Deepvariant --no-channel-insert-size Error


I am getting the following error when trying to use “–no-channel-insert-size” with deepvariant on version 4.0.0-1 using “pbrun deepvariant --in-bam in.bam --ref ref.fna --out-variants out.vcf --no-channel-insert-size”:

[PB Error 2022-Dec-03 22:50:11][src/] The dimension of model file doesn’t match with the confirguation. Model:100,221,7. Confirguation:100,221,6, exiting.
[PB Error 2022-Dec-03 22:50:12][./inc/common.h:108] NvInfer ERROR: CUDA initialization failure with error 4. Please check your CUDA installation: Installation Guide Linux :: CUDA Toolkit Documentation, exiting.

Considering the error (“the dimension of model file doesn’t match with the confirguation … 100,221,7 … 100,221,6”), it appears that the pre-trained weights for Deepvariant don’t match the model that is being constructed; that is, it is probably trying to use the weights with the channel-insert-size layer on a model that does not have the layer (or vice-versa). Can you please help with loading the correct weights/model?

Hey @orianl,

I’ve just replicated your results and submitted a bug to the engineering team to investigate. Please hang tight. Thank you for letting us know.

Thank you! Is there anything that can be done in the meantime? I’ve noticed that there is an option for manually selecting the model via “–pb-model-file”, could that be used to load the v1.3 Deepvariant model? If so, where can that model file be found?

Hi @orianl,

Baseline Google DeepVariant changed the defaulted channel size to 7 in v1.4 release (it was 6 before). As we updated DeepVariant version in Parabricks v4.0 to be compatible with v1.4 of the baseline, we now have an engine that only support channel size 7.

You can indeed use an engine from Parabricks v3.8. If you have access to this version you can download the engine from inside the container and use it as input for --pb-model-file