sasha@torchlab ~ % whoami
Hi, this is Sasha Chernenkoff
Machine Learning | Genomics | Design
sasha@torchlab ~ % python build_sequences_bed.py --genome hg38.fa --window 262144
Tiling hg38 into fixed 262,144 bp windows...
Wrote 41,623 windows -> sequences.bed
sasha@torchlab ~ % python dataset.py --patients manifest.tsv --vcf cohort.clean.vcf.gz
Building pseudo-diploid seq1/seq2 from germline SNPs...
Split -> chr6/7 test | chr8/9 val | remaining train (825/110/110)
Wrote 12 HDF5 shards -> petl_dataset.h5
sasha@torchlab ~ % sbatch train_petl.slurm
Submitted batch job 66841506
sasha@torchlab ~ % cat logs/train_66841506.out
PETL 2-GPU LoRA training run
--- Environment ---
Python: /home/scher/envs/petl-torch/bin/python
PyTorch 2.9.1, CUDA available: True, GPUs: 2
Node: rlf9403
NVIDIA A100-SXM4-40GB, 40960 MiB
NVIDIA A100-SXM4-40GB, 40960 MiB
-------------------
Optimizer (residual head): base lr=1e-05 (3,842), delta lr=0.0003 (3,842), no-decay (1), LoRA lr=0.0001 (655,360)
[epoch 0] train=0.812 val=0.774 patient_r=0.583 gene_r=0.412 pearson=0.612
[epoch 1] train=0.501 val=0.489 patient_r=0.617 gene_r=0.548 pearson=0.701
...
[epoch 39] train=0.198 val=0.213 patient_r=0.669 gene_r=0.706 pearson=0.834
Best checkpoint saved: petl_best.pt
Duration: 8h 10m. Trainer exit code: 0.
sasha@torchlab ~ %