Large image pyramids using dask
Don Venable
turbod33 at gmail.com
Mon Aug 17 10:15:12 EDT 2015
Thanks for the feedback, Matt. Will check this out later this week.
On Friday, August 14, 2015 at 8:29:17 PM UTC-5, Matthew Rocklin wrote:
>
> Nice work. A couple of random notes from a dask.array perspective.
>
> The pyramid_reduce_hdf5 function loads the data from disk, then writes it
> back out. We call this many times with successively smaller datasets. It
> might be nice to read the data only once during this process and do all of
> the writing in a single stream. To do this you might want to look at the
> `da.to_hdf5` function rather than the `.to_hdf5` method.
>
> Note that the map_overlap method overlaps a bit of every block from each
> of its neighbors. This effectively induces a copy which may be slow-ish in
> some cases. If your resizing is aligned very nicely with the block
> structure then this overlap may not be necessary. You might find the
> `da.coarsen` function of interest in this case.
>
> Looking forward to seeing what happens with this,
> -matt
>
>
> On Friday, August 14, 2015 at 12:59:06 PM UTC-7, Don Venable wrote:
>>
>> I've been using scikit-image and dask to play with very large image
>> rasters, and I have a gist for creating an image-pyramid using dask. As the
>> chunk-shapes change during downsampling, you I couldn't use the
>> apply_parallel() function from the existing PR on using dask +
>> scikit-image. Was talking with @blink1073 and he mentioned that it might be
>> useful/interesting to the community, so wanted to see what you thought.
>>
>> Here's a link:
>>
>> https://gist.github.com/venabled/669caeb2615e9ed1d952
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-image/attachments/20150817/ea529b2b/attachment.html>
More information about the scikit-image
mailing list