def _init(logdir, train_mode, *, sampling_seed=None, _cfg):
set_random_seed(_cfg.get('random_seed'))
encoding = _cfg['encoding'].configure()
model = _cfg['model'].configure(RNNSeq2Seq,
train_mode=train_mode,
vocabulary=encoding.vocabulary,
sampling_seed=sampling_seed)
trainer = _cfg['trainer'].configure(BasicTrainer,
dataset_manager=model.dataset_manager,
training_ops=model.training_ops,
logdir=logdir,
write_summaries=train_mode)
if train_mode:
# Configure the dataset manager with the training and validation data.
_cfg['data_prep'].configure(
prepare_train_and_val_data,
dataset_manager=model.dataset_manager,
train_generator=_cfg['train_data'].configure(_load_data,
encoding=encoding),
val_generator=_cfg['val_data'].configure(_load_data,
encoding=encoding),
output_types=(tf.int32, tf.int32, tf.int32),
output_shapes=([None], [None], [None]))
return model, trainer, encoding
def __init__(self, logdir, train_mode, sampling_seed=None):
random_seed = self._cfg.get('random_seed', None)
set_random_seed(random_seed)
self.logdir = logdir
self.input_encoding = self._cfg['input_encoding'].configure()
self.output_encoding = self._cfg['output_encoding'].configure()
num_rows = getattr(self.input_encoding, 'num_rows', None)
self.input_shapes = (([num_rows, None] if num_rows else [None]),
[None], [None], [None])
self.input_types = (tf.float32 if num_rows else tf.int32, tf.int32,
tf.int32, tf.int32)
self.dataset_manager = DatasetManager(
output_types=self.input_types,
output_shapes=tuple([None, *shape] for shape in self.input_shapes))
self.model = self._cfg['model'].configure(
Model,
dataset_manager=self.dataset_manager,
train_mode=train_mode,
vocabulary=self.output_encoding.vocabulary,
sampling_seed=sampling_seed)
self._load_checkpoint = self._cfg.get('load_checkpoint', None)
if self._load_checkpoint and self.model.training_ops is not None:
self.model.training_ops.init_op = ()
self.trainer = self._cfg['trainer'].configure(
BasicTrainer,
session=tf.Session(),
dataset_manager=self.dataset_manager,
training_ops=self.model.training_ops,
logdir=logdir,
write_summaries=train_mode)
if train_mode:
# Configure the dataset manager with the training and validation data.
train_loader = self._cfg['train_data'].configure(
TrainLoader, random_seed=random_seed)
val_loader = self._cfg['val_data'].configure(
TrainLoader, random_seed=random_seed, reseed=True)
self._cfg['data_prep'].configure(
prepare_train_and_val_data,
dataset_manager=self.dataset_manager,
train_generator=self._load_data(train_loader, training=True),
val_generator=self._load_data(val_loader),
output_types=self.input_types,
output_shapes=self.input_shapes)
def __init__(self, logdir, train_mode):
set_random_seed(self._cfg.get('random_seed', None))
self.input_encoding = self._cfg['input_encoding'].configure()
self.output_encoding = self._cfg['output_encoding'].configure()
#with open(self._cfg.get('style_list')) as f:
with open("./seq2seq/data/parallel/styles") as f:
style_list = [line.rstrip('\n') for line in f]
self.style_vocabulary = Vocabulary(style_list,
pad_token=None,
start_token=None,
end_token=None)
self.input_shapes = ([self.input_encoding.num_rows,
None], [], [None], [None])
self.input_types = (tf.float32, tf.int32, tf.int32, tf.int32)
self.dataset_manager = DatasetManager(
output_types=self.input_types,
output_shapes=tuple([None, *shape] for shape in self.input_shapes))
self.model = self._cfg['model'].configure(
CNNRNNSeq2Seq,
dataset_manager=self.dataset_manager,
train_mode=train_mode,
vocabulary=self.output_encoding.vocabulary,
style_vocabulary=self.style_vocabulary)
self.trainer = self._cfg['trainer'].configure(
BasicTrainer,
dataset_manager=self.dataset_manager,
training_ops=self.model.training_ops,
logdir=logdir,
write_summaries=train_mode)
self._load_data_kwargs = dict(input_encoding=self.input_encoding,
output_encoding=self.output_encoding,
style_vocabulary=self.style_vocabulary)
if train_mode:
# Configure the dataset manager with the training and validation data.
self._cfg['data_prep'].configure(
prepare_train_and_val_data,
dataset_manager=self.dataset_manager,
train_generator=self._cfg['train_data'].configure(
load_data, log=True, **self._load_data_kwargs),
val_generator=self._cfg['val_data'].configure(
load_data, **self._load_data_kwargs),
output_types=self.input_types,
output_shapes=self.input_shapes)
def _init(cfg, logdir, train_mode, **kwargs):
set_random_seed(cfg.get('random_seed', None))
encoding = cfg['encoding'].configure()
with open(cfg.get('style_list')) as f:
style_list = [line.rstrip('\n') for line in f]
style_vocabulary = Vocabulary(style_list,
pad_token=None,
start_token=None,
end_token=None)
model = cfg['model'].configure(RNNSeq2Seq,
train_mode=train_mode,
vocabulary=encoding.vocabulary,
style_vocabulary=style_vocabulary)
trainer = cfg['trainer'].configure(BasicTrainer,
dataset_manager=model.dataset_manager,
training_ops=model.training_ops,
logdir=logdir,
write_summaries=train_mode)
if train_mode:
# Configure the dataset manager with the training and validation data.
load_data_kwargs = dict(input_encoding=encoding,
output_encoding=encoding,
style_vocabulary=style_vocabulary)
cfg['data_prep'].configure(
prepare_train_and_val_data,
dataset_manager=model.dataset_manager,
train_generator=cfg['train_data'].configure(load_data,
log=True,
**load_data_kwargs),
val_generator=cfg['val_data'].configure(load_data,
**load_data_kwargs),
output_types=(tf.int32, tf.int32, tf.int32, tf.int32),
output_shapes=([None], [], [None], [None]))
return model, trainer, encoding, style_vocabulary