RADWIMPSっぽい曲LSTMで自動生成してみた（１）（歌詞と、コード進行のみ） ref: https://qiita.com/ahpjop/items/a1d2d159c614258828ab

code_lstm.py

from keras.layers import Dense, Activation, LSTM
from keras.optimizers import RMSprop
from keras.utils.data_utils import get_file
import numpy as np
import random
import sys

chars = sorted(list(set(text)))
print('Total chars:', len(chars))

char_indices = dict((c,i) for i,c in enumerate(chars))
indices_char = dict((i,c) for i,c in enumerate(chars))

maxlen = 1
step = 3
sentences = []
next_chars = []
for i in range(0, len(text) - maxlen, step):
    sentences.append(text[i: i + maxlen])
    next_chars.append(text[i + maxlen])

# テキストのベクトル化
X = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool)
y = np.zeros((len(sentences), len(chars)), dtype=np.bool)

for i, sentence in enumerate(sentences):
    for t, char in enumerate(sentence):
        X[i, t, char_indices[char]] = 1
    y[i, char_indices[next_chars[i]]] = 1


# モデルを定義する
model = Sequential()
model.add(LSTM(128, input_shape=(maxlen, len(chars))))
model.add(Dense(len(chars)))
model.add(Activation('softmax'))
optimizer = RMSprop(lr=0.01)
model.compile(loss='categorical_crossentropy', optimizer=optimizer)

def sample(preds, temperature=1.0):
    preds = np.asarray(preds).astype('float64')
    preds = np.log(preds) / temperature
    exp_preds = np.exp(preds)
    preds = exp_preds / np.sum(exp_preds)
    probas = np.random.multinomial(1, preds, 1)
    return np.argmax(probas)

for iteration in range(1,10):
    print()
    print('-' *50)
    print('繰り返し回数: ', iteration)
    model.fit(X, y, batch_size=128, epochs=1)

    start_index = random.randint(0, len(text)-maxlen-1)

    for diversity in [0.2, 0.5, 1.0, 1.2]:
        print()
        print('-----diveristy', diversity)

        generated = ''
        sentence = text[start_index: start_index + maxlen]
        generated += sentence
        print('----- Seedを生成しました: "' + sentence + '"')
        sys.stdout.write(generated)

        for i in range(40):
            x = np.zeros((1,maxlen,len(chars)))
            for t, char in enumerate(sentence):
                x[0, t, char_indices[char]] = 1.

            preds = model.predict(x, verbose=0)[0]
            next_index = sample(preds, diversity)
            next_char = indices_char[next_index]

            generated += next_char
            sentence = sentence[1:] + next_char

            sys.stdout.write(next_char)
            sys.stdout.flush()
        print()

file3.txt

C　→　D　→　A　→　C　→　C →　D　→　A　→　D　→　D　→　A　→　C　→　A　→　D
C　→　G　→　G　→　Bm →　C　→　A →　C　→　A　→　D　→　D　→　A　→　D 
C　→　Bm7　→　Bm　→　E　→　Bm　→　E　→　A　→　D　→　A　→　Bm7
C　→　G　→　A#　→　G　→　G　→　A　→　C#　→　C　→　F#　→　Em　→　E

file4.txt

F　→　G　→　A　→　D　→　Bm　→　G　→　G#　→　A　→　D　→　C　→　B　→　A
F　→　G　→　A　→　G　→　Em　→　G　→　G　→　A　→　Bm　→　E　→　A　→　G
F　→　C　→　G　→　Em　→　C　→　A　→　G　→　A　→　G　→　C　→　D　→　G

file5.txt

G　→　C　→　D　→　C　→　D　→　C　→　Em　→　C　→　G　→　F　→　C　→　Bm
G　→　A　→　G　→　D　→　E　→　Em　→　G　→　Bm7　→　Em　→　G
G　→　C　→　A　→　G　→　D　→　D　→　G　→　G　→　A　→　C　→　A　→　E 

lyrics_lstm.py

from keras.models import Sequential
from keras.layers import Dense, Activation, LSTM
from keras.optimizers import RMSprop
from keras.utils.data_utils import get_file
import numpy as np
import random
import sys

path = './rad_lyrics.txt'
text = open(path, "r").read()

chars = sorted(list(set(text)))
char_indices = dict((c,i) for i,c in enumerate(chars))
indices_char = dict((i,c) for i,c in enumerate(chars))

maxlen = 40
step = 3
sentences = []
next_chars = []
for i in range(0, len(text) - maxlen, step):
    sentences.append(text[i: i + maxlen])
    next_chars.append(text[i + maxlen])

# テキストのベクトル化
X = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool)
y = np.zeros((len(sentences), len(chars)), dtype=np.bool)

for i, sentence in enumerate(sentences):
    for t, char in enumerate(sentence):
        X[i, t, char_indices[char]] = 1
    y[i, char_indices[next_chars[i]]] = 1

# モデルを定義する
model = Sequential()
model.add(LSTM(128, input_shape=(maxlen, len(chars))))
model.add(Dense(len(chars)))
model.add(Activation('softmax'))
optimizer = RMSprop(lr=0.01)
model.compile(loss='categorical_crossentropy', optimizer=optimizer)

def sample(preds, temperature=1.0):
    preds = np.asarray(preds).astype('float64')
    preds = np.log(preds) / temperature
    exp_preds = np.exp(preds)
    preds = exp_preds / np.sum(exp_preds)
    probas = np.random.multinomial(1, preds, 1)
    return np.argmax(probas)

for iteration in range(1,120):
    print()
    print('-' *50)
    print('繰り返し回数: ', iteration)
    model.fit(X, y, batch_size=128, epochs=1)

    start_index = random.randint(0, len(text)-maxlen-1)

    for diversity in [0.2, 0.5, 1.0, 1.2]:
        print()
        print('-----diveristy', diversity)

        generated = ''
        sentence = text[start_index: start_index + maxlen]
        generated += sentence
        print('----- Seedを生成しました: "' + sentence + '"')
        sys.stdout.write(generated)

        for i in range(100):
            x = np.zeros((1,maxlen,len(chars)))
            for t, char in enumerate(sentence):
                x[0, t, char_indices[char]] = 1.

            preds = model.predict(x, verbose=0)[0]
            next_index = sample(preds, diversity)
            next_char = indices_char[next_index]

            generated += next_char
            sentence = sentence[1:] + next_char

            sys.stdout.write(next_char)
            sys.stdout.flush()
        print()

markov.py

from janome.tokenizer import Tokenizer
import json

# テキストファイルを読み込む
sjis = open('rad_lyrics.txt', 'rb').read()
text = sjis.decode('utf_8')

# テキストを形態素解析読み込みます
t = Tokenizer()
words = t.tokenize(text)

# 辞書を生成
def make_dic(words):
    tmp = ["@"]
    dic = {}
    for i in words:
        word = i.surface
        if word == "" or word == "\r\n" or word == "\n": continue
        tmp.append(word)
        if len(tmp) < 3: continue
        if len(tmp) > 3: tmp = tmp[1:]
        set_word3(dic, tmp)
        if word == "。":
            tmp = ["@"]
            continue
    return dic

# 三要素のリストを辞書として登録
def set_word3(dic, s3):
    w1, w2, w3 = s3
    if not w1 in dic: dic[w1] = {}
    if not w2 in dic[w1]: dic[w1][w2] = {}
    if not w3 in dic[w1][w2]: dic[w1][w2][w3] = 0
    dic[w1][w2][w3] += 1

dic = make_dic(words)
json.dump(dic, open("markov-blog.json", "w", encoding="utf-8"))

##自動生成
import json
dic = open("markov-blog.json" , "r")
dic = json.load(dic)

tweets_list = []
import random
def word_choice(sel):
    keys = sel.keys()
    ran = random.choice(list(keys))
    return ran

def make_sentence(dic):
    ret = []
    if not "@" in dic: return "no dic"
    top = dic["@"]
    w1 = word_choice(top)
    w2 = word_choice(top[w1])
    ret.append(w1)
    ret.append(w2)
    while True:
        w3 = word_choice(dic[w1][w2])
        ret.append(w3)
        if w3 == "。": break
        w1, w2 = w2, w3
    tweets_list.append(ret)
    return "".join(ret)

for i in range(1):
    s = make_sentence(dic)
    tweets_list.append(s)
    print(s)