Add memoization to dynamic Callable class #1063

philippjfr · 2017-01-15T21:19:20Z

In order to address #1049 the Callable class now memoizes the last returned value from a DynamicMap, based on the args, kwargs and nested stream values.

Needs docstrings and tests. @jordansamuels this seems to address your issue. Small example for testing:

import holoviews as hv
import numpy as np
import pandas as pd
from holoviews.operation.datashader import datashade
from holoviews.streams import PositionXY

hv.notebook_extension('bokeh')

%opts RGB [width=800, height=450] VLine HLine (line_color='black' line_width=0.5)

n=100000
def f(n):
    return np.cumsum(np.random.randn(n))
df = pd.DataFrame({'t': range(n), 'y1': f(n), 'y2': f(n), 'y3' : .10 * f(n)} )

def p(y):
    return hv.Curve(df, kdims=['t'], vdims=[y])

ts = datashade(p('y1'))
ch = hv.DynamicMap(lambda x, y: hv.HLine(y) * hv.VLine(x) * hv.Text(x+0.15, y, '%.3f' % x + ", %.3f" % y),
              kdims=[], streams=[PositionXY()])
(ts * datashade(p('y2')) * ch + ts).cols(1)

jbednar · 2017-01-16T05:45:08Z

Wow! And the total amount of code even went down. Fabulous!

philippjfr · 2017-01-16T13:54:55Z

@jlstevens Requested review, the code is fairly straightforward and I've added at least one test to ensure the memoization behaves correctly. There are already a number of tests to check memoization works during plotting, stopping the callback from being called repeatedly.

jlstevens · 2017-01-16T15:10:21Z

holoviews/core/spaces.py

-    to a DynamicMap.
+    to a DynamicMap. Additionally a Callable will memoize the last
+    returned value based on the arguments to the function and the
+    state of all streams on its inputs, avoiding calling the function


I think either 'avoiding calls to the function' or 'to avoid calling the function ...' would be read better.

jlstevens · 2017-01-16T15:10:44Z

holoviews/core/spaces.py

+    to a DynamicMap. Additionally a Callable will memoize the last
+    returned value based on the arguments to the function and the
+    state of all streams on its inputs, avoiding calling the function
+    unncessarily.
    """


jlstevens · 2017-01-16T15:12:10Z

holoviews/core/traversal.py

-
-
-class dimensionless_cache(object):
-    """


Glad to get rid of this!

jlstevens · 2017-01-16T15:15:17Z

holoviews/core/spaces.py

@@ -500,6 +517,8 @@ class DynamicMap(HoloMap):
       """)

    def __init__(self, callback, initial_items=None, **params):
+        if not isinstance(callback, Callable) and not isinstance(callback, types.GeneratorType):
+            callback = Callable(callable_function=callback)


This means all callbacks will now be wrapped in Callable. This is a change from the previous behavior, and probably a good one (more consistent, at least with __mull__) though we should now document this. It would also be good to have an example of where a user might want to supply Callable themselves....

For instance, we might want an example of the user declaring a Callable with their own set of inputs.

jlstevens · 2017-01-16T15:18:12Z

holoviews/core/spaces.py

    def __call__(self, *args, **kwargs):
-        return self.callable_function(*args, **kwargs)
+        inputs = [inp for inp in self.inputs if isinstance(inp, DynamicMap)]


Very minor point - I would call the loop variable i given that inp is a bit weird and you probably don't want to clash with the input built-in. Alternatively, this is one instance of a pure filter so you could consider using that (it would be shorter).

jlstevens · 2017-01-16T15:22:03Z

holoviews/core/spaces.py

-        return self.callable_function(*args, **kwargs)
+        inputs = [inp for inp in self.inputs if isinstance(inp, DynamicMap)]
+        streams = [s for inp in inputs for s in get_nested_streams(inp)]
+        values = tuple(tuple(sorted(s.contents.items())) for s in streams)


This means it is memoizing on the stream parameters in ascending alphanumeric order. I think that is fine (and makes sense!) but it is worth noting that this is one particular policy for how streams parameters are ordered into a tuple key. This is important to stay consistent if we ever need it somewhere else and I am now wondering if it might be worth having a utility to codify the idea....up to you.

It really doesn't matter much, it just needs to be consistently ordered. If we're using it somewhere else it can use the same scheme or another scheme without having any effect here.

True - my point is that we should try to be consistent anyway!

jlstevens · 2017-01-16T15:31:35Z

Looks good!

I'm very happy this turned out to be far more straightforward than I anticipated. Overall only seven lines were added and the new approach is far more intuitive to me than dimensionless_cache ever was.

I think my only worry is that is that (from what I understand) Callback will memoise at every level so dmap1 * dmap2 * dmap3 is memoising 3 + 2 + 1 i.e 6 times (3 leaves, two binary muls, one triple mul). The main worry then is unnecessary/excessive memory usage.

philippjfr · 2017-01-16T15:35:05Z

I think my only worry is that is that, from what I understand, Callback will memoise at every level so dmap1 * dmap2 * dmap3 is memoising 3 + 2 + 1 i.e 6 times (3 leaves, two binary muls, one triple mul). The main worry then is unnecessary/excessive memory usage.

That's accurate although each level already caches these values in the DynamicMap cache so I don't think there will be much being cached additionally. It's also worth noting that each level simply wraps the previous level, so it's not holding copies of the data, so the overhead should simply be in the parameterized objects that are created, which should be fairly cheap.

jlstevens · 2017-01-16T16:36:57Z

Tests are passing so I will merge now.

philippjfr added 2 commits January 15, 2017 21:09

Dynamic Callable memoizes last returned value

c771640

Removed dimensionless_cache

601d1f8

philippjfr added the status: WIP label Jan 15, 2017

Added docstring and unit test about Callable memoization

f0a66c7

philippjfr removed the status: WIP label Jan 16, 2017

philippjfr requested a review from jlstevens January 16, 2017 13:53

jlstevens reviewed Jan 16, 2017

View reviewed changes

holoviews/core/traversal.py

class dimensionless_cache(object):

"""

Copy link

Contributor

jlstevens Jan 16, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Glad to get rid of this!

jlstevens reviewed Jan 16, 2017

View reviewed changes

Cleaned up dynamic Callable

8a701c8

philippjfr force-pushed the callable_cache branch from e40e0fc to 8a701c8 Compare January 16, 2017 16:01

jlstevens merged commit 4deedc8 into master Jan 16, 2017

philippjfr mentioned this pull request Jan 17, 2017

Dynamic Callback memoization breaks for non-hashable stream values #1067

Closed

philippjfr deleted the callable_cache branch January 27, 2017 02:53

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add memoization to dynamic Callable class #1063

Add memoization to dynamic Callable class #1063

philippjfr commented Jan 15, 2017 •

edited

jbednar commented Jan 16, 2017

philippjfr commented Jan 16, 2017 •

edited

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens Jan 16, 2017 •

edited

philippjfr Jan 16, 2017

jlstevens Jan 16, 2017

jlstevens commented Jan 16, 2017 •

edited

philippjfr commented Jan 16, 2017

jlstevens commented Jan 16, 2017

Add memoization to dynamic Callable class #1063

Add memoization to dynamic Callable class #1063

Conversation

philippjfr commented Jan 15, 2017 • edited

jbednar commented Jan 16, 2017

philippjfr commented Jan 16, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jlstevens Jan 16, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jlstevens commented Jan 16, 2017 • edited

philippjfr commented Jan 16, 2017

jlstevens commented Jan 16, 2017

philippjfr commented Jan 15, 2017 •

edited

philippjfr commented Jan 16, 2017 •

edited

jlstevens Jan 16, 2017 •

edited

jlstevens commented Jan 16, 2017 •

edited