Compare performance of IterativeDecoder and ChunkedDecoder

Both IterativeDecoder and ChunkedDecoder work well with encoded sequences. Additionally, an IterativeDecoder can replace the functionality of a ChunkedDecoder as the iterator it returns can easily be converted to a list. Because of this we will compare the efficiency of IterativeDecoder and ChunkedDecoder to see if we should replace the functionality of ChunkedDecoder with IterativeDecoder.

from hum.gen.sine_mix import freq_based_stationary_wf
from recode import (ChunkedEncoder, 
                    ChunkedDecoder, 
                    IterativeDecoder, 
                    StructCodecSpecs)

Create a synthetic waveform to test both decoders on:

DFLT_N_SAMPLES = 21 * 2048
DFLT_SR = 44100

wf_mix = freq_based_stationary_wf(freqs=(200, 400, 600, 800), weights=None,
                             n_samples = DFLT_N_SAMPLES*300, sr = DFLT_SR)

Define the decoders and encode the waveform:

specs = StructCodecSpecs('d')
encoder = ChunkedEncoder(frame_to_chk=specs.frame_to_chk)
decoder = ChunkedDecoder(
    chk_size_bytes=specs.chk_size_bytes,
    chk_to_frame=specs.chk_to_frame,
    n_channels=specs.n_channels
)
idecoder = IterativeDecoder(chk_to_frame = specs.chk_to_frame_iter)
b = encoder(wf_mix)

Helper function so the output of the decoders are in the same format:

def it_to_list(it):
    frame = list(it)
    frame = [item for tup in frame for item in tup]
    return frame

Performance of ChunkedDecoder:

%timeit -r 2 -n 5 list(decoder(b))

5.59 s ± 14.5 ms per loop (mean ± std. dev. of 2 runs, 5 loops each)

Performance of IterativeDecoder:

%timeit -r 2 -n 5 it_to_list(idecoder(b))

2.04 s ± 13.4 ms per loop (mean ± std. dev. of 2 runs, 5 loops each)

From these results, we can clearly see that we should replace the functionality of ChunkedDecoder with IterativeDecoder as it is nearly three times more efficient.