def test_trainable_variables(self):
r"""Tests the functionality of automatically collecting trainable
variables.
"""
# case 1: bert base
encoder = RoBERTaEncoder()
self.assertEqual(len(encoder.trainable_variables), 2 + 2 + 12 * 16 + 2)
_, _ = encoder(self.inputs)
# case 2: bert large
hparams = {"pretrained_model_name": "roberta-large"}
encoder = RoBERTaEncoder(hparams=hparams)
self.assertEqual(len(encoder.trainable_variables), 2 + 2 + 24 * 16 + 2)
_, _ = encoder(self.inputs)
# case 3: self-designed bert
hparams = {
"encoder": {
"num_blocks": 6,
},
"pretrained_model_name": None,
}
encoder = RoBERTaEncoder(hparams=hparams)
self.assertEqual(len(encoder.trainable_variables), 2 + 2 + 6 * 16 + 2)
_, _ = encoder(self.inputs)
def test_hparams(self):
r"""Tests the priority of the encoder arch parameter.
"""
# case 1: set "pretrained_mode_name" by constructor argument
hparams = {
"pretrained_model_name": "roberta-large",
}
encoder = RoBERTaEncoder(pretrained_model_name="roberta-base",
hparams=hparams)
self.assertEqual(encoder.hparams.encoder.num_blocks, 12)
_, _ = encoder(self.inputs)
# case 2: set "pretrained_mode_name" by hparams
hparams = {
"pretrained_model_name": "roberta-large",
"encoder": {
"num_blocks": 6,
}
}
encoder = RoBERTaEncoder(hparams=hparams)
self.assertEqual(encoder.hparams.encoder.num_blocks, 24)
_, _ = encoder(self.inputs)
# case 3: set to None in both hparams and constructor argument
hparams = {
"pretrained_model_name": None,
"encoder": {
"num_blocks": 6,
},
}
encoder = RoBERTaEncoder(hparams=hparams)
self.assertEqual(encoder.hparams.encoder.num_blocks, 6)
_, _ = encoder(self.inputs)
# case 4: using default hparams
encoder = RoBERTaEncoder()
self.assertEqual(encoder.hparams.encoder.num_blocks, 12)
_, _ = encoder(self.inputs)
def test_soft_ids(self):
r"""Tests soft ids.
"""
hparams = {
"pretrained_model_name": None,
}
encoder = RoBERTaEncoder(hparams=hparams)
inputs = torch.rand(self.batch_size, self.max_length, 50265)
outputs, pooled_output = encoder(inputs)
outputs_dim = encoder.hparams.encoder.dim
self.assertEqual(
outputs.shape,
torch.Size([self.batch_size, self.max_length, outputs_dim]))
self.assertEqual(pooled_output.shape,
torch.Size([self.batch_size, encoder.output_size]))
def default_hparams():
r"""Returns a dictionary of hyperparameters with default values.
.. code-block:: python
{
# (1) Same hyperparameters as in RoBertaEncoder
...
# (2) Additional hyperparameters
"num_classes": 2,
"logit_layer_kwargs": None,
"clas_strategy": "cls_time",
"max_seq_length": None,
"dropout": 0.1,
"name": "roberta_classifier"
}
Here:
1. Same hyperparameters as in
:class:`~texar.torch.modules.RoBERTaEncoder`.
See the :meth:`~texar.torch.modules.RoBERTaEncoder.default_hparams`.
An instance of RoBERTaEncoder is created for feature extraction.
2. Additional hyperparameters:
`"num_classes"`: int
Number of classes:
- If **> 0**, an additional `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 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.
`"clas_strategy"`: str
The classification strategy, one of:
- **cls_time**: Sequence-level classification based on the
output of the first time step (which is the `CLS` token).
Each sequence has a class.
- **all_time**: Sequence-level classification based on
the output of all time steps. Each sequence has a class.
- **time_wise**: Step-wise classification, i.e., make
classification for each time step based on its output.
`"max_seq_length"`: int, optional
Maximum possible length of input sequences. Required if
`clas_strategy` is `all_time`.
`"dropout"`: float
The dropout rate of the RoBERTa encoder output.
`"name"`: str
Name of the classifier.
"""
hparams = RoBERTaEncoder.default_hparams()
hparams.update({
"num_classes": 2,
"logit_layer_kwargs": None,
"clas_strategy": "cls_time",
"max_seq_length": None,
"dropout": 0.1,
"name": "roberta_classifier"
})
return hparams
def test_encode(self):
r"""Tests encoding.
"""
# case 1: bert base
hparams = {
"pretrained_model_name": None,
}
encoder = RoBERTaEncoder(hparams=hparams)
inputs = torch.randint(30521, (self.batch_size, self.max_length))
outputs, pooled_output = encoder(inputs)
outputs_dim = encoder.hparams.encoder.dim
self.assertEqual(
outputs.shape,
torch.Size([self.batch_size, self.max_length, outputs_dim]))
self.assertEqual(pooled_output.shape,
torch.Size([self.batch_size, encoder.output_size]))
# case 2: self-designed bert
hparams = {
'pretrained_model_name': None,
'embed': {
'dim': 96,
},
'position_embed': {
'dim': 96,
},
'encoder': {
'dim': 96,
'multihead_attention': {
'num_units': 96,
'output_dim': 96,
},
'poswise_feedforward': {
'layers': [{
'kwargs': {
'in_features': 96,
'out_features': 96 * 4,
'bias': True,
},
'type': 'Linear',
}, {
"type": "BertGELU"
}, {
'kwargs': {
'in_features': 96 * 4,
'out_features': 96,
'bias': True,
},
'type': 'Linear',
}]
},
},
'hidden_size': 96,
}
encoder = RoBERTaEncoder(hparams=hparams)
outputs, pooled_output = encoder(inputs)
outputs_dim = encoder.hparams.encoder.dim
self.assertEqual(
outputs.shape,
torch.Size([self.batch_size, self.max_length, outputs_dim]))
self.assertEqual(pooled_output.shape,
torch.Size([self.batch_size, encoder.output_size]))
def test_model_loading(self):
r"""Tests model loading functionality."""
for pretrained_model_name in RoBERTaEncoder.available_checkpoints():
encoder = RoBERTaEncoder(
pretrained_model_name=pretrained_model_name)
_, _ = encoder(self.inputs)
