A python decorator for caching the output of a repeated slow computation to disk, and bypassing future computations based on a md5 hash.
This is most useful for scientific pipelines where you repeatedly process some input data into an intermediate data structure, then do stuff on the result. Since the processing is the same each time and the 'stuff' is what's interesting, we would like to bypass the processing after the first call.
Using a single decorator, this module stores the output of the processing step to a pickle file after the first time, and will load it each subsequent time.
Because data is stored to disk, it is persistent across different runs of the interpreter. This module hides away all the logic of checking if the file exists, if it contains the correct data, or if you should recompute the data in the cache (do so by changing the expected hash).
In contrast with other methods that hash the function's input, we check the checksum of the resulting file. This is useful in many cases where the input isn't actually what matters, for example if the input is a the filename of the data to parse.
from results_filecache import results_cache
# Define an expected md5sum for the resulting file.
expected_md5 = '89ae471d2783a5bdaca71cd91d2f3274'
cache_decorator = partial(results_cache.cached_call,
expected_hash=expected_md5)
# Wrap your processing code with the decorator. If a cache file exists
# with the correct hash, it will be loaded and returned immediately
# instead of processing. Otherwise, it will process the data and save
# the result to the cache.
@cache_decorator
def process(x):
print('Running a big task on: {}'.format(x))
# Do stuff and return it here.
return
data = process(x) # This will do the processing work and stores data.
data = process(x) # This time, the data will be loaded from file.
Obviously, update the expected_md5 once you actually know what the
result should be (run md5sum on the cache file).
Note that you can either call process() several times in the same
program, or call it repeatedly on the same input across different runs.
See example.py for two ways of wrapping your processing code.
To facilitate dependency tracking in ROS, this python module is provided as a ROS package. This is simply to integrate into ROS's build system; none of the code actually depends on ROS.