File size: 1,959 Bytes
1bd2ac7 | 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 64 65 | ---
library_name: pytorch
tags:
- eeg
- braindecode
- pytorch
- keras-conversion
- eegsym
license: other
---
# EEGSym original Keras 8-channel weights converted to PyTorch
This repository contains a PyTorch `state_dict` converted from the original
Keras `EEGSym_pretrained_weights_8_electrode.h5` checkpoint for the 8-electrode
EEGSym configuration.
The checkpoint targets the faithful PyTorch port in:
`replicability.eegsym_faithful.FaithfulEEGSym`
It is not a drop-in checkpoint for the current upstream `braindecode.models.EEGSym`
topology. The upstream model is structurally close, but it differs from the
authors' released Keras architecture in enough places that the original weights
cannot be loaded 1:1 without an architecture-compatibility path.
## Expected Input
- Shape: `(batch, 8, 384)`
- Sampling rate: 128 Hz
- Canonical channel order: `F3, C3, P3, Cz, Pz, F4, C4, P4`
- Output: 2 logits
## Verification
- Keras H5 SHA256: `f60ddd220cf48dd18dd9706c5389c70a244fe878610021c2ed97d40105145b2b`
- State values excluding `num_batches_tracked`: `144440`
- Trainable parameters: `142784`
- H5 datasets consumed: `146/146`
- PyTorch tensors filled: `146/146`
- Forward probe output shape: `(2, 2)`
- Source replicability commit: `a76ee9bf66204b4ebd07746aeb48933dd4272daa`
## Loading
```python
import torch
from replicability.eegsym_faithful import FaithfulEEGSym
model = FaithfulEEGSym(n_chans=8, n_outputs=2, n_times=384)
state_dict = torch.load("pytorch_model.bin", map_location="cpu")
model.load_state_dict(state_dict)
model.eval()
```
## Provenance
The original Keras checkpoint was obtained from the authors' EEGSym reference
code mirrored under the replicability evidence tree:
`research/evidence/source_code_repos/EEGSym/EEGSym/EEGSym_pretrained_weights_8_electrode.h5`
This conversion only changes tensor layout from Keras/HDF5 conventions to the
faithful PyTorch module layout; it does not retrain or modify the learned values.
|