Upload metadata.json with huggingface_hub
Browse files- metadata.json +62 -0
metadata.json
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": {
|
| 3 |
+
"name": "FormantNet",
|
| 4 |
+
"source": "expdir/mvt33_f6z1sTpF10/weights.sigmoid195-20.143 (TF checkpoint)",
|
| 5 |
+
"architecture": "LSTM(257,512,return_sequences=True) -> Dense(20,sigmoid)",
|
| 6 |
+
"input": {
|
| 7 |
+
"name": "input",
|
| 8 |
+
"shape": ["batch", "time", 257],
|
| 9 |
+
"dtype": "float32",
|
| 10 |
+
"note": "normalized log-spectral envelope, 257 bins (32ms/16kHz window)"
|
| 11 |
+
},
|
| 12 |
+
"output": {
|
| 13 |
+
"name": "dense",
|
| 14 |
+
"shape": ["batch", "time", 20],
|
| 15 |
+
"dtype": "float32",
|
| 16 |
+
"note": "raw sigmoid [0,1]; rescale with FN_model.get_rescale_fn() to get Hz/BW/dB"
|
| 17 |
+
},
|
| 18 |
+
"opset": 15,
|
| 19 |
+
"total_params": 1587220,
|
| 20 |
+
"framework_origin": "tensorflow-macos 2.13 / keras 2.13"
|
| 21 |
+
},
|
| 22 |
+
"variants": {
|
| 23 |
+
"fp32": {
|
| 24 |
+
"file": "formantnet.onnx",
|
| 25 |
+
"size_mb": 6.361,
|
| 26 |
+
"max_abs_diff": 0.0,
|
| 27 |
+
"mean_abs_diff": 0.0,
|
| 28 |
+
"max_rel_diff": 0.0,
|
| 29 |
+
"threshold_abs": 0.0001,
|
| 30 |
+
"pass": true
|
| 31 |
+
},
|
| 32 |
+
"fp16": {
|
| 33 |
+
"file": "formantnet_fp16.onnx",
|
| 34 |
+
"size_mb": 3.184,
|
| 35 |
+
"max_abs_diff": 4.09e-04,
|
| 36 |
+
"mean_abs_diff": 2.49e-05,
|
| 37 |
+
"max_rel_diff": 4.52e-03,
|
| 38 |
+
"threshold_abs": 0.005,
|
| 39 |
+
"pass": true
|
| 40 |
+
},
|
| 41 |
+
"int8": {
|
| 42 |
+
"file": "formantnet_int8.onnx",
|
| 43 |
+
"size_mb": 1.608,
|
| 44 |
+
"max_abs_diff": 9.15e-02,
|
| 45 |
+
"mean_abs_diff": 8.71e-03,
|
| 46 |
+
"max_rel_diff": 0.737,
|
| 47 |
+
"threshold_abs": 0.15,
|
| 48 |
+
"pass": true,
|
| 49 |
+
"note": "max_rel_diff high due to near-zero sigmoid outputs; abs_diff within threshold (same pattern as DeepFormants-onnx int8)"
|
| 50 |
+
}
|
| 51 |
+
},
|
| 52 |
+
"post_processing_outside_onnx": [
|
| 53 |
+
"FFT / spectral envelope extraction (FN_data.py)",
|
| 54 |
+
"Input normalization by training-set mean/std (Normfile)",
|
| 55 |
+
"Output rescaling to Hz/BW/dB (FN_model.get_rescale_fn)",
|
| 56 |
+
"Formant sorting by mean frequency",
|
| 57 |
+
"Binomial smoothing (BIN_SMOOTH_PASSES=10)"
|
| 58 |
+
],
|
| 59 |
+
"conversion_script": "convert_to_onnx.py",
|
| 60 |
+
"quantization_script": "quantize_onnx.py",
|
| 61 |
+
"validation_script": "validate_onnx.py"
|
| 62 |
+
}
|