quinngroup/dr1dl-dask

Dictionary Learning in Dask

dr1dl-dask's Issues

Implement CuPy/Cuda

How to make this function work with or without a GPU available. What is the best way? Checking if this https://github.com/anderskm/gputil will return anything (None type if no GPUs available) i.e.

GPUtil.showUtilization()

OR:

try: 
  import cupy
  make cupy array
else:
  Make dask array

Titile of Thesis/Joss paper

I want this thread to be used for coming up with names for for my thesis and JOSS submission.

Thesis:
Rank 1 Dictionary Learning: Comparing Distributed Frameworks
Comparing Dask and Spark: Scalable Rank 1 Dictionary Learning

JOSS Paper:
dR1DL-Dask: Scalable Rank 1 Dictionary Learning in Dask

Issues with CuPy and the Sparse Implementation

Currently implemented is pydata's sparse.COO but it isn't compatable with cupy yet! It is part of their future plans, but currently not.

Other option but not really. Cupy has scipy coo_matrix implemented but it doesn't support 1D vectors.

The S Matrix is a cupy array, so the 1D array must be cupy or compatible with cupy and numpy aka sparse.COO

Outer Product with sparse vector and dask array

A = da.random.random(200) 
B = da.random.random(200)
B[B<.90] = 0
B = B.map_blocks(sparse.COO)
da.outer(A,B.compute().todense())

This test works great. A is a dask array and B is a Dask Sparse array. By calling B.compute() it converts it back to a sparse.COO array. In order for the outer function to work, we have to have the dense the sparse array by calling todense().

In the code, by the time the sparse array goes through calculation da.outer will no longer accept the type.

Command Line arguments

In the Pyspark implementation, there were two arguments that I really had to do anything with.

1. Partitions --> Chunks
   • Dask Chunks
   • If an int x is supplied then each chunk will be (x,x)
   • A tuple of things can be applied as well so give Dask a more specific structure of all the chunks
   • I changed the default value to 'auto' because that is the default for Dask.
2. execmem
   • I can't really decide if this is needed. I have checked the parameters of the client and there isnt a way to establish how much memory a worker should have. Plus when run locally, Dask sets its regardless in the LocalCluster class.