def main():
parser = argparse.ArgumentParser()
parser.add_argument("--input",
default=None,
type=str,
help="run demo chatbot")
args = parser.parse_args()
input_sentence = args.input
tokenizer = tfds.features.text.SubwordTextEncoder.load_from_file(
vocab_filename)
# Vocabulary size plus start and end token
VOCAB_SIZE = tokenizer.vocab_size + 2
model = Transformer(num_layers=NUM_LAYERS,
units=UNITS,
d_model=D_MODEL,
num_heads=NUM_HEADS,
vocab_size=VOCAB_SIZE,
dropout=DROPOUT,
name='transformer')
demo_sentense = 'How are you'
predict(demo_sentense, tokenizer, model, True)
model.load_weights(save_weight_path)
model.summary()
tf.keras.utils.plot_model(model,
to_file='transformer.png',
show_shapes=True)
predict(input_sentence, tokenizer, model)
from models import Transformer
from utils import tokenizer, parse_wav, generate_input
import tensorflow as tf
import numpy as np
transformer = Transformer(2,
512,
8,
2048,
4337,
pe_input=4096,
pe_target=512,
rate=0.1,
training=False)
checkpoint_path = "results/asr-transfer/2/checkpoint-{:02d}".format(37)
transformer.load_weights(checkpoint_path)
def pre_process(wav_file):
log_bank = parse_wav(wav_file)
input, mask, target = generate_input(log_bank)
return input, mask, target
def post_process(predicted_ids):
y_predict = u"".join([tokenizer.id2token(id) for id in predicted_ids])
return y_predict
def trans_mp3_to_wav(filepath):
song = AudioSegment.from_mp3(filepath)