Download test_rawnetlite.py from SpeechAntiSpoofingBenchmarks/RawNetLite: direct link, hf CLI and curl.
- Browser
- Download file 2.73 kB
-
https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawNetLite/resolve/main/test_rawnetlite.py
- Command line
-
hf download hf://SpeechAntiSpoofingBenchmarks/RawNetLite/test_rawnetlite.py
-
curl -L -o test_rawnetlite.py https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawNetLite/resolve/main/test_rawnetlite.py
2.73 kB
| import hashlib | |
| import io | |
| import os | |
| import numpy as np | |
| import pytest | |
| import torch | |
| import _net | |
| import rawnetlite as R | |
| SHA256 = "6154d91f34f0a49d58e30bac602915f6b6853419419804be63b5274603891a9b" | |
| N_PARAMS = 240_769 # trainable parameters (BN running stats / num_batches_tracked are buffers) | |
| def test_prep_peak_norm_then_prefix(): | |
| x = np.concatenate([np.full(48000, 0.1, np.float32), np.array([-0.5], np.float32), np.zeros(999, np.float32)]) | |
| out = _net.prep_np(x) # peak over the WHOLE clip (0.5 lies beyond the 3 s crop) | |
| assert out.shape == (48000,) and out.dtype == np.float32 and np.allclose(out, 0.2) | |
| def test_prep_short_right_zero_pad(): | |
| x = np.linspace(-0.25, 0.5, 1001, dtype=np.float32) | |
| out = _net.prep_np(x) | |
| assert np.allclose(out[:1001], x / 0.5) and not out[1001:].any() and abs(out.max() - 1.0) < 1e-7 | |
| assert not _net.prep_np(np.zeros(5000, np.float32)).any() # silent clip -> zeros (source: NaN) | |
| assert not _net.prep_np(np.zeros(0, np.float32)).any() | |
| def test_checkpoint_sha256(): | |
| assert hashlib.sha256(open(R._CKPT, "rb").read()).hexdigest() == SHA256 | |
| def test_strict_load_all_tensors_used(): | |
| sd = R.load_state() | |
| assert len(sd) == 61 | |
| net = _net.RawNetLite() | |
| assert set(net.state_dict()) == set(sd) | |
| r = net.load_state_dict(sd, strict=True) | |
| assert not r.missing_keys and not r.unexpected_keys | |
| def model(): | |
| m = R.RawNetLite(); m.load() | |
| yield m | |
| m.unload() | |
| def test_eval_and_params(model): | |
| assert not any(mod.training for mod in model.net.modules()) | |
| assert sum(p.numel() for p in model.net.parameters()) == N_PARAMS | |
| def test_logit_matches_source_sigmoid(model): | |
| x = torch.randn(3, 1, 48000, device=model.device) * 0.3 | |
| with torch.no_grad(): | |
| assert torch.allclose(torch.sigmoid(model.net.forward_logit(x)), model.net(x), atol=1e-6) | |
| def test_shapes_and_batch_matches_single(model): | |
| rng = np.random.default_rng(0) | |
| a = [rng.standard_normal(n).astype(np.float32) * s for n, s in ((48000, 0.1), (20000, 0.5), (90000, 0.02), (3000, 0.3))] | |
| batched = model.score_batch(a, [16000] * 4) | |
| single = [model.score_batch([x], [16000])[0] for x in a] | |
| assert len(batched) == 4 and all(np.isfinite(batched)) | |
| assert np.allclose(batched, single, atol=1e-4) | |
| def test_deterministic(model): | |
| a = [np.random.randn(40000).astype(np.float32) * 0.1 for _ in range(3)] | |
| assert model.score_batch(a, [16000] * 3) == model.score_batch(a, [16000] * 3) | |
| def test_scale_invariant(model): | |
| x = np.random.default_rng(2).standard_normal(30000).astype(np.float32) * 0.1 | |
| s1, s2 = model.score_batch([x, 3 * x], [16000] * 2) | |
| assert abs(s1 - s2) < 1e-4 # peak normalisation | |