keep_prob, model.word_point:m, model.sentence_point:z}
train_loss,t_accs,yy,yyy, summary, _ = sess.run([model.cost, model.accuracy,model.y_results,
model.y_tr, merged, model.train_op], feed_dict=feed)
e_avg_loss.append(train_loss)
t_acc.append(t_accs)
global_step = e * FLAGS.num_batches + b
if global_step % 20 == 0:
train_writer.add_summary(summary, e * FLAGS.num_batches + b)
num_t = 0
for i in range(len(yy)):
if yy[i] == yyy[i] and yy[i] != 4:
num_t+=1
num_tt.append(num_t*1.0/len(yy))
# w.write('predict '+str(len(yy))+'\n')
# for y in yy:
# w.write(str(y)+'\t')
# w.write('\ntrue '+str(len(yyy))+'\n')
# for ys in yyy:
# w.write(str(ys)+'\t')
# w.write('\n')
# w.close()
# model test
test_data_loader.reset_batch()
test_accuracy = []
test_a = []
for i in xrange(test_data_loader.num_batches):
test_x, test_y, test_z, test_m = test_data_loader.next_batch(test_x_batch,test_y_batch,test_interaction_point,test_word_point)
feed = {model.input_data:test_x, model.targets:test_y, model.output_keep_prob:1.0,model.word_point:test_m, model.sentence_point:test_z}
accuracy,y_p,y_r = sess.run([model.accuracy,model.y_results,model.y_tr],feed_dict=feed)
test_accuracy.append(accuracy)
num_test = 0
for j in range(len(y_p)):
if y_p[j] == y_r[j] and y_p[j] != 4:
num_test+=1
test_a.append(num_test*1.0/len(y_p))
end = time.time()
num_tt_acc = np.average(num_tt)
num_test_acc = np.average(test_a)
avg_loss = np.average(e_avg_loss)
print('e{},loss = {:.3f}, train_acc = {:.3f}, test_acc = {:.3f}, time/epoch'.format(e,avg_loss,num_tt_acc,num_test_acc,end - start ))
#print and save
# avg_loss = np.average(e_avg_loss)
# t_avg_acc = np.average(t_acc)
# log.info('epoch {}, train_loss = {:.3f},train_acc = {:.3f} test_accuracy:{:.3f}, time/epoch = {:.3f}'.format(e, avg_loss,t_avg_acc,np.average(test_accuracy), end - start))
checkpoint_path = os.path.join(FLAGS.save_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=e)
def main(_):
assert FLAGS.source_train_path, ("--source_train_path is required.")
assert FLAGS.target_train_path, ("--target_train_path is required.")
# Create vocabularies.
source_vocab_path = os.path.join(os.path.dirname(FLAGS.source_train_path),
"vocabulary.source")
target_vocab_path = os.path.join(os.path.dirname(FLAGS.source_train_path),
"vocabulary.target")
utils.create_vocabulary(source_vocab_path, FLAGS.source_train_path, FLAGS.source_vocab_size)
utils.create_vocabulary(target_vocab_path, FLAGS.target_train_path, FLAGS.target_vocab_size)
# Read vocabularies.
source_vocab, rev_source_vocab = utils.initialize_vocabulary(source_vocab_path)
target_vocab, rev_target_vocab = utils.initialize_vocabulary(target_vocab_path)
# Read parallel sentences.
parallel_data = utils.read_data(FLAGS.source_train_path, FLAGS.target_train_path,
source_vocab, target_vocab)
# Read validation data set.
if FLAGS.source_valid_path and FLAGS.target_valid_path:
valid_data = utils.read_data(FLAGS.source_valid_path, FLAGS.target_valid_path,
source_vocab, target_vocab)
# Initialize BiRNN.
config = Config(len(source_vocab),
len(target_vocab),
FLAGS.embedding_size,
FLAGS.state_size,
FLAGS.hidden_size,
FLAGS.num_layers,
FLAGS.learning_rate,
FLAGS.max_gradient_norm,
FLAGS.use_lstm,
FLAGS.use_mean_pooling,
FLAGS.use_max_pooling,
FLAGS.source_embeddings_path,
FLAGS.target_embeddings_path,
FLAGS.fix_pretrained)
model = BiRNN(config)
# Build graph.
model.build_graph()
# Train model.
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
train_iterator = utils.TrainingIteratorRandom(parallel_data, FLAGS.num_negative)
train_summary_writer = tf.summary.FileWriter(os.path.join(FLAGS.checkpoint_dir, "train"), sess.graph)
if FLAGS.source_valid_path and FLAGS.target_valid_path:
valid_iterator = utils.EvalIterator(valid_data)
valid_summary_writer = tf.summary.FileWriter(os.path.join(FLAGS.checkpoint_dir, "valid"), sess.graph)
epoch_loss = 0
epoch_completed = 0
batch_completed = 0
num_iter = int(np.ceil(train_iterator.size / FLAGS.batch_size * FLAGS.num_epochs))
start_time = time.time()
print("Training model on {} sentence pairs per epoch:".
format(train_iterator.size, valid_iterator.size))
for step in xrange(num_iter):
source, target, label = train_iterator.next_batch(FLAGS.batch_size)
source_len = utils.sequence_length(source)
target_len = utils.sequence_length(target)
feed_dict = {model.x_source: source,
model.x_target: target,
model.labels: label,
model.source_seq_length: source_len,
model.target_seq_length: target_len,
model.input_dropout: FLAGS.keep_prob_input,
model.output_dropout: FLAGS.keep_prob_output,
model.decision_threshold: FLAGS.decision_threshold}
_, loss_value, epoch_accuracy,\
epoch_precision, epoch_recall = sess.run([model.train_op,
model.mean_loss,
model.accuracy[1],
model.precision[1],
model.recall[1]],
feed_dict=feed_dict)
epoch_loss += loss_value
batch_completed += 1
# Write the model's training summaries.
if step % FLAGS.steps_per_checkpoint == 0:
summary = sess.run(model.summaries, feed_dict=feed_dict)
train_summary_writer.add_summary(summary, global_step=step)
# End of current epoch.
if train_iterator.epoch_completed > epoch_completed:
epoch_time = time.time() - start_time
epoch_loss /= batch_completed
epoch_f1 = utils.f1_score(epoch_precision, epoch_recall)
epoch_completed += 1
print("Epoch {} in {:.0f} sec\n"
" Training: Loss = {:.6f}, Accuracy = {:.4f}, "
"Precision = {:.4f}, Recall = {:.4f}, F1 = {:.4f}"
.format(epoch_completed, epoch_time,
epoch_loss, epoch_accuracy,
epoch_precision, epoch_recall, epoch_f1))
# Save a model checkpoint.
checkpoint_path = os.path.join(FLAGS.checkpoint_dir, "model.ckpt")
model.saver.save(sess, checkpoint_path, global_step=step)
# Evaluate model on the validation set.
if FLAGS.source_valid_path and FLAGS.target_valid_path:
eval_epoch(sess, model, valid_iterator, valid_summary_writer)
# Initialize local variables for new epoch.
batch_completed = 0
epoch_loss = 0
sess.run(tf.local_variables_initializer())
start_time = time.time()
print("Training done with {} steps.".format(num_iter))
train_summary_writer.close()
valid_summary_writer.close()
from model import BiRNN from options import options import tensorflow as tf import numpy as np from train import train, train_step options = options() opts = options.parse() biLSTM = BiRNN(opts) #load the data (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data() #make it so that the shape is [Batch, 1, 784] that can be fed into the LSTM module x_train = x_train.reshape([-1, 1, 28 * 28]).astype(np.float32) x_test = x_test.reshape([-1, 1, 28 * 28]).astype(np.float32) #transform the input so that the values lie between 0 and 1 x_train = x_train / 255.0 x_test = x_test / 255.0 #make the classes integers y_train = tf.cast(y_train, tf.int64) y_test = tf.cast(y_test, tf.int64) #set the buffer / batch sizes and shuffle the data BUFFER_SIZE = 60000 BATCH_SIZE = opts.batch_size
