Ears for coding agents: let AI hear, review and make sound. CPU-only, open source.
"""bf16 vs fp32 CLAP on this CPU: embedding agreement and speed (models load twice; no cache is used or written).
EARSHOT_THREADS=2 .venv/Scripts/python.exe eval/semantics/precision.py # -> eval/semantics/precision.json
Input: the first 10 s of 12 real recordings (2 per category folder of the Clawd project's library, read-only), each
embedded in fp32 and in bf16; cosine between the two embeddings of the same clip, and seconds per 10 s window
(batch 4, after one warm-up batch). Whether bf16 is used at all depends on the CPU: earshot.models.clap.native_bf16.
"""
import json
import os
import sys
import time
import numpy as np
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.dirname(os.path.dirname(HERE)))
from earshot import audio, config # noqa: E402
from earshot.models import clap # noqa: E402
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) # eval/: local data locations
import evalpaths as EP # noqa: E402
ASSETS = EP.ASSETS
CATS = ['ambience', 'foley', 'keyboard', 'motion', 'music', 'toy']
def clips():
out = []
for c in CATS:
fs = sorted(f for f in os.listdir(os.path.join(ASSETS, c)) if os.path.splitext(f)[1].lower() in audio.AUDIO_EXTS)
for f in fs[:2]:
out.append(audio.load(os.path.join(ASSETS, c, f), mono=True, end=10.0))
return out
def run(precision, xs):
config.PRECISION = precision
c = clap.Clap()
c._forward(xs[:4], 'repeatpad', 4) # warm-up
t = time.time()
e = c._forward(xs, 'repeatpad', 4)
return c.precision, e, (time.time() - t) / len(xs)
def main():
import torch
xs = clips()
p32, e32, s32 = run('fp32', xs)
p16, e16, s16 = run('bf16', xs)
cos = np.sum(e32 * e16, axis=1) / (np.linalg.norm(e32, axis=1) * np.linalg.norm(e16, axis=1))
res = {'torch': torch.__version__, 'cpu_capability': torch.backends.cpu.get_cpu_capability(),
'native_bf16': clap.native_bf16(torch), 'threads': config.threads(), 'clips': len(xs),
'used': [p32, p16], 'cosine_mean': round(float(cos.mean()), 5), 'cosine_min': round(float(cos.min()), 5),
's_per_window_fp32': round(s32, 3), 's_per_window_bf16': round(s16, 3),
'speedup_bf16': round(s32 / s16, 2) if s16 else None}
json.dump(res, open(os.path.join(HERE, 'precision.json'), 'w'), indent=1)
print(res)
if __name__ == '__main__':
main()