Chris Seymour

@aqrit I think I am missing something. I can see the delta coding but I'm confused on the signed integer support as the functions still expect `uint16_t`. Is the idea...

> Using FSE directly may be better? I spent some time looking at `fse` and it seems `streamvbyte + zstd` is hard to beat *(encoding)*. | | Time (s) |...

Hello @hd2326 you can give bonito your custom trained remora onnx model with `bonito basecaller --modified-base-model custom.onnx` _(it doesn't need converting)_.

You need both models _(a bonito basecalling model [tar+toml] and a remora modbase [onnx] model)_ but it's one command - ``` bonito basecaller [email protected] /data/reads --modified-base-model custom.onnx > calls.bam ```

Hey @andreaswallberg This detailed analysis was really helpful - the issues you highlighted should be resolved with the `[email protected]` model that was just release in `v0.3.7`.

Hi @AlbertoOsorio Can you confirm you are passing a directory that contains one or more `*.fast5` files inside? By default, bonito will only call the files inside the directory specified....

Hey @linzho Thanks for raising - we have seen some strandedness internally and are investigating, I will update this issue when we have more details.

PyTorch only provides prebuilt versions against CUDA 10.2 and 11.1 right now, so that's the main driver here - you can install CUDA 11.1 alongside 11.4 however.

CUDA 11.3 builds are now available `ont-bonito-cuda113`.

Hey @visanuwan You can specify the location of the training data with `--directory` ``` $ bonito train --amp /data/training/model-dir --directory /data/training/training-data-dir ```