# Update clock (next request)
# total_time = sum([e.value for e in map(Event.from_int, seq) if e.type == 'time_shift'])
# self.server_state['tick_interval'] = total_time / 1000
return seq
def decode(self, token):
return Event.from_int(token)
if __name__ == '__main__':
# import utils
print(tf.executing_eagerly())
src = tf.constant(
[utils.fill_with_placeholder([1, 2, 3, 4], max_len=2048)])
trg = tf.constant(
[utils.fill_with_placeholder([1, 2, 3, 4], max_len=2048)])
src_mask, trg_mask, lookup_mask = utils.get_masked_with_pad_tensor(
2048, src, trg)
print(lookup_mask)
print(src_mask)
mt = MusicTransformer(debug=True,
embedding_dim=par.embedding_dim,
vocab_size=par.vocab_size)
mt.save_weights('my_model.h5', save_format='h5')
mt.load_weights('my_model.h5')
result = mt.generate([27, 186, 43, 213, 115, 131], length=100)
print(result)
from deprecated import sequence
out_tar = y
inp_tar = y[:, :-1]
inp_tar = tf.concat([start_token, inp_tar], -1)
return x, inp_tar, out_tar
def reset_metrics(self):
for metric in self.custom_metrics:
metric.reset_states()
return
if __name__ == '__main__':
import utils
print(tf.executing_eagerly())
src = tf.constant([utils.fill_with_placeholder([1,2,3,4],max_len=2048)])
trg = tf.constant([utils.fill_with_placeholder([1,2,3,4],max_len=2048)])
src_mask, trg_mask, lookup_mask = utils.get_masked_with_pad_tensor(2048, src,trg)
print(lookup_mask)
# print(src_mask, trg_mask)
mt = MusicTransformer(debug=True, embedding_dim=par.embedding_dim, vocab_size=par.vocab_size)
mt.save_weights('my_model.h5', save_format='h5')
mt.load_weights('my_model.h5')
# print('compile...')
# mt.compile(optimizer='adam', loss=callback.TransformerLoss(debug=True))
# # print(mt.train_step(inp=src, tar=trg))
#
# print('start training...')
# for i in range(2):
# mt.train_on_batch(x=src, y=trg)
result = mt.generate([27, 186, 43, 213, 115, 131], length=100)
