0
私はテンソルフローによってtflearnのlstm imdb exampleを実装しています。TFlearnを実装するimdb lstm example by tensorflow
tflearnモデルと同じデータセット、アーキテクチャ、ハイパーパラメータ（埋め込みサイズ、文章の長さなど）を使用しましたが、私のモデルのパフォーマンスはtflearnの例より劣ります（10エポック後、約52％の精度であり、例は80％近くであった）。
例の適切なパフォーマンスを達成するためのアドバイスをいただければ幸いです。
以下が私のコードです：
import tensorflow as tf
from tflearn.data_utils import to_categorical, pad_sequences
from tflearn.datasets import imdb
from tensorflow.contrib.rnn import BasicLSTMCell
import time
n_class = 2
n_words = 10000
EMBEDDING_SIZE = 128
HIDDEN_SIZE = 128
MAX_LENGTH = 100
lr = 1e-3
epoch = 10
TRAIN_SIZE = 22500
validation_size = 2500
batch_size = 128
KP = 0.8
# IMDB Dataset loading
train, test, _ = imdb.load_data(path='imdb.pkl', n_words=n_words,
valid_portion=0.1, sort_by_len=False)
trainX, trainY = train
validationX, validationY = test
testX, testY = _
# Data preprocessing
# Sequence padding
trainX = pad_sequences(trainX, maxlen=MAX_LENGTH, value=0.)
validationX = pad_sequences(validationX, maxlen=MAX_LENGTH, value=0.)
testX = pad_sequences(testX, maxlen=MAX_LENGTH, value=0.)
# Converting labels to binary vectors
trainY = to_categorical(trainY, n_class)
validationY = to_categorical(validationY, n_class)
testY = to_categorical(testY, n_class)
graph = tf.Graph()
with graph.as_default():
# input
text = tf.placeholder(tf.int32, [None, MAX_LENGTH])
labels = tf.placeholder(tf.float32, [None, n_class])
keep_prob = tf.placeholder(tf.float32)
embeddings_var = tf.Variable(tf.truncated_normal([n_words, EMBEDDING_SIZE]), trainable=True)
text_embedded = tf.nn.embedding_lookup(embeddings_var, text)
print(text_embedded.shape) # [batch_size, length, embedding_size]
word_list = tf.unstack(text_embedded, axis=1)
cell = BasicLSTMCell(HIDDEN_SIZE)
dropout_cell = tf.contrib.rnn.DropoutWrapper(cell, input_keep_prob=keep_prob, output_keep_prob=keep_prob)
outputs, encoding = tf.nn.static_rnn(dropout_cell, word_list, dtype=tf.float32)
logits = tf.layers.dense(outputs[-1], n_class, activation=None)
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(lr).minimize(loss)
prediction = tf.argmax(logits, 1)
accuracy = tf.reduce_mean(tf.cast(tf.equal(prediction, tf.argmax(labels, 1)), tf.float32))
train_steps = epoch * TRAIN_SIZE // batch_size + 1
print("Train steps: ", train_steps)
with tf.Session(graph=graph) as sess:
tf.global_variables_initializer().run()
print("Initialized!")
s = time.time()
offset = 0
for step in range(train_steps):
offset = (offset * step) % (TRAIN_SIZE - batch_size)
batch_text = trainX[offset: offset + batch_size, :]
batch_label = trainY[offset: offset + batch_size, :]
fd = {text: batch_text, labels: batch_label, keep_prob: KP}
_, l, acc = sess.run([optimizer, loss, accuracy], feed_dict=fd)
if step % 100 == 0:
print("Step: %d loss: %f accuracy: %f" % (step, l, acc))
if step % 500 == 0:
v_l, v_acc = sess.run([loss, accuracy], feed_dict={
text: validationX,
labels: validationY,
keep_prob: 1.0
})
print("------------------------------------------------")
print("Validation: step: %d loss: %f accuracy: %f" % (step, v_l, v_acc))
print("------------------------------------------------")
print("Training finished, time consumed:", time.time() - s, " s")
print("Test accuracy: %f" % accuracy.eval(feed_dict={
text: testX,
labels: testY,
keep_prob: 1.0
}))