| |
| |
| |
| import io, os, sys |
| import numpy as np, pyarrow.parquet as pq, soundfile as sf, torch |
| from transformers import Wav2Vec2Config, Wav2Vec2Model |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, HERE) |
| import _net |
| W = __import__([m for m in ("prosdd_asv19", "prosdd_asv5") if os.path.exists(os.path.join(HERE, m + ".py"))][0]) |
| src = sys.argv[1] |
| ckpt = sys.argv[2] if len(sys.argv) > 2 else W._CKPT |
| sys.path.append(src) |
| Wav2Vec2Model.from_pretrained = classmethod(lambda cls, *a, **k: cls(Wav2Vec2Config.from_json_file(_net.XLSR_CONFIG))) |
| import model_stage2realfake, main_eval, data_utils_eval |
|
|
| torch.backends.cudnn.allow_tf32 = False |
| torch.backends.cuda.matmul.allow_tf32 = False |
| dev = "cuda" |
| ref = model_stage2realfake.ProSDDStage2(classifier_pool="mean", T_target=200) |
| state = torch.load(ckpt, map_location="cpu", weights_only=True) |
| r = ref.load_state_dict({k.replace("module.", ""): v for k, v in state.items()}, strict=False) |
| print("source strict=False: missing", r.missing_keys, "unexpected", r.unexpected_keys) |
| ref = ref.to(dev).eval() |
| m = W.__dict__[[n for n in dir(W) if n.startswith("ProSDDASV")][0]]() |
| m.load() |
|
|
| t = pq.read_table(os.path.join(HERE, "..", "J-SPAW_LA", "data", sorted(os.listdir(os.path.join(HERE, "..", "J-SPAW_LA", "data")))[-1]), |
| columns=["audio"]).slice(0, 16).column("audio").to_pylist() |
| auds = [sf.read(io.BytesIO(x["bytes"]), dtype="float32")[0] for x in t] |
| auds.append(auds[0][:20000]); auds.append(np.concatenate([auds[1]] * 4)) |
| ours = np.array(m.score_batch(auds, [16000] * len(auds))) |
| theirs = [] |
| with torch.no_grad(): |
| for a in auds: |
| w = data_utils_eval.pad(torch.from_numpy(a).float())[None].to(dev) |
| theirs.append(float(main_eval.inference_forward(ref, w)[:, 1])) |
| theirs = np.array(theirs) |
| print("lens", [len(a) for a in auds]) |
| print("ours ", np.round(ours, 4)) |
| print("source", np.round(theirs, 4)) |
| print("max_abs_diff", float(np.max(np.abs(ours - theirs)))) |
|
|
