def test_can_pass_callbacks_to_tsne_object(self):
callback = MagicMock()
callback2 = MagicMock()
# We don't want individual callbacks to be iterable
del callback.__iter__
del callback2.__iter__
# Should be able to pass a single callback
TSNE(callbacks=callback,
callbacks_every_iters=1,
early_exaggeration_iter=0,
n_iter=1).fit(self.x)
self.assertEqual(callback.call_count, 1)
# Should be able to pass a list callbacks
callback.reset_mock()
TSNE(callbacks=[callback],
callbacks_every_iters=1,
early_exaggeration_iter=0,
n_iter=1).fit(self.x)
self.assertEqual(callback.call_count, 1)
# Should be able to change the callback on the object
callback.reset_mock()
tsne = TSNE(callbacks=callback,
callbacks_every_iters=1,
early_exaggeration_iter=0,
n_iter=1)
tsne.callbacks = callback2
tsne.fit(self.x)
callback.assert_not_called()
self.assertEqual(callback2.call_count, 1)
def test_raises_error_on_unrecognized_metric(self):
"""Unknown distance metric should raise error"""
tsne = TSNE(metric='imaginary', neighbors='exact')
with self.assertRaises(ValueError):
tsne.prepare_initial(self.x)
tsne = TSNE(metric='imaginary', neighbors='approx')
with self.assertRaises(ValueError):
tsne.prepare_initial(self.x)
def test_unfitted_pca_model(self):
"""Using PCA initialization in `transform` should fail when the initial
embedding was initialized with PCA."""
tsne = TSNE(initialization='random')
embedding = tsne.fit(self.x)
# Transforming using `pca` init on embedding that did not use
# `pca` init did not fail
with self.assertRaises(AssertionError):
embedding.transform(self.x_test, initialization='pca')
def test_low_variance(self):
"""Low variance of the initial embedding is very important for the
convergence of tSNE."""
# Cycle through various initializations
initializations = ['random', 'pca']
allowed = 1e-3
for init in initializations:
tsne = TSNE(initialization=init, perplexity=2)
embedding = tsne.prepare_initial(self.x)
np.testing.assert_array_less(
np.var(embedding, axis=0), allowed,
'using the `%s` initialization' % init)
def check_error_approx():
x, y = get_mouse_60k(1500)
tsne = TSNE(
perplexity=20,
learning_rate=100,
early_exaggeration=12,
n_jobs=4,
theta=0.5,
initialization='pca',
metric='euclidean',
n_components=2,
n_iter=750,
early_exaggeration_iter=250,
neighbors='exact',
negative_gradient_method='bh',
min_num_intervals=10,
ints_in_interval=2,
late_exaggeration_iter=0,
late_exaggeration=4,
callbacks=ErrorLogger(),
)
embedding = tsne.prepare_initial(x, initialization='random')
errors = ErrorApproximations(embedding.affinities.P)
logger = ErrorLogger()
embedding.optimize(
250,
exaggeration=12,
callbacks=[errors, logger],
callbacks_every_iters=5,
inplace=True,
)
embedding.optimize(
750,
exaggeration=None,
callbacks=[errors, logger],
callbacks_every_iters=5,
inplace=True,
)
errors.report()
plot(embedding, y)
x = list(range(len(errors.exact_errors)))
plt.semilogy(x, errors.exact_errors, label='Exact')
plt.semilogy(x, errors.bh_errors, label='BH')
plt.semilogy(x, errors.fft_errors, label='FFT')
plt.legend()
plt.show()
def transform(n_jobs=4, grad='bh', neighbors='approx'):
# iris = datasets.load_iris()
# x, y = iris['data'], iris['target']
x, y = get_mnist(20000)
x_train, x_test, y_train, y_test = train_test_split(x,
y,
test_size=0.33,
random_state=42)
tsne = TSNE(
n_components=2,
perplexity=30,
learning_rate=100,
early_exaggeration=12,
n_jobs=n_jobs,
theta=0.5,
initialization='random',
metric='euclidean',
n_iter=750,
early_exaggeration_iter=250,
neighbors=neighbors,
negative_gradient_method=grad,
min_num_intervals=10,
ints_in_interval=2,
late_exaggeration_iter=0,
late_exaggeration=4,
callbacks=[ErrorLogger()],
)
start = time.time()
embedding = tsne.fit(x_train)
print('tsne train', time.time() - start)
plt.subplot(121)
plot(embedding, y_train, show=False, ms=3)
start = time.time()
partial_embedding = embedding.transform(x_test, perplexity=20)
# partial_embedding = embedding.get_partial_embedding_for(
# x_test, perplexity=10, initialization='random')
# partial_embedding.optimize(200, exaggeration=2, inplace=True, momentum=0.1)
print('tsne transform', time.time() - start)
plt.subplot(122)
plot(embedding, y_train, show=False, ms=3, alpha=0.25)
plt.gca().set_color_cycle(None)
plot(partial_embedding, y_test, show=False, ms=3, alpha=0.8)
plt.show()
def test_embedding_optimize(self, param_name, param_value,
gradient_descent):
# type: (str, Any, MagicMock) -> None
# Make sure mock still conforms to signature
gradient_descent.return_value = (1, MagicMock())
# `optimize` requires us to specify the `n_iter`
params = {'n_iter': 50, param_name: param_value}
tsne = TSNE()
embedding = tsne.prepare_initial(self.x)
embedding.optimize(**params, inplace=True)
self.assertEqual(1, gradient_descent.call_count)
check_call_contains_kwargs(gradient_descent.mock_calls[0], params)
def test_embedding_transform(self, param_name, param_value,
gradient_descent):
# type: (str, Any, MagicMock) -> None
# Make sure mock still conforms to signature
gradient_descent.return_value = (1, MagicMock())
# Perform initial embedding - this is tested above
tsne = TSNE()
embedding = tsne.fit(self.x)
gradient_descent.reset_mock()
embedding.transform(self.x_test, **{param_name: param_value})
# Early exaggeration training loop
if param_name in ('early_exaggeration_iter', 'early_exaggeration'):
call_idx = 0
# Main training loop
elif param_name in ('n_iter', 'final_momentum'):
call_idx = 1
# If general parameter, should be applied to every call
else:
call_idx = 0
self.assertEqual(2, gradient_descent.call_count)
check_call_contains_kwargs(
gradient_descent.mock_calls[call_idx],
{param_name: param_value},
)
def setUpClass(cls):
cls.tsne = TSNE(early_exaggeration_iter=20, n_iter=100)
# Set up two modalities, if we want to viually inspect test results
cls.x = np.vstack((
np.random.normal(+1, 1, (100, 4)),
np.random.normal(-1, 1, (100, 4)),
))
cls.x_test = np.random.normal(0, 1, (25, 4))
def run_graph():
graph = nx.read_edgelist(join(DATA_DIR, 'dolphins.edges'))
affinities = NxGraphAffinities(graph)
tsne = TSNE()
y_coords = tsne.generate_initial_coordinates(affinities.P,
initialization='random')
embedding = TSNEEmbedding(
y_coords, affinities, {
'callbacks': None,
'negative_gradient_method': 'bh',
'dof': 1,
'momentum': 0,
'learning_rate': 100
})
embedding.optimize(1000)
plt.plot(embedding[:, 0], embedding[:, 1], 'o')
plt.show()
def test_same_partial_embedding_on_fixed_random_state(self):
tsne = TSNE(random_state=1, initialization='random')
embedding = tsne.fit(self.x)
partial1 = embedding.prepare_partial(self.x_test,
initialization='random')
partial2 = embedding.prepare_partial(self.x_test,
initialization='random')
np.testing.assert_array_equal(
partial1, partial2,
'Same random state produced different partial embeddings')
def test_mismatching_embedding_dimensions_simple_api(self):
# Fit
tsne = TSNE(n_components=2, initialization=self.x[:10, :2])
with self.assertRaises(ValueError,
msg='fit::ncorrect number of points'):
tsne.fit(self.x[:25])
with self.assertRaises(ValueError,
msg='fit::ncorrect number of dimensions'):
TSNE(n_components=2, initialization=self.x[:10, :4])
# Transform
tsne = TSNE(n_components=2, initialization='random')
embedding = tsne.fit(self.x)
with self.assertRaises(ValueError,
msg='transform::incorrect number of points'):
embedding.transform(X=self.x[:5], initialization=self.x[:10, :2])
with self.assertRaises(
ValueError, msg='transform::incorrect number of dimensions'):
embedding.transform(X=self.x, initialization=self.x[:, :4])
def test_iris(self):
iris = datasets.load_iris()
x, y = iris['data'], iris['target']
# Evaluate tSNE optimization using a KNN classifier
knn = KNeighborsClassifier(n_neighbors=10)
tsne = TSNE(perplexity=30, initialization='random', random_state=0)
# Prepare a random initialization
embedding = tsne.prepare_initial(x)
# KNN should do poorly on a random initialization
knn.fit(embedding, y)
predictions = knn.predict(embedding)
self.assertTrue(accuracy_score(predictions, y) < .5)
# Optimize the embedding for a small number of steps so tests run fast
embedding.optimize(50, inplace=True)
# Similar points should be grouped together, therefore KNN should do well
knn.fit(embedding, y)
predictions = knn.predict(embedding)
self.assertTrue(accuracy_score(predictions, y) > .95)
def test_same_results_on_fixed_random_state_pca_init(self):
"""Results should be exactly the same if we provide a random state."""
tsne1 = TSNE(random_state=1, initialization='pca')
embedding1 = tsne1.fit(self.x)
tsne2 = TSNE(random_state=1, initialization='pca')
embedding2 = tsne2.fit(self.x)
np.testing.assert_array_equal(
embedding1, embedding2,
'Same random state produced different initial embeddings')
def test_nndescent_distances(self, param_name, metric,
nndescent: MagicMock):
"""Distance metrics should be properly passed down to NN descent"""
assert param_name == 'metric'
tsne = TSNE(metric=metric, neighbors='approx')
# We don't care about what happens later, just that the NN method is
# properly called
nndescent.side_effect = InterruptedError()
try:
# Haversine distance only supports two dimensions
tsne.prepare_initial(self.x[:, :2])
except InterruptedError:
pass
self.assertEqual(nndescent.call_count, 1)
check_call_contains_kwargs(nndescent.mock_calls[0], {'metric': metric})
def test_nndescent_mahalanobis_distance(self, nndescent: MagicMock):
"""Distance metrics and additional params should be correctly passed down to NN descent"""
metric = 'mahalanobis'
C = np.cov(self.x)
tsne = TSNE(metric=metric, metric_params={'V': C}, neighbors='approx')
# We don't care about what happens later, just that the NN method is
# properly called
nndescent.side_effect = InterruptedError()
try:
tsne.prepare_initial(self.x)
except InterruptedError:
pass
self.assertEqual(nndescent.call_count, 1)
check_call_contains_kwargs(nndescent.mock_calls[0], {'metric': metric})
def test_constructor(self, param_name, param_value, gradient_descent):
# type: (str, Any, MagicMock) -> None
# Make sure mock still conforms to signature
gradient_descent.return_value = (1, MagicMock())
# Early exaggeration training loop
if param_name == 'early_exaggeration_iter':
check_param_name = 'n_iter'
call_idx = 0
elif param_name == 'early_exaggeration':
check_param_name = 'exaggeration'
call_idx = 0
elif param_name == 'initial_momentum':
check_param_name = 'momentum'
call_idx = 0
# Main training loop
elif param_name == 'n_iter':
check_param_name = param_name
call_idx = 1
elif param_name == 'final_momentum':
check_param_name = 'momentum'
call_idx = 1
# Early exaggeration training loop
elif param_name == 'late_exaggeration_iter':
check_param_name = 'n_iter'
call_idx = 2
elif param_name == 'late_exaggeration':
check_param_name = 'exaggeration'
call_idx = 2
# If general parameter, should be applied to every call
else:
check_param_name = param_name
call_idx = 0
TSNE(**{param_name: param_value}).fit(self.x)
self.assertEqual(3, gradient_descent.call_count)
check_call_contains_kwargs(gradient_descent.mock_calls[call_idx],
{check_param_name: param_value})
def run(perplexity=30, learning_rate=100, n_jobs=4):
x, y = get_mouse_60k()
# x, y = get_fashion_mnist()
angle = 0.5
ee = 12
metric = 'euclidean'
print(x.shape)
start = time.time()
tsne = TSNE(
perplexity=perplexity,
learning_rate=learning_rate,
early_exaggeration=ee,
n_jobs=n_jobs,
theta=angle,
initialization='random',
metric=metric,
n_components=2,
n_iter=750,
early_exaggeration_iter=250,
neighbors='approx',
negative_gradient_method='fft',
min_num_intervals=10,
ints_in_interval=1,
late_exaggeration_iter=0,
late_exaggeration=2.,
callbacks=ErrorLogger(),
)
# x = PCA(n_components=50).fit_transform(x)
embedding = tsne.fit(x)
print('-' * 80)
print('tsne', time.time() - start)
plt.title('tsne')
plot(embedding, y)
return
x = np.ascontiguousarray(x.astype(np.float64))
from fitsne import FItSNE
start = time.time()
embedding = FItSNE(
x,
2,
perplexity=perplexity,
stop_lying_iter=250,
ann_not_vptree=True,
early_exag_coeff=ee,
nthreads=n_jobs,
theta=angle,
)
print('-' * 80)
print('fft interp %.4f' % (time.time() - start))
plt.title('fft interp')
plot(embedding, y)
plt.show()
return
init = PCA(n_components=2).fit_transform(x)
start = time.time()
embedding = MulticoreTSNE(early_exaggeration=ee,
learning_rate=learning_rate,
perplexity=perplexity,
n_jobs=n_jobs,
cheat_metric=False,
angle=angle,
init=init,
metric=metric,
verbose=True).fit_transform(x)
print('-' * 80)
print('mctsne', time.time() - start)
plt.title('mctsne')
plot(embedding, y)
plt.show()
start = time.time()
embedding = SKLTSNE(
early_exaggeration=ee,
learning_rate=learning_rate,
angle=angle,
perplexity=perplexity,
init='pca',
metric=metric,
).fit_transform(x)
print('-' * 80)
print('sklearn', time.time() - start)
plt.title('sklearn')
plot(embedding, y)
plt.show()
def setUpClass(cls):
cls.tsne = TSNE()
cls.x = np.random.randn(100, 4)
cls.x_test = np.random.randn(25, 4)
def setUpClass(cls):
cls.tsne = TSNE()
random_state = np.random.RandomState(42)
cls.x = random_state.randn(100, 4)
cls.x_test = random_state.randn(25, 4)
def test_fitted_pca_model(self):
"""Using PCA initialization in `transform` should work when the initial
embedding was initialized with PCA."""
tsne = TSNE(initialization='pca')
embedding = tsne.fit(self.x)
embedding.transform(self.x_test, initialization='pca')