def __init__(self, in_channels: int, in_features: Optional[int] = None,
hparams: Optional[Union[HParams, Dict[str, Any]]] = None):
super().__init__(hparams=hparams)
encoder_hparams = utils.dict_fetch(hparams,
Conv1DEncoder.default_hparams())
self._encoder = Conv1DEncoder(in_channels=in_channels,
in_features=in_features,
hparams=encoder_hparams)
# Add an additional dense layer if needed
self._num_classes = self._hparams.num_classes
if self._num_classes > 0:
if self._hparams.num_dense_layers <= 0:
self._encoder.append_layer({"type": "Flatten"})
logit_kwargs = self._hparams.logit_layer_kwargs
if logit_kwargs is None:
logit_kwargs = {}
elif not isinstance(logit_kwargs, HParams):
raise ValueError(
"hparams['logit_layer_kwargs'] must be a dict.")
else:
logit_kwargs = logit_kwargs.todict()
logit_kwargs.update({"out_features": self._num_classes})
self._encoder.append_layer({"type": "Linear",
"kwargs": logit_kwargs})
def __init__(self, hparams=None):
ClassifierBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
encoder_hparams = utils.dict_fetch(
hparams, Conv1DEncoder.default_hparams())
self._encoder = Conv1DEncoder(hparams=encoder_hparams)
# Add an additional dense layer if needed
self._num_classes = self._hparams.num_classes
if self._num_classes > 0:
if self._hparams.num_dense_layers <= 0:
self._encoder.append_layer({"type": "Flatten"})
logit_kwargs = self._hparams.logit_layer_kwargs
if logit_kwargs is None:
logit_kwargs = {}
elif not isinstance(logit_kwargs, HParams):
raise ValueError(
"hparams['logit_layer_kwargs'] must be a dict.")
else:
logit_kwargs = logit_kwargs.todict()
logit_kwargs.update({"units": self._num_classes})
if 'name' not in logit_kwargs:
logit_kwargs['name'] = "logit_layer"
self._encoder.append_layer(
{"type": "Dense", "kwargs": logit_kwargs})
def test_encode(self):
r"""Tests encode.
"""
inputs_1 = torch.ones([128, 32, 300])
encoder_1 = Conv1DEncoder(in_channels=inputs_1.size(1),
in_features=inputs_1.size(2))
self.assertEqual(len(encoder_1.layers), 4)
self.assertTrue(
isinstance(encoder_1.layer_by_name("MergeLayer"),
tx.core.MergeLayer))
for layer in encoder_1.layers[0].layers:
self.assertTrue(isinstance(layer, nn.Sequential))
outputs_1 = encoder_1(inputs_1)
self.assertEqual(outputs_1.size(), torch.Size([128, 256]))
self.assertEqual(outputs_1.size(-1), encoder_1.output_size)
inputs_2 = torch.ones([128, 64, 300])
hparams = {
# Conv layers
"num_conv_layers": 2,
"out_channels": 128,
"kernel_size": [[3, 4, 5], 4],
"other_conv_kwargs": {
"padding": 0
},
# Pooling layers
"pooling": "AvgPool1d",
"pool_size": 2,
"pool_stride": 1,
# Dense layers
"num_dense_layers": 3,
"out_features": [128, 128, 10],
"dense_activation": "ReLU",
"other_dense_kwargs": None,
# Dropout
"dropout_conv": [0, 1, 2],
"dropout_dense": 2
}
network_2 = Conv1DEncoder(in_channels=inputs_2.size(1),
in_features=inputs_2.size(2),
hparams=hparams)
# dropout-merge-dropout-conv-avgpool-dropout-flatten-
# (Sequential(Linear,ReLU))-(Sequential(Linear,ReLU))-dropout-linear
self.assertEqual(len(network_2.layers), 1 + 1 + 1 + 3 + 4 + 1)
self.assertTrue(
isinstance(network_2.layer_by_name("MergeLayer"),
tx.core.MergeLayer))
for layer in network_2.layers[1].layers:
self.assertTrue(isinstance(layer, nn.Sequential))
outputs_2 = network_2(inputs_2)
self.assertEqual(outputs_2.size(), torch.Size([128, 10]))
self.assertEqual(outputs_2.size(-1), network_2.output_size)
def test_unknown_seq_length(self):
"""Tests use of pooling layer when the seq_length dimension of inputs
is `None`.
"""
encoder_1 = Conv1DEncoder()
inputs_1 = tf.placeholder(tf.float32, [64, None, 300])
outputs_1 = encoder_1(inputs_1)
self.assertEqual(outputs_1.shape, [64, 128])
hparams = {
# Conv layers
"num_conv_layers": 2,
"filters": 128,
"kernel_size": [[3, 4, 5], 4],
# Pooling layers
"pooling": "AveragePooling",
"pool_size": [2, None],
# Dense layers
"num_dense_layers": 1,
"dense_size": 10,
}
encoder = Conv1DEncoder(hparams)
# nlayers = nconv-pool + nconv + npool + ndense + ndropout + flatten
self.assertEqual(len(encoder.layers), 1 + 1 + 1 + 1 + 1 + 1)
self.assertTrue(
isinstance(encoder.layer_by_name('pool_2'),
tx.core.AverageReducePooling1D))
inputs = tf.placeholder(tf.float32, [64, None, 300])
outputs = encoder(inputs)
self.assertEqual(outputs.shape, [64, 10])
hparams_2 = {
# Conv layers
"num_conv_layers": 1,
"filters": 128,
"kernel_size": 4,
"other_conv_kwargs": {
'data_format': 'channels_first'
},
# Pooling layers
"pooling": "MaxPooling",
"other_pool_kwargs": {
'data_format': 'channels_first'
},
# Dense layers
"num_dense_layers": 1,
"dense_size": 10,
}
encoder_2 = Conv1DEncoder(hparams_2)
inputs_2 = tf.placeholder(tf.float32, [64, 300, None])
outputs_2 = encoder_2(inputs_2)
self.assertEqual(outputs_2.shape, [64, 10])
def test_encode(self):
"""Tests encode.
"""
encoder_1 = Conv1DEncoder()
self.assertEqual(len(encoder_1.layers), 4)
self.assertTrue(
isinstance(encoder_1.layer_by_name("conv_pool_1"),
tx.core.MergeLayer))
for layer in encoder_1.layers[0].layers:
self.assertTrue(isinstance(layer, tx.core.SequentialLayer))
inputs_1 = tf.ones([64, 16, 300], tf.float32)
outputs_1 = encoder_1(inputs_1)
self.assertEqual(outputs_1.shape, [64, 128])
hparams = {
# Conv layers
"num_conv_layers": 2,
"filters": 128,
"kernel_size": [[3, 4, 5], 4],
"other_conv_kwargs": {
"padding": "same"
},
# Pooling layers
"pooling": "AveragePooling",
"pool_size": 2,
"pool_strides": 1,
# Dense layers
"num_dense_layers": 3,
"dense_size": [128, 128, 10],
"dense_activation": "relu",
"other_dense_kwargs": {
"use_bias": False
},
# Dropout
"dropout_conv": [0, 1, 2],
"dropout_dense": 2
}
encoder_2 = Conv1DEncoder(hparams)
# nlayers = nconv-pool + nconv + npool + ndense + ndropout + flatten
self.assertEqual(len(encoder_2.layers), 1 + 1 + 1 + 3 + 4 + 1)
self.assertTrue(
isinstance(encoder_2.layer_by_name("conv_pool_1"),
tx.core.MergeLayer))
for layer in encoder_2.layers[1].layers:
self.assertTrue(isinstance(layer, tx.core.SequentialLayer))
inputs_2 = tf.ones([64, 16, 300], tf.float32)
outputs_2 = encoder_2(inputs_2)
self.assertEqual(outputs_2.shape, [64, 10])
def default_hparams() -> Dict[str, Any]:
r"""Returns a dictionary of hyperparameters with default values.
.. code-block:: python
{
# (1) Same hyperparameters as in Conv1DEncoder
...
# (2) Additional hyperparameters
"num_classes": 2,
"logit_layer_kwargs": {
"use_bias": False
},
"name": "conv1d_classifier"
}
Here:
1. Same hyperparameters as in :class:`~texar.modules.Conv1DEncoder`.
See the :meth:`~texar.modules.Conv1DEncoder.default_hparams`.
An instance of :class:`~texar.modules.Conv1DEncoder` is created for
feature extraction.
2. Additional hyperparameters:
`"num_classes"`: int
Number of classes:
- If `> 0`, an additional :torch_nn:`Linear`
layer is appended to the encoder to compute the logits over
classes.
- If `<= 0`, no dense layer is appended. The number of
classes is assumed to be equal to ``out_features`` of the
final dense layer size of the encoder.
`"logit_layer_kwargs"`: dict
Keyword arguments for the logit :torch_nn:`Linear` layer
constructor, except for argument ``out_features`` which is set
to ``"num_classes"``. Ignored if no extra logit layer is
appended.
`"name"`: str
Name of the classifier.
"""
hparams = Conv1DEncoder.default_hparams()
hparams.update({
"name": "conv1d_classifier",
"num_classes": 2, # set to <=0 to avoid appending output layer
"logit_layer_kwargs": {
"in_features": hparams["out_features"],
"bias": True
}
})
return hparams
def default_hparams():
"""Returns a dictionary of hyperparameters with default values.
"""
hparams = Conv1DEncoder.default_hparams()
hparams.update({
"name": "conv1d_classifier",
"num_classes": 2, #set to <=0 to avoid appending output layer
"logit_layer_kwargs": {"use_bias": False}
})
return hparams
def default_hparams():
"""Returns a dictionary of hyperparameters with default values.
.. code-block:: python
{
# (1) Same hyperparameters as in Conv1DEncoder
...
# (2) Additional hyperparameters
"num_classes": 2,
"logit_layer_kwargs": {
"use_bias": False
},
"name": "conv1d_classifier"
}
Here:
1. Same hyperparameters as in :class:`~texar.modules.Conv1DEncoder`.
See the :meth:`~texar.modules.Conv1DEncoder.default_hparams`.
An instance of Conv1DEncoder is created for feature extraction.
2. Additional hyperparameters:
"num_classes" : int
Number of classes:
- If **`> 0`**, an additional :tf_main:`Dense <layers/Dense>` \
layer is appended to the encoder to compute the logits over \
classes.
- If **`<= 0`**, no dense layer is appended. The number of \
classes is assumed to be the final dense layer size of the \
encoder.
"logit_layer_kwargs" : dict
Keyword arguments for the logit Dense layer constructor,
except for argument "units" which is set to "num_classes".
Ignored if no extra logit layer is appended.
"name" : str
Name of the classifier.
"""
hparams = Conv1DEncoder.default_hparams()
hparams.update({
"name": "conv1d_classifier",
"num_classes": 2, #set to <=0 to avoid appending output layer
"logit_layer_kwargs": {
"use_bias": False
}
})
return hparams
