Running selene wih the original deepSEA model

[Marcel Mück]

Hello everybody,
I was going through the instructions in `tutorials/variants_and_visualizations/selene_variants_and_vis_tutorial/variants_and_visualizations.ipynb`. In there, variant effects are predicted from a vcf file using the example `DeeperDeepSEA` model. I was now trying to get the same effect prediction to run on the model from the original deepSEA paper (Jian Zhou, Olga G. Troyanskaya 2015). For this I used the model architecture from 'https://github.com/FunctionLab/selene/blob/master/models/deepsea.py' as well as the model weights from the original paper (https://zenodo.org/records/1466993/files/deepsea_predict.pth). However, the model weights seemed to be incompatible. When I run

```

model_predict = AnalyzeSequences(
    NonStrandSpecific(DeepSEA(1000,919)),
    "deepsea_predict.pth",
    sequence_length=1000,
    features=distinct_features,
    reference_sequence=Genome("hg38.fa"),
    use_cuda=False  # update this to False if you do not have CUDA on your machine.
)

```

I get

```

RuntimeError: 

Error(s) in loading state_dict for NonStrandSpecific: 

size mismatch for model.conv_net.0.weight: copying a param with shape torch.Size([320, 4, 1, 8]) from checkpoint, the shape in current model is torch.Size([320, 4, 8]). 

size mismatch for model.conv_net.4.weight: copying a param with shape torch.Size([480, 320, 1, 8]) from checkpoint, the shape in current model is torch.Size([480, 320, 8]). 

size mismatch for model.conv_net.8.weight: copying a param with shape torch.Size([960, 480, 1, 8]) from checkpoint, the shape in current model is torch.Size([960, 480, 8]).

 size mismatch for model.classifier.0.weight: copying a param with shape torch.Size([925, 50880]) from checkpoint, the shape in current model is torch.Size([919, 50880]). 

size mismatch for model.classifier.0.bias: copying a param with shape torch.Size([925]) from checkpoint, the shape in current model is torch.Size([919]). 

size mismatch for model.classifier.2.weight: copying a param with shape torch.Size([919, 925]) from checkpoint, the shape in current model is torch.Size([919, 919]).

```

So the dimensions seem to be off. Did anyone have any idea how I could get selene to run on the original deepSEA model or did anyone run into similar problems?

Thank you and regards
Marcel Mück from DKFZ (German Cancer research Centre)