Reduce batch size & general questions

[Romeo1-1]

Hello, and thank you for this very interesting and exciting package !

I have 2 issues and 1 general question:

1. I am working on a Windows Laptop with WSL2. I tried running your Docker Image but was always running through this issue:

docker: Error response from daemon: failed to create shim task: OCI runtime create failed: runc create failed: unable to start container process: error during container init: error running hook #0: error running hook: exit status 1, stdout: , stderr: Auto-detected mode as 'legacy'
nvidia-container-cli: mount error: file creation failed: /var/lib/docker/overlay2/775f320175cbf2f849eb43680d71fb362bd58a8c7b33ea54ab99002f75bc476a/merged/usr/lib/x86_64-linux-gnu/libnvidia-ml.so.1: file exists: unknown

This is an issue that does not seem related only to your package. As mentionned here, NVIDIA/nvidia-docker#1699 (comment) I rebuilt a new docker image with this Dockerfile and it now works

FROM deepair1/deepair:latest

RUN rm -rf \
    /usr/lib/x86_64-linux-gnu/libcuda.so* \
    /usr/lib/x86_64-linux-gnu/libnvcuvid.so* \
    /usr/lib/x86_64-linux-gnu/libnvidia-*.so* \
    /usr/lib/firmware \
    /usr/local/cuda/compat/lib

2. When I am trying to run your exemples, I always immediatly run out of memory. This is the end of the error message I receive.

tensorflow.python.framework.errors_impl.ResourceExhaustedError: Exception encountered when calling layer "output" (type TFBertOutput).

failed to allocate memory [Op:AddV2]

Call arguments received:
  • hidden_states=tf.Tensor(shape=(3, 5, 4096), dtype=float32)
  • input_tensor=tf.Tensor(shape=(3, 5, 1024), dtype=float32)
  • training=False

I am running your exemple with a Laptop, with a Nvidia GTX 1050Ti GPU, an i5 7th Gen with 4 cores (2.5Ghz) and 16Gb of RAM. After trying some workarounds (tensorflow/tensorflow#51354), I think the issue come from the batch_size that is set to 3 (if i understand correctly), but I couldn't find any way to simply lower it. Where could I modify it ?

3. I am very interested in trying your packages with TCR that were sequenced with VDJ 10x genomic solution. I read thoroughly your pre-print but I am very new to deep learning. If I wanted to run specific sequences from my personal TCR runs against selected epitopes, I could directly use your model, right ? And as i understood it, for the runs I would just need the TCR chains sequences, but I would also need the sequences and AlphaFold2 results of my epitopes of interest ?

Thank you very much in advance for your help