def test_fine_tuneable(self):
tw = TextWiser(Embedding.TfIdf(min_df=2),
Transformation.SVD(n_components=2),
dtype=torch.float32,
is_finetuneable=True)
tw.fit(docs)
embeddings1 = tw._imp[1].V.data.clone().detach()
# Give a fake task to train embeddings on
# Have a linear layer with a single output after pooling
linear = nn.Linear(2, 1, bias=False)
model = nn.Sequential(tw, linear).to(device).train()
y_pred = model(docs)
# Use ones as the target
y_act = torch.ones_like(y_pred)
# Optimize MSE using SGD
criterion = nn.MSELoss()
optimizer = optim.SGD(model.parameters(), lr=1e-3)
# Calculate the loss & gradients
optimizer.zero_grad()
loss = criterion(y_pred, y_act)
loss.backward()
# The embedding layer should have gradients now
self.assertIsNotNone([p for p in tw._imp[1].named_parameters()
][0][1].grad)
# Update weights
optimizer.step()
# The weights should be updated if fine_tune is true, else it should be the same
self.assertFalse(torch.allclose(embeddings1, tw._imp[1].V.data))
def test_v_in_parameters(self):
n_components = 2 # Restrict the # of components
tw = TextWiser(Embedding.TfIdf(min_df=2),
Transformation.SVD(n_components=n_components),
dtype=torch.float32)
tw.fit(docs)
self.assertIn('_imp.1.V', [p[0] for p in tw.named_parameters()])
def test_save_load(self):
try:
os.environ[
'TF_CPP_MIN_LOG_LEVEL'] = '3' # shut tensorflow up during testing
# Create a model with a downstream task
tw = TextWiser(Embedding.USE(), dtype=torch.float32).fit(docs)
model = nn.Sequential(tw, nn.Linear(512, 1)).to(device)
# Get results of the model
expected = model(docs)
# Save the model to a temporary file
with NamedTemporaryFile() as file:
state_dict = model.state_dict()
self.assertNotIn('0._imp.0.use', state_dict)
torch.save(state_dict, file) # Use string name of the file
# Get rid of the original model
del tw
del model
# Create the same model
tw = TextWiser(Embedding.USE(), dtype=torch.float32)
tw.fit()
model = nn.Sequential(tw, nn.Linear(512, 1)).to(device)
# Load the model from file
file.seek(0)
model.load_state_dict(torch.load(file, map_location=device))
# Do predictions with the loaded model
predicted = model(docs)
self.assertTrue(torch.allclose(predicted, expected, atol=1e-6))
except ModuleNotFoundError:
print('No Tensorflow found. Skipping the test. ...',
end=" ",
flush=True)
def test_save_load(self):
# Create a model with a downstream task
tw = TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en-turian'),
[
Transformation.SVD(n_components=2),
Transformation.Pool(pool_option=PoolOptions.mean)
],
dtype=torch.float32)
tw.fit(docs)
model = nn.Sequential(tw, nn.Linear(2, 1)).to(device)
# Get results of the model
expected = model(docs)
# Save the model to a temporary file
with NamedTemporaryFile() as file:
torch.save(model.state_dict(), file) # Use string name of the file
# Get rid of the original model
del tw
del model
# Create the same model
tw = TextWiser(Embedding.Word(
word_option=WordOptions.word2vec, pretrained='en-turian'), [
Transformation.SVD(n_components=2),
Transformation.Pool(pool_option=PoolOptions.mean)
],
dtype=torch.float32)
tw.fit()
model = nn.Sequential(tw, nn.Linear(2, 1)).to(device)
# Load the model from file
file.seek(0)
model.load_state_dict(torch.load(file, map_location=device))
# Do predictions with the loaded model
predicted = model(docs)
self.assertTrue(torch.allclose(predicted, expected, atol=1e-6))
def test_set_params(self):
# Set the arguments in container classes
tw = TextWiser(Embedding.TfIdf(min_df=5),
Transformation.NMF(n_components=30),
lazy_load=True)
tw.set_params(embedding__min_df=10,
transformations__0__n_components=10)
self.assertEqual(tw.embedding.min_df, 10)
self.assertEqual(tw.transformations[0].n_components, 10)
# Set the arguments in implementation
tw = TextWiser(Embedding.Doc2Vec(vector_size=2, min_count=1,
workers=1))
tw.fit(docs)
tw.set_params(_imp__0__seed=10)
self.assertEqual(tw._imp[0].seed, 10)
# Set the arguments in a schema
schema = {'transform': ['tfidf', ['nmf', {'n_components': 30}]]}
tw = TextWiser(Embedding.Compound(schema=schema))
tw.set_params(embedding__schema__transform__0__min_df=10,
embedding__schema__transform__1__n_components=10)
self.assertEqual(tw.embedding.schema['transform'][0][1]['min_df'], 10)
self.assertEqual(
tw.embedding.schema['transform'][1][1]['n_components'], 10)
# Replace a part of the schema in a list
tw.set_params(embedding__schema__transform__0='bow')
self.assertEqual(tw.embedding.schema['transform'][0], 'bow')
# Replace a part of the schema
tw.set_params(embedding__schema__transform=['bow'])
self.assertEqual(tw.embedding.schema['transform'][0], 'bow')
def test_lazy_load(self):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
tw = TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en-turian'),
lazy_load=True)
self.assertIsNone(tw._imp)
tw.fit(docs)
self.assertIsNotNone(tw._imp)
tw = TextWiser(Embedding.Word(word_option=WordOptions.word2vec,
pretrained='en-turian'),
lazy_load=True,
dtype=torch.float32,
is_finetuneable=True)
self.assertIsNone(tw._imp)
tw.fit_transform(docs)
self.assertIsNotNone(tw._imp)