python - 改变RNN中的 'temperature'生成文本

Question

我最近关注了一个 tutorial使 RNN 生成文本:我准确地复制了 python 代码并且也有点理解它。我的模型已经训练了 20 个时期，它产生了一个包含 3 个单词的长循环:

"and the wour and the wour and the wour..."

我读过 Andrej Kaparthy 的 blog改变 RNN 的温度会改变其置信度:

Decreasing the temperature from 1 to some lower number (e.g. 0.5) makes the RNN more confident, but also more conservative in its samples.

我想更改此温度级别以降低 RNN 的置信度，以便它创建新模式，但由于这是我的第一个机器学习项目，我不知道如何操作。

这是我的 Python/keras 代码:

生成文本文件:

# Generate Text
import sys
import numpy
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import LSTM
from keras.callbacks import ModelCheckpoint
from keras.utils import np_utils

filename = "king_lear.txt"
raw_text = open(filename).read()
raw_text = raw_text.lower()

chars = sorted(list(set(raw_text)))
char_to_int = dict((c, i) for i, c in enumerate(chars))
int_to_char = dict((i, c) for i, c in enumerate(chars))

n_chars = len(raw_text)
n_vocab = len(chars)
print "Total Characters: ", n_chars
print "Total Vocab: ", n_vocab

seq_length = 100
dataX = []
dataY = []
for i in range(0, n_chars - seq_length, 1):
    seq_in = raw_text[i:i + seq_length]
    seq_out = raw_text[i + seq_length]
    dataX.append([char_to_int[char] for char in seq_in])
    dataY.append(char_to_int[seq_out])
n_patterns = len(dataX)
print "Total Patterns: ", n_patterns

X = numpy.reshape(dataX, (n_patterns, seq_length, 1))

X = X / float(n_vocab)

y = np_utils.to_categorical(dataY)

model = Sequential()
model.add(LSTM(256, input_shape=(X.shape[1], X.shape[2]), return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(256))
model.add(Dropout(0.2))
model.add(Dense(y.shape[1], activation='softmax'))

filename = "weights-improvement-08-2.0298-bigger.hdf5"
model.load_weights(filename)
model.compile(loss='categorical_crossentropy', optimizer='adam')

start = numpy.random.randint(0, len(dataX)-1)
pattern = dataX[start]
print "Seed:"
print "\"", ''.join([int_to_char[value] for value in pattern]), "\""

for i in range(60):
    x = numpy.reshape(pattern, (1, len(pattern), 1))
    x = x / float(n_vocab)
    prediction = model.predict(x, verbose=0)
    index = numpy.argmax(prediction)
    result = int_to_char[index]
    seq_in = [int_to_char[value] for value in pattern]
    sys.stdout.write(result)
    pattern.append(index)
    pattern = pattern[1:len(pattern)]

print "\nDone."

学习文件:

# Learn Sentences
import numpy
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import LSTM
from keras.callbacks import ModelCheckpoint
from keras.utils import np_utils



filename = "king_lear.txt"
raw_text = open(filename).read()
raw_text = raw_text.lower()

chars = sorted(list(set(raw_text)))
char_to_int = dict((c, i) for i, c in enumerate(chars))

n_chars = len(raw_text)
n_vocab = len(chars)
print "Total Characters: ", n_chars
print "Total Vocab: ", n_vocab

seq_length = 100
dataX = []
dataY = []
for i in range(0, n_chars - seq_length, 1):
    seq_in = raw_text[i:i + seq_length]
    seq_out = raw_text[i + seq_length]
    dataX.append([char_to_int[char] for char in seq_in])
    dataY.append(char_to_int[seq_out])
n_patterns = len(dataX)
print "Total Patterns: ", n_patterns

X = numpy.reshape(dataX, (n_patterns, seq_length, 1))

X = X / float(n_vocab)

y = np_utils.to_categorical(dataY)

model = Sequential()
model.add(LSTM(256, input_shape=(X.shape[1], X.shape[2]), return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(256))
model.add(Dropout(0.2))
model.add(Dense(y.shape[1], activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam')

filepath="weights-improvement-{epoch:02d}-{loss:.4f}-bigger.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='loss', verbose=1, save_best_only=True, mode='min')
callbacks_list = [checkpoint]

model.fit(X, y, epochs=50, batch_size=64, callbacks=callbacks_list)

请帮我做这件事。如果这篇文章有任何问题，请随时纠正我，因为这是我的第一个问题。非常感谢。

Answer 1

查看 this issue在 Keras GitHub 上。您可以在 softmax 之前添加一个 Lambda 层以除以温度:

model.add(Lambda(lambda x: x / temp))

根据 Wiki :

For high temperatures, all actions have nearly the same probability and the lower the temperature, the more expected rewards affect the probability. For a low temperature, the probability of the action with the highest expected reward tends to 1.