def test_batched_tensor(self):
b = builder.MetaNetworkBuilder()
with b:
x = b.prebatch_input(shape=(4, 5), dtype=tf.float32)
batched = builder.batched(x)
self.assertEqual(batched.shape.as_list(), [None, 4, 5])
b.preprocessor(())
def test_batched_ragged(self):
b = builder.MetaNetworkBuilder()
with b:
x = b.prebatch_input(shape=(None, 5), dtype=tf.float32)
batched = builder.batched(x)
self.assertTrue(isinstance(batched, tf.RaggedTensor))
self.assertEqual(batched.shape.as_list(), [None, None, 5])
b.preprocessor(())
def _test_batched_ragged_ragged(self, row_split_vals, row_splits_size):
from tensorflow.python.ops.ragged import ragged_tensor_shape as rts # pylint: disable=no-name-in-module
rts.broadcast_to
b = builder.MetaNetworkBuilder()
with b:
values = b.prebatch_input(shape=(None, 5), dtype=tf.float32)
# values = tf.keras.layers.Lambda(lambda x: 2*x)(values)
row_splits = b.prebatch_input(shape=(row_splits_size,),
dtype=tf.int64)
x = tf.RaggedTensor.from_row_splits(values, row_splits)
b.as_batched_model_input(x)
b.preprocessor(())
def _rebuild_model(self, input_spec, output_spec):
from more_keras.meta_models import builder as b
with b.MetaNetworkBuilder() as builder:
self._model = self._model_fn(input_spec, output_spec)
self.compile_model()
# build preprocessor
spec = self.problem.element_spec
if len(spec) == 2:
labels_spec = spec[1]
weights_spec = None
else:
labels_spec, weights_spec = spec[1:]
labels, weights = tf.nest.map_structure(
lambda spec: None if spec is None else builder.batched(
builder.prebatch_input(shape=spec.shape, dtype=spec.dtype)),
(labels_spec, weights_spec))
self._preprocessor = builder.preprocessor(labels, weights)
def test_transposed_consistent(self):
with b.MetaNetworkBuilder():
self._test_transposed_consistent()
def test_sampled_consistent(self):
with b.MetaNetworkBuilder():
self._test_sampled_consistent()