jakirkham
commented
4 months ago One test was slightly off from its tolerance on CI:
```python
____________ test_tsne_distance_metrics_on_sparse_input[cosine-fft] ____________
[gw7] linux -- Python 3.10.14 /opt/conda/envs/test/bin/python
method = 'fft', metric = 'cosine'
@pytest.mark.parametrize("method", ["fft", "barnes_hut", "exact"])
@pytest.mark.parametrize(
"metric", ["l2", "euclidean", "cityblock", "l1", "manhattan", "cosine"]
)
def test_tsne_distance_metrics_on_sparse_input(method, metric):
data, labels = make_blobs(
n_samples=1000, n_features=64, centers=5, random_state=42
)
data_sparse = scipy.sparse.csr_matrix(data)
cuml_tsne = TSNE(
n_components=2,
random_state=1,
n_neighbors=DEFAULT_N_NEIGHBORS,
method=method,
learning_rate_method="none",
min_grad_norm=1e-12,
perplexity=DEFAULT_PERPLEXITY,
metric=metric,
)
if method == "fft":
sk_tsne = skTSNE(
n_components=2,
random_state=1,
min_grad_norm=1e-12,
method="barnes_hut",
perplexity=DEFAULT_PERPLEXITY,
metric=metric,
init="random",
)
else:
sk_tsne = skTSNE(
n_components=2,
random_state=1,
min_grad_norm=1e-12,
method=method,
perplexity=DEFAULT_PERPLEXITY,
metric=metric,
init="random",
)
cuml_embedding = cuml_tsne.fit_transform(data_sparse)
nans = np.sum(np.isnan(cuml_embedding))
sk_embedding = sk_tsne.fit_transform(data_sparse)
cu_trust = trustworthiness(data, cuml_embedding, metric=metric)
sk_trust = trustworthiness(data, sk_embedding, metric=metric)
> assert cu_trust > 0.85
E assert 0.8022727822580645 > 0.85
```
Restarted it
In any event this is unrelated to this change
KyleFromNVIDIA
After updating everything to CUDA 12.5.1, use
shared-workflows@branch-24.08again.Contributes to https://github.com/rapidsai/build-planning/issues/73