setting up inference to use pre trained weights

[iamrishu11]

So i was going through the github and hugging face link and available items and then tried to create the inference script to test the model.

Steps:

1. I firstly cloned the repo and its content !git clone https://github.com/pulp-bio/BioFoundation.git
2. installed the required packages from requirements.txt
3. downloaded the weights from the hugging face via the following command

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="thorir/FEMBA",
    repo_type="model",
    local_dir="checkpoints/FEMBA"
)

4. installed mamba-ssm and its dependencies
5. wrote the script to test the artifact detection model

import torch
from safetensors.torch import load_file
from models.FEMBA import FEMBA

# Model configuration
model = FEMBA(
    seq_length=1280,
    num_channels=22,
    num_classes=2,
    exp=4,
    embed_dim=79,
    num_blocks=4,
    classification_type='bc'
)

# Load safetensors checkpoint
state_dict = load_file("/content/checkpoints/FEMBA/TUAR/FEMBA_base.safetensors")
model.load_state_dict(state_dict)

model.eval()

# Dummy input
x = torch.randn(1, 22, 1280)
mask = torch.zeros_like(x, dtype=torch.bool)

# Inference
with torch.no_grad():
    preds, _ = model(x, mask)

print(preds)

but im getting this error module 'torch' has no attribute 'version' multiple times, tried restarting the colab session.
even though the torch is installed and working

import torch
print(torch.__version__)

result is 2.8.0+cu126

PS : this is the colab notebook Link.

[iamrishu11]

am i doing something wrong in testing it?, bcs i only found the doc for retraining the model or fine tuning it. Would really appreciate the help.

[Thoriri]

Hi,

The error you're getting is because you call

from models.FEMBA import FEMBA

From a notebook outside BioFoundation.

In the collab you're using you can do this instead:

from BioFoundation.models.FEMBA import FEMBA

There are some other issues with the code. For FEMBA-Base, we have an embedding dimension of 35 and 12 blocks; this should be defined.

Additionally, since the model you are loading is the one that has been pre-trained on TUEG without TUAR, you need to fine-tune it first. To do this, I highly recommend starting with the actual BioFoundation as a core repository. This means that you would define your checkpoint 'pretrained_checkpoint_path' (check config/experiment/FEMBA_finetune.yaml).

[iamrishu11]

i fixed these issue and now to pretrain the model i am preparing the dataset , i have the tuh_eeg_artifact/v3.0.1/edf/03_tcp_ar_a/
folder it have around 7 edf files

['aaaaabnn_s002_t004_seiz.csv',
 'aaaaatjz_s001_t000_seiz.csv',
 'aaaaabnn_s002_t001_seiz.csv',
 'aaaaabnn_s002_t002_seiz.csv',
 'aaaaabnn_s002_t001.csv',
 'aaaaadao_s001_t003.csv',
 'aaaaagxr_s017_t000.csv',
 'aaaaadao_s001_t004.csv',
 'aaaaatjz_s001_t000.csv',
 'aaaaabnn_s002_t004.csv',
 'aaaaabnn_s002_t002.csv',
 'aaaaabnn_s002_t001.edf',
 'aaaaabnn_s002_t004.edf',
 'aaaaadao_s001_t004.edf',
 'aaaaagxr_s017_t000.edf',
 'aaaaadao_s001_t003.edf',
 'aaaaabnn_s002_t002.edf',
 'aaaaatjz_s001_t000.edf']

for testing purpose im only using the small dataset now i will have to convert them into .pkl first by running make_datasets/process_raw_eeg.py and then convert those .pkl to .hdf5 format?

am i correct?

[iamrishu11]

i ran this

!python /content/BioFoundation/make_datasets/process_raw_eeg.py tuab \
    --root_dir /content/DATA/TUH_structured \
    --output_dir /content/DATA/TUH_processed

and it gave me this output

Setting up TUAB dataset processing...
Found 7 files to process using 24 processes.
  0% 0/7 [00:00<?, ?it/s]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
 14% 1/7 [00:15<01:34, 15.80s/it]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
 43% 3/7 [00:16<00:17,  4.40s/it]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
 57% 4/7 [00:17<00:09,  3.03s/it]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
 71% 5/7 [00:17<00:04,  2.10s/it]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
 86% 6/7 [00:17<00:01,  1.56s/it]NOTE: pick_channels() is a legacy function. New code should use inst.pick(...).
100% 7/7 [00:19<00:00,  2.86s/it]
Processing complete.

but i can't find any .pkl files inside the TUH_processed dir

this is my notebook , if u need here

[Thoriri]

I highly recommend following the README at https://github.com/pulp-bio/BioFoundation/tree/main/make_datasets line by line.

I do not fully get what the issue is. You ran process_raw_eeg.py, calling TUAB but supplying it TUAR data (why)?

From the output, it looks like the processing went fine. Do you not see any files, error log, or anything anywhere? (Next step would be to convert to HDF5, as detailed in the README.)

I ran this for example here on my side:
python make_datasets/process_raw_eeg.py tuar --root_dir /EEG/TUAR/edf --outp ut_dir TUAR_processed/TUAR_data Setting up TUAR dataset processing... Found 303 files to process using 24 processes. 0%| | 0/303 [00:00<?, ?it/s]
And this processes the data and put the .pkl files inside TUAR_processed/TUAR_data/processed.

[Thoriri]

Please take a look at the new 'make_datasets' folder and the updated documentation! I hope it's clear -- I'll close this issue for now. If it's still a problem, just reopen it.