File size: 1,969 Bytes
856d3a4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | {
"model": {
"name": "FormantNet",
"source": "expdir/mvt33_f6z1sTpF10/weights.sigmoid195-20.143 (TF checkpoint)",
"architecture": "LSTM(257,512,return_sequences=True) -> Dense(20,sigmoid)",
"input": {
"name": "input",
"shape": ["batch", "time", 257],
"dtype": "float32",
"note": "normalized log-spectral envelope, 257 bins (32ms/16kHz window)"
},
"output": {
"name": "dense",
"shape": ["batch", "time", 20],
"dtype": "float32",
"note": "raw sigmoid [0,1]; rescale with FN_model.get_rescale_fn() to get Hz/BW/dB"
},
"opset": 15,
"total_params": 1587220,
"framework_origin": "tensorflow-macos 2.13 / keras 2.13"
},
"variants": {
"fp32": {
"file": "formantnet.onnx",
"size_mb": 6.361,
"max_abs_diff": 0.0,
"mean_abs_diff": 0.0,
"max_rel_diff": 0.0,
"threshold_abs": 0.0001,
"pass": true
},
"fp16": {
"file": "formantnet_fp16.onnx",
"size_mb": 3.184,
"max_abs_diff": 4.09e-04,
"mean_abs_diff": 2.49e-05,
"max_rel_diff": 4.52e-03,
"threshold_abs": 0.005,
"pass": true
},
"int8": {
"file": "formantnet_int8.onnx",
"size_mb": 1.608,
"max_abs_diff": 9.15e-02,
"mean_abs_diff": 8.71e-03,
"max_rel_diff": 0.737,
"threshold_abs": 0.15,
"pass": true,
"note": "max_rel_diff high due to near-zero sigmoid outputs; abs_diff within threshold (same pattern as DeepFormants-onnx int8)"
}
},
"post_processing_outside_onnx": [
"FFT / spectral envelope extraction (FN_data.py)",
"Input normalization by training-set mean/std (Normfile)",
"Output rescaling to Hz/BW/dB (FN_model.get_rescale_fn)",
"Formant sorting by mean frequency",
"Binomial smoothing (BIN_SMOOTH_PASSES=10)"
],
"conversion_script": "convert_to_onnx.py",
"quantization_script": "quantize_onnx.py",
"validation_script": "validate_onnx.py"
}
|