Hanbin Lee

I have tested after dropping the first column as you suggested but getting something nearly the same:  The code: ``` X

I supplied 2000 features (feature selection was performed using Seurat's default vst) . By the way, I really thank you for the super fast replies. Trying adjusting the lr. p.s....

 Adjusting the learning rate didn't help (X with dropped intercepts, openTSNE) The next figure was produced after reducing the perplexity but still the same  This one is a...

Not sure but may be some initialization scheme through (computationally much cheaper) PCA might help. Is there an option in GLM-PCA that takes initial latent coordinates?

I think we can close this issue with 0f7fa20 ?

The most straightforward mode of parallelization is splitting the job over windows. After splitting, one can add the results (or average them by some weight) to get the final result....

To answer your question, yes, malloc does matter. Here's the result from `seq_length=1e7` and `num_individuals=1e4` where the weight matrix `100` dimensions.  However, I think it's not necessary to change...

@petrelharp Here's the code.

A simple test case for the `windows` feature. ``` demography = msprime.Demography() demography.add_population(name="A", initial_size=5_000) demography.add_population(name="B", initial_size=5_000) demography.add_population(name="C", initial_size=1_000) demography.add_population_split(time=1000, derived=["A", "B"], ancestral="C") ts = msprime.sim_ancestry( samples={"A": 500, "B": 500}, sequence_length=1e6,...

I just noticed that `centre` doesn't work with `nodes` option. The new commit fixed this problem.