Open lsc64 opened 3 months ago
I strong recommend to not pass the computed array darr.compute()
to pairwise_distance
. It should just be darr
. This will otherwise materialize the entire array into memory and will require you to send all of this over the network.
I suspect that the bug you are running into is actually already fixed in https://github.com/dask/distributed/pull/8507 but you wouldn't have a good time submitting 12TB from your client to the scheduler.
I strong recommend to not pass the computed array
darr.compute()
topairwise_distance
. It should just bedarr
. This will otherwise materialize the entire array into memory and will require you to send all of this over the network.I suspect that the bug you are running into is actually already fixed in dask/distributed#8507 but you wouldn't have a good time submitting 12TB from your client to the scheduler.
Unfortunately that functions requires a dask.array and a numpy.array, otherwise it would of course be nicer to not do that. https://github.com/dask/dask-ml/blob/b95ba909c6dcd37c566f5193ba0b918396edaaee/dask_ml/metrics/pairwise.py#L58
if isinstance(Y, da.Array):
raise TypeError("`Y` must be a numpy array")
If I batch the materialized array into 100k slices (which reduces the graph size) it works, so you're probably right!
hists = []
batch_size = 100000
for batch in tqdm(range(darr.shape[0] // batch_size)):
distances = pairwise_distances(
darr,
darr[
batch * batch_size : min((batch + 1) * batch_size, darr.shape[0])
].compute(),
metric="cosine",
)
hist, bins = da.histogram(distances, bins=100, range=[0, 2])
hists.append(hist)
da.compute(hists) # works, still computes everything at once
Do I have the patch if I install from source?
Unfortunately that functions requires a dask.array and a numpy.array,
Sorry, I missed that. I haven't tried to understand your batching code to ensure if it is correct. If it is, maybe you want to contribute this to dask-ml because a "proper" dask algorithm works similarly. I don't know enough about the pairwise_distance algorithm to tell
However, what I can tell you is that if you include a 12TB array in the map_blocks
call of https://github.com/dask/dask-ml/blob/b95ba909c6dcd37c566f5193ba0b918396edaaee/dask_ml/metrics/pairwise.py#L60-L67 this will replicate that array to the scheduler and all dask workers. I doubt this is what you want to do.
Do I have the patch if I install from source?
I just checked and this was already released in 2024.2.1 (the version you are running on). By breaking up the array you are avoiding all sorts of problems so if this is possible, go for it.
Sorry, I missed that. I haven't tried to understand your batching code to ensure if it is correct. If it is, maybe you want to contribute this to dask-ml because a "proper" dask algorithm works similarly. I don't know enough about the pairwise_distance algorithm to tell
No worries, I just like to leave code snippets in case anyone has the same issue, so they're not faced the unhelpful "nvm I solved it". I can open a PR at some point and discuss this over there.
However, what I can tell you is that if you include a 12TB array in the map_blocks call of https://github.com/dask/dask-ml/blob/b95ba909c6dcd37c566f5193ba0b918396edaaee/dask_ml/metrics/pairwise.py#L60-L67 this will replicate that array to the scheduler and all dask workers. I doubt this is what you want to do.
For X.map_blocks(fn, Y)
the array Y
gets fully replicated, X
or both? But that array/those arrays are not materialized right?
The docs give an example, which is exactly what I/the function wants to achieve (just for huge arrays and lambda a,b,: distance(a,b)
)
d = da.arange(5, chunks=2)
e = da.arange(5, chunks=2)
f = da.map_blocks(lambda a, b: a + b**2, d, e)
f.compute()
I just checked and this was already released in 2024.2.1 (the version you are running on). By breaking up the array you are avoiding all sorts of problems so if this is possible, go for it.
We need bigger graphs!! /s (but maybe actually)
Describe the issue: I'm trying to compute a histogram over a 12 TB array of pairwise distances and it fails. Returns either
ValueError: memoryview is too large
or just cancels
Minimal Complete Verifiable Example:
Anything else we need to know?: Just computing the histogram of such a large matrix works
Environment: