def test_fit_transform(self):
tw = TextWiser(Embedding.TfIdf(min_df=2),
Transformation.LDA(n_components=2),
dtype=torch.float32)
expected = torch.tensor(
[[0.7724367976, 0.2275632024], [0.5895692706, 0.4104307294],
[0.2381444573, 0.7618555427]],
dtype=torch.float32)
self._test_fit_transform(tw, expected)
self._reset_seed()
self._test_fit_before_transform(tw, expected)
def test_finetune_validation(self):
# Nothing is fine-tuneable if dtype is numpy
with self.assertRaises(TypeError):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en_turian'),
dtype=np.float32,
is_finetuneable=True)
# Word2Vec is fine-tuneable
try:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en-turian'),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
except ValueError:
self.fail("Word2vec is fine tuneable")
# ELMo is not fine-tuneable, and should raise an error
with self.assertRaises(ValueError):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
TextWiser(Embedding.Word(word_option=WordOptions.elmo),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
# TfIdf is not fine-tuneable, and should raise an error
with self.assertRaises(ValueError):
TextWiser(Embedding.TfIdf(),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
# TfIdf is not fine-tuneable, but SVD is
try:
TextWiser(Embedding.TfIdf(),
Transformation.SVD(),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
except ValueError:
self.fail("SVD is fine tuneable")
# LDA cannot propagate gradients, so the whole thing is not fine-tuneable
with self.assertRaises(ValueError):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en'),
Transformation.LDA(),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
schema = {
'concat': [{
'transform': [('word2vec', {
'pretrained': 'en-turian'
}), ('pool', {
'pool_option': 'max'
})]
}, {
'transform': ['tfidf', ('nmf', {
'n_components': 30
})]
}]
}
# Word2Vec is fine-tuneable, therefore the whole schema is fine-tuneable
try:
TextWiser(Embedding.Compound(schema=schema),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
except ValueError:
self.fail(
"Any fine-tuneable weights is enough for the model to be fine-tuneable"
)
# TfIdf is not fine-tuneable, but SVD is
schema = {'transform': ['tfidf', 'svd']}
try:
TextWiser(Embedding.Compound(schema=schema),
dtype=torch.float32,
is_finetuneable=True,
lazy_load=True)
except ValueError:
self.fail("SVD is fine tuneable")
def test_min_components(self):
with self.assertRaises(ValueError):
TextWiser(Embedding.TfIdf(min_df=2),
Transformation.LDA(n_components=1),
dtype=torch.float32)