def test_compute_embedding(check_asserts=True):
np.random.seed(0)
random.seed(0)
# sample data from the digits 8x8 pixels dataset
digits = datasets.load_digits()
data = digits.data
n_samples, n_features = data.shape
low_dim = 2
# baseline score using a random 2D projection
projection = random_project(data, target_dim=2, rng=np.random)
score = local_match(data, projection, query_size=50, ratio=0.1, seed=0)
assert_almost_equal(score, 0.12, 2)
# compute an embedding of the data
embedder = SDAEmbedder(
(n_features, 40, 15, low_dim),
noise=0.1,
reconstruction_penalty=0.0,
embedding_penalty=1,
sparsity_penalty=0.0,
learning_rate=0.1,
seed=0,
)
embedder.pre_train(data, epochs=500, batch_size=5)
code = embedder.encode(data)
assert_equal(code.shape, (n_samples, low_dim))
# compare nearest neighbors
score = local_match(data, code, query_size=50, ratio=0.1, seed=0)
assert_almost_equal(score, 0.33, 1)
pl.clf()
n_features = 30
n_samples = 1000
print "Generating embedded swissroll with n_features=%d and n_samples=%d" % (
n_features, n_samples)
data, manifold = swissroll.load(
n_features=n_features,
n_samples=n_samples,
n_turns=1.2,
radius=1.,
hole=False,
)
score_manifold_data = local_match(
data, manifold, query_size=50, ratio=1, seed=0)
print "kNN score match manifold/data:", score_manifold_data
# build model to extract the manifold and learn a mapping / encoder to be able
# to reproduce this on test data
embedder = SDAEmbedder((n_features, 10, 2),
noise=0.1,
reconstruction_penalty=1.0,
embedding_penalty=0.1,
sparsity_penalty=0.0,
learning_rate=0.1, seed=0)
# use the randomly initialized encoder to measure the baseline
code = embedder.encode(data)
score_code_data = local_match(data, code, query_size=50, ratio=1, seed=0)
print "kNN score match after pre-training code/data:", score_code_data
n_samples = 5000
n_manifolds = 100
print "Generating %d embedded swissrolls with n_features=%d, n_samples=%d" % (
n_manifolds, n_features, n_samples)
data, manifolds, colors = multirolls.load(
n_features=n_features,
n_samples=n_samples,
n_manifolds=n_manifolds,
)
# compute a baseline evaluation of the manifolds (ground truth)
stacked_manifolds = np.vstack([m + [10 * i, 0]
for i, m in enumerate(manifolds)])
score = local_match(data, stacked_manifolds, query_size=50, ratio=1, seed=0)
print "kNN score match manifolds/data (ground truth):", score
# compute the score of a projection
projection = random_project(data, target_dim=2, rng=np.random)
score = local_match(data, projection, query_size=50, ratio=1, seed=0)
print "kNN score match projection/data (baseline):", score
# reshuffle the data since stochastic gradient descent assumes I.I.D. samples
perm = np.random.permutation(data.shape[0])
data, colors = data[perm], colors[perm]
# build model to extract the manifolds and learn a mapping / encoder to be able
# to reproduce this on test data
embedder = SDAEmbedder((n_features, 30, 10, 2), noise=0.1,
embedding_penalty=1.0,
embedding_penalty=0.0,
sparsity_penalty=0.0,
learning_rate=0.1, seed=0)
print "Training encoder to extract a semantic preserving 2D mapping"
start = time.time()
embedder.pre_train(data, slice_=slice(None, None), epochs=1000, batch_size=100)
print "done in %ds" % (time.time() - start)
# evaluation of the quality of the embedding by comparing kNN queries from the
# original (high dim) data and the low dim code on the one hand, and from the
# ground truth low dim manifold and the low dim code on the other hand
fig = pl.figure(1)
code = embedder.encode(data)
score_code_data = local_match(data, code, query_size=50, ratio=1, seed=0)
print "kNN score match after pre-training code/data:", score_code_data
_, _, corr = pairwise_distances(data, code, ax=fig.add_subplot(1, 1, 1),
title="pre-training")
print "Pairwise distances correlation:", corr
## fine tuning
#print "Fine tuning encoder to unroll the embedded data..."
#start = time.time()
#embedder.fine_tune(data, epochs=100, batch_size=5)
#print "done in %ds" % (time.time() - start)
#code = embedder.encode(data)
#score_code_data = local_match(data, code, query_size=50, ratio=1, seed=0)
#print "kNN score match after fine-tuning code/data:", score_code_data
#_, _, corr = pairwise_distances(data, code, ax=fig.add_subplot(3, 1, 3),
