def _build_dense_hparams(self):
ndense = self._hparams.num_dense_layers
dense_size = _to_list(self._hparams.dense_size, 'dense_size', ndense)
other_kwargs = self._hparams.other_dense_kwargs or {}
if isinstance(other_kwargs, HParams):
other_kwargs = other_kwargs.todict()
if not isinstance(other_kwargs, dict):
raise ValueError("hparams['other_dense_kwargs'] must be a dict.")
dense_hparams = []
activation_fn = get_activation_fn(
self._hparams.dense_activation,
self._hparams.dense_activation_kwargs)
for i in range(ndense):
if i == ndense - 1:
activation_fn = get_activation_fn(
self._hparams.final_dense_activation,
self._hparams.final_dense_activation_kwargs)
kwargs_i = {
"units": dense_size[i],
"activation": activation_fn,
"name": "dense_%d" % (i + 1)
}
kwargs_i.update(other_kwargs)
dense_hparams.append({"type": "Dense", "kwargs": kwargs_i})
return dense_hparams
def test_get_activation_fn(self):
"""Tests.
"""
fn = layers.get_activation_fn()
self.assertEqual(fn, tf.identity)
fn = layers.get_activation_fn('relu')
self.assertEqual(fn, tf.nn.relu)
inputs = tf.random_uniform([64, 100], -5, 20, dtype=tf.int32)
fn = layers.get_activation_fn('leaky_relu')
fn_output = fn(inputs)
ref_output = tf.nn.leaky_relu(inputs)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
fn_output_, ref_output_ = sess.run([fn_output, ref_output])
np.testing.assert_array_equal(fn_output_, ref_output_)
fn = layers.get_activation_fn('leaky_relu', kwargs={'alpha': 0.1})
fn_output = fn(inputs)
ref_output = tf.nn.leaky_relu(inputs, alpha=0.1)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
fn_output_, ref_output_ = sess.run([fn_output, ref_output])
np.testing.assert_array_equal(fn_output_, ref_output_)
def _build(self, inputs):
"""Transforms inputs with an MLP layer and packs the results to have
the same structure as specified by :attr:`output_size`.
Args:
inputs: Input (structure of) tensors to be transformed. Must be a
Tensor of shape `[batch_size, ...]` or a (nested) tuple of
such Tensors. That is, the first dimension of (each) tensor
must be the batch dimension.
Returns:
A Tensor or a (nested) tuple of Tensors of the same structure of
`output_size`.
"""
activation_fn = layers.get_activation_fn(self.hparams.activation_fn)
output = _mlp_transform(inputs, self._output_size, activation_fn)
if not self._built:
self._add_internal_trainable_variables()
self._built = True
return output
def _build_conv1d_hparams(self, pool_hparams):
"""Creates the hparams for each of the conv layers usable for
:func:`texar.tf.core.layers.get_layer`.
"""
nconv = self._hparams.num_conv_layers
if len(pool_hparams) != nconv:
raise ValueError("`pool_hparams` must be of length %d" % nconv)
filters = _to_list(self._hparams.filters, 'filters', nconv)
if nconv == 1:
kernel_size = _to_list(self._hparams.kernel_size)
if not isinstance(kernel_size[0], (list, tuple)):
kernel_size = [kernel_size]
elif nconv > 1:
kernel_size = _to_list(self._hparams.kernel_size, 'kernel_size',
nconv)
kernel_size = [_to_list(ks) for ks in kernel_size]
other_kwargs = self._hparams.other_conv_kwargs or {}
if isinstance(other_kwargs, HParams):
other_kwargs = other_kwargs.todict()
if not isinstance(other_kwargs, dict):
raise ValueError("hparams['other_conv_kwargs'] must be a dict.")
conv_pool_hparams = []
activation_fn = get_activation_fn(self._hparams.conv_activation,
self._hparams.conv_activation_kwargs)
for i in range(nconv):
hparams_i = []
names = []
for ks_ij in kernel_size[i]:
name = uniquify_str("conv_%d" % (i + 1), names)
names.append(name)
conv_kwargs_ij = {
"filters": filters[i],
"kernel_size": ks_ij,
"activation": activation_fn,
"name": name
}
conv_kwargs_ij.update(other_kwargs)
hparams_i.append({"type": "Conv1D", "kwargs": conv_kwargs_ij})
if len(hparams_i) == 1:
conv_pool_hparams.append([hparams_i[0], pool_hparams[i]])
else: # creates MergeLayer
mrg_kwargs_layers = []
for hparams_ij in hparams_i:
seq_kwargs_j = {"layers": [hparams_ij, pool_hparams[i]]}
mrg_kwargs_layers.append({
"type": "SequentialLayer",
"kwargs": seq_kwargs_j
})
mrg_hparams = {
"type": "MergeLayer",
"kwargs": {
"layers": mrg_kwargs_layers,
"name": "conv_pool_%d" % (i + 1)
}
}
conv_pool_hparams.append(mrg_hparams)
return conv_pool_hparams