| # Float-source parity vs the published ITW score file: run from this dir: | |
| # python source_parity_float.py <module> <ssl_name> <N> (reads the original float wavs, not the PCM_16 parquet) | |
| import os, sys, numpy as np, soundfile as sf, torch, importlib | |
| sys.path.insert(0, os.getcwd()) | |
| W = importlib.import_module(sys.argv[1]); ssl = sys.argv[2]; N = int(sys.argv[3]) | |
| SRC = "/home/kirill/mnt/users_4tb/datasets/release_in_the_wild" | |
| pub = {} | |
| for line in open(f"../_ckpt_dl/issf__spoof-superb-scores/raw/linear_head/in_the_wild/{ssl}.txt"): | |
| p = line.split(); pub[int(p[0])] = float(p[-1]) | |
| ids = list(range(N)) | |
| aud = [sf.read(f"{SRC}/{i}.wav", dtype="float32")[0] for i in ids] | |
| m = [c for n, c in vars(W).items() if n.startswith("SpoofSUPERB") and isinstance(c, type)][0](); m.load() | |
| pv = np.array([pub[i] for i in ids]) | |
| for tf in (False, True): | |
| torch.backends.cudnn.allow_tf32 = tf | |
| s = np.array(sum((m.score_batch(aud[k:k+16], [16000]*16) for k in range(0, N, 16)), [])) | |
| print(f"float source, cudnn_tf32={tf}: max|d|={np.abs(pv-s).max():.3g} mean|d|={np.abs(pv-s).mean():.3g}") | |