johnire80 commented 12 months ago

Hello,

I am trying to run DNAscent on a cluster with Nvidia A100 GPUs. Is this possible?

DNAscent uses tensorflow 1.15x but, according to what I have read, the A100 can only use tensorflow 2x (the A100 needs cuda 11 & cudnn 8 which apparently cant use tensorflow 1.15x).

However, using tensorflow 2.4, cuda 11 & cudnn 8, I can get DNAscent running on a A100, but the results are strange, producing 2 floating point numbers rather than a float and sequence string (see below). Does this suggest that minimal changes to DNAscent could get it working with tensorflow 2.4?

Alignment guppy_output_all/pass/alignment.sorted.bam

Genome GCA_002163515.1_ASM216351v1_genomic.fa

Index index.dnascent

Threads 1

Compute GPU

Mode CNN

MappingQuality 20

MappingLength 200

SystemStartTime 20/11/2023 15:18:10

Software /dnascent/DNAscent/

Version 3.1.2

Commit d57082f35f7c3e9dab0ac61cc60510ee43ccac1e

2057e48f-732b-4228-9560-fea0709f321a CM007964.1 0 306 fwd 1f3a1973-3350-4b47-ba69-3e0e79f3a451 CM007964.1 0 303 fwd ed5096a0-d54f-4f5d-ab27-260e3444739c CM007964.1 0 305 fwd 31cc4b53-81b9-41a5-afac-29e2c03f2615 CM007964.1 0 306 fwd f7a1e6cf-cbdb-470a-9c4d-746a48b8aa1e CM007964.1 0 303 fwd a2adcee3-91c0-449d-a653-a7e969e00afe CM007964.1 0 46253 fwd 306 0.040656 0.113539 309 0.064228 0.189773 315 0.046056 0.182314 320 0.039238 0.156609

Regards,

John Ireland

Institute of Genetics and Cancer The University of Edinburgh

MBoemo commented 12 months ago

I think you're confusing v2 with v3. Version 2 only detects BrdU so the output is a read header followed by two columns (genome position on the left and probability of BrdU on the right). You're using v3.1.2 which detects BrdU and EdU, hence you get two probabilities (one for each analogue) at each genomic position.

3.1.2 uses Tensorflow 2.4.1 which we routinely run on A100s.