| # Conversion record |
|
|
| How `model.safetensors` was produced from Google's original TensorFlow artifact. |
| These scripts are kept as a record. They are not needed to use the model. |
|
|
| 1. `inspect_google_tensorflow.py`: lists the original SavedModel's graph and |
| variables (TensorFlow 2.x, CPU). |
| 2. `google_tensorflow_reference.py`: extracts the 96 trained variables, the |
| Hann window and the mel matrix exactly as the graph computes them, and runs |
| the original model on reference clips. The reference clips came from the |
| training project's private feature cache. That step cannot be reproduced |
| from this repository, but its outputs are summarised in `../qualification/`. |
| 3. `convert_google_weights.py`: maps the variables to the PyTorch module in |
| `naptrap.google` (HWIO→OIHW kernels, and the first kernel's inference-time |
| zero-mean subtraction applied ahead of time). It writes `model.pt` |
| (sha256 `20c27209…b246`). |
| 4. `to_safetensors.py`: repackages `model.pt` as `model.safetensors`. Every |
| tensor is checked to be bit-identical. |
|
|
| The `../qualification/` reports were measured against `model.pt`. Its tensors |
| are identical to `model.safetensors`. |
|
|
