06_treinar_modelo_dnn.py

import argparse
import os
import numpy as np
from gensim import models
from keras.callbacks import ModelCheckpoint, LambdaCallback, TensorBoard
from keras.layers import Dense, Flatten, Dropout, Embedding
from keras.models import Sequential, load_model
from keras import optimizers
import pickle

parser = argparse.ArgumentParser()
parser.add_argument("-epochs", type=int, default=1)
parser.add_argument("-units1", type=int, default=30)
parser.add_argument("-units2", type=int, default=30)
parser.add_argument("-batchsize", type=int, default=64)
parser.add_argument("-gpu", type=bool, default=False)
parser.add_argument("-t", type=int, default=5)
parser.add_argument("-lr", type=float, default=0.001)

args = vars(parser.parse_args())
print(args)

with open('modelos/docs_tokenizados_vagalume.pkl', 'rb') as fp:
    docs_tokenizados = pickle.load(fp)

caminho_modelo_w2v = "modelos/w2v_sertanejo_28kmusicas_tweet_tknzr_window_15_mincount_10.model"
if not os.path.isfile(caminho_modelo_w2v):
    raise FileExistsError("Modelo w2v nao encontrado.")

w2v_model = models.Word2Vec.load(caminho_modelo_w2v)
vocab = list(w2v_model.wv.vocab)


def word2idx(word):
    return w2v_model.wv.vocab[word].index


def idx2word(idx):
    return w2v_model.wv.index2word[idx]


with open('modelos/dados_treino_mincount_10_vagalume.pkl', 'rb') as fp:
    dados_treino = pickle.load(fp)

x = dados_treino[0]
y = dados_treino[1]

tamanho_sentencas = 15
pretrained_weights = w2v_model.wv.vectors
tamanho_vocab = pretrained_weights.shape[0]
tamanho_vetor_w2v = pretrained_weights.shape[1]  # 350
print("Tamanho vocab e w2v vector: ", (tamanho_vocab, tamanho_vetor_w2v))
print("Quantidade de sentencas para treino: {}".format(len(x)))

units1 = args["units1"]
units2 = args["units2"]
learning_rate = args["lr"]
nome_modelo = "dnn-w2v-wordlevel-mincount10-{}-len-{}-{}-lr-{}-sertanejo-vagalume.model".format(tamanho_sentencas, units1, units2, learning_rate)
caminho_modelo_lstm = "modelos/{}".format(nome_modelo)


if os.path.isfile(caminho_modelo_lstm):
    print("Carregando modelo lstm previo...")
    model = load_model(caminho_modelo_lstm)
else:
    print("Criando novo modelo totalmente conectado DNN")
    model = Sequential([
        Embedding(input_dim=tamanho_vocab, input_length=tamanho_sentencas, output_dim=tamanho_vetor_w2v, weights=[pretrained_weights]),
        Flatten(),
        Dense(units1),
        Dropout(0.1),
        Dense(units2),
        Dropout(0.1),
        Dense(tamanho_vocab, activation='softmax')
    ])

    optimizer = optimizers.Adam(lr=learning_rate)
    model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=["accuracy"])


def gerar_texto(epoch, logs):
    print('----- Gerando texto apos a Epoch: %d' % epoch)

    idx_sentenca = np.random.randint(0, len(x))
    idx_tokens = list(x[idx_sentenca])
    seed_tokens = [idx2word(i) for i in idx_tokens]
    print("Seed:")
    print("\"", " ".join(seed_tokens), "\"\n")
    temperatura = args["t"]

    aux = 1
    for i in range(500):

        xt = np.zeros((1, tamanho_sentencas), dtype=np.int32)

        for t, token in enumerate(seed_tokens):
            xt[0, t] = word2idx(token)

        prediction = model.predict(xt, verbose=0)

        if temperatura > 0 and aux % temperatura == 0:
            ordered = (-prediction[0]).argsort()[:3]
            indice = np.random.choice(ordered)
        else:
            indice = (np.argmax(prediction))

        result = idx2word(indice)
        print("\n" if result == "nlinha" else result, end=" ")
        seed_tokens.append(result)
        seed_tokens = seed_tokens[1:len(seed_tokens)]
        aux = aux + 1

    print("\n\nFIM\n\n", )


checkpoint = ModelCheckpoint(caminho_modelo_lstm, monitor='loss', verbose=1, save_best_only=True, mode='min')
print_callback = LambdaCallback(on_epoch_end=gerar_texto)
tensorboard = TensorBoard(log_dir="logs/{}".format(nome_modelo))
callbacks_list = [checkpoint, print_callback, tensorboard]

print(model.summary())

model.fit(x, y, epochs=args["epochs"], batch_size=args["batchsize"], callbacks=callbacks_list)