upload LiSenNet checkpoint + ONNX (fp32, int8 dyn/static)
Browse files
README.md
CHANGED
|
@@ -51,8 +51,7 @@ reproduction lands within ~0.06 of the paper's reported ~3.07.
|
|
| 51 |
| ------------------------- | --- |
|
| 52 |
| `g_best` | PyTorch checkpoint (`{"generator": state_dict}`) |
|
| 53 |
| `g_best_fp32.onnx` | FP32 ONNX of the mask sub-network: `feat (B,3,T,F)` -> `est_mag (B,T,F)`, dynamic batch/time |
|
| 54 |
-
| `
|
| 55 |
-
| `g_best_int8_static.onnx` | Static full int8 (QDQ, per-tensor, percentile calibration; QAT-grade) |
|
| 56 |
| `config.json` | Training + architecture config (STFT, channels, blocks) |
|
| 57 |
|
| 58 |
The ONNX graphs cover only the mask sub-network (the pure tensor-op core). The
|
|
@@ -89,7 +88,7 @@ from huggingface_hub import hf_hub_download
|
|
| 89 |
|
| 90 |
REPO = "claroche1/LiSenNet"
|
| 91 |
sess = ort.InferenceSession(
|
| 92 |
-
hf_hub_download(REPO, "
|
| 93 |
providers=["CPUExecutionProvider"],
|
| 94 |
)
|
| 95 |
# input feat: (B, 3, T, F=257) = [compressed_mag, group_delay/pi, ifd/pi]
|
|
|
|
| 51 |
| ------------------------- | --- |
|
| 52 |
| `g_best` | PyTorch checkpoint (`{"generator": state_dict}`) |
|
| 53 |
| `g_best_fp32.onnx` | FP32 ONNX of the mask sub-network: `feat (B,3,T,F)` -> `est_mag (B,T,F)`, dynamic batch/time |
|
| 54 |
+
| `g_best_int8_static.onnx` | Static full int8 (QDQ, per-tensor, percentile calibration; embedded-deployable) |
|
|
|
|
| 55 |
| `config.json` | Training + architecture config (STFT, channels, blocks) |
|
| 56 |
|
| 57 |
The ONNX graphs cover only the mask sub-network (the pure tensor-op core). The
|
|
|
|
| 88 |
|
| 89 |
REPO = "claroche1/LiSenNet"
|
| 90 |
sess = ort.InferenceSession(
|
| 91 |
+
hf_hub_download(REPO, "g_best_int8_static.onnx"), # or g_best_fp32.onnx
|
| 92 |
providers=["CPUExecutionProvider"],
|
| 93 |
)
|
| 94 |
# input feat: (B, 3, T, F=257) = [compressed_mag, group_delay/pi, ifd/pi]
|