參考

實踐上與上面還是有差別，通過實踐理解更深刻了

其實就是看到cs231n課堂上用rnn生成文字，然後就想用keras試試看，順便複習複習python語法，有點生疏了，小技巧比較多

課堂的例子(需要翻牆)不得不說人家一百行程式碼效果就非常好，而且是純python，不用任何框架輔助，佩服

import keras
import word2vec

import numpy as np 
from keras.utils import plot_model
from keras.preprocessing.image import ImageDataGenerator
from keras.models import *
from keras.layers import *
from keras.callbacks import *
from keras import backend as K
import h5py

filename = "input.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 'vocal: ',n_vocab
# data = open('input.txt','r').read()
# data=data.lower()
#上面基本都是仿照課堂上老師給的原始碼
seq_length = 32
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

# reshape X to be [samples, time steps, features]
X = np.reshape(dataX, (n_patterns, seq_length, 1))
# X = X / n_vocab #歸一化後效果不好
#下面可以用函式直接轉成多元分類的 ，例如:valY = np_utils.to_categorical(valY, num_classes=NUM_CLASS)
Y = []
for i in range(n_patterns):
    y = np.zeros((n_vocab, 1))
    y[dataY[i]] = 1
    Y.append(y)
Y = np.reshape(Y, (n_patterns, n_vocab))

print Y.shape

#設定檢查點，儲存權重
filepath="weights-improvement-{epoch:02d}-{loss:.4f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='loss', verbose=1, save_best_only=True, mode='min')
callbacks_list = [checkpoint]


model = Sequential()
model.add(LSTM(64, input_shape=(X.shape[1], X.shape[2]),return_sequences=True))
# # model.add(LSTM(32,return_sequences=True))
# model.add(LSTM(8))

# model.add(LSTM(
#     batch_input_shape=(None, TIME_STEPS, INPUT_SIZE),       # Or: input_dim=INPUT_SIZE, input_length=TIME_STEPS,
#     output_dim=CELL_SIZE,
#     return_sequences=True,      # True: output at all steps. False: output as last step.
#     stateful=True,              # True: the final state of batch1 is feed into the initial state of batch2
# ))

# model.add(Dropout(0.2))


model.add(Dense(n_vocab,activation='softmax'))
adam=keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
adagrad=keras.optimizers.Adagrad(lr=0.001, epsilon=1e-06)
model.compile(loss='categorical_crossentropy', optimizer='adam')

print model.layers[1].input  #use the index of layer to find the input and output shape
print model.layers[1].output

plot_model(model, to_file='model.png')
#嘗試過多層rnn和單層不同寬度，效果都不怎麼好，而且收斂很慢，而且這樣的實現和老師的程式碼演算法上還是有很大區別的，最終效果loss在0.1以下會生產一些單詞，句子基本不可讀
#model.fit(X, Y, nb_epoch=20, batch_size=128, callbacks=callbacks_list)
# model.fit(X, Y, epochs=500, batch_size=128)
# model.save('word_pre.h5')