In many applications, elements from a large dataset, that cannot be stored in memory, need to be sampled (with or without replacement). This simple package simplifies this process by allowing storage of large datasets by splitting them over multiple files in a directory. Once that's done, any dataset element can be queried. Further, most python objects as dataset elements can be used as pickles are used at the backend.

As mentioned before, a dataset will be represented by a directory that contains two sets of files, meta and data. The file identifiers are randomly generated and therefore, two datasets can be added together to create a larger dataset without additional work. This also allows multiple applications to add elements to the dataset in parallel. An example is multiple simulations running in parallel.

pip3 install ra-pickles

First import all the packages

from ra_pickles import RandomAccessPicklesReader, RandomAccessPicklesWriter
import shutil
import os
import numpy as np

So to create a dataset in test_folder, use the following code. 100 here specifies number of dataset samples in each file. If there are more than 100 files, multiple files will be created. Heuristically, this number can be chosen such that resulting data files are approximately 1 GB, although this wouldn't affect performance much as a complete file is not fully loaded into memory. On the other hand, a too small number can also affect performance as some file systems create issues if there are too many files.

fold = 'test_folder'
writer = RandomAccessPicklesWriter(100, fold)

Now entries to the dataset can be added as follows:

for i in range(30):
    # d is the dataset element which can be any python object
    d = np.ones(4) * i
    
    # and you add it to the dataset
    writer.add(d)
writer.close()

Here, 30 numpy arrays have been added.

To sample elements in the dataset, use the following code:

reader = RandomAccessPicklesReader(fold)
print("Total", reader.get_total())

print("Trying to retreive")
for i in range(30):
    print("Retrieving",i)
    a = reader.get_element(i)
    print(i, a[0])

The index retrieval index can be randomly generated --[0, reader.get_total() ) to sample random elements.

Multiple elements can also be retrieved in parallel for fast access. To do so, first, retrieval threads must be started (and closed at the end):

reader.start_parallel_retrieval_threads(n_theads=5)
data = reader.get_multi_in_parallel([1,2,3])
reader.close_parallel_retrieval_threads()
reader.close()

Good luck!