CUDA out of memory in unexpected cases

[havu73]

Hi! Thank you so much for providing this implementation and documentations. I have been looking for the pytorch implementation of enformer for a while.

When I run the provided code test_pretrained.py using 8G of GPU, the model behaves as expected. The input sequence dimension is [131072, 4], which means the input DNA sequence is of length 131072 bp.

However, when I started to write a script that would take the embedding of some genomic sequences, which according to enformer paper and the default parameters, is of length 196608 bp, then I started to get errors of CUDA out of memory. This happens even if my batch size is 1. It also happens when I increases my memory request to 32Gb (instead of 8G as before). If I run the same code to get embeddings for input sequences (batchsize=1 to maximum of 8), then it can run on a CPU with 32Gb of memory.

What confused me is that the code test_pretrained.py seems to run fine with just 8GB and CUDA, but when I just try to get embeddings on 1 input sequence (which theoretically less computing than what is done in test_pretrained.py), I get such an error. I have been looking at various platforms to fix this, but I have not been successful. Do you have any suggestions about why that is the case?

Thank you so much. Below is a snapshot of my code to get embeddings:

import torch
from enformer_pytorch import Enformer, seq_indices_to_one_hot, GenomeIntervalDataset
from torch.utils.data import DataLoader

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Enformer.from_pretrained('enformer-official-rough')
model = model.to(device)

ds = GenomeIntervalDataset(
    bed_file = '/data/gnomad/v2.1.1_hg19/variants/MAF/chr1_maf.bed.gz',                       
    fasta_file = '/data/ucsc/sequence/hg19/chr1.fa.gz',                        
    return_seq_indices = True,                          
    shift_augs = None,                            
    context_length = 196_608,
    chr_bed_to_fasta_map = {
        'chr1': 'chr1'
    }
)

[lucidrains]

@havu73 hello! while the network was trained at 191k, you can still run it at 131k (or even less, provided it is divisible by the total downsample factors of the conv stem, which is 128), if that is what will fit into memory

it would have a bit less context, but perhaps good to see if you get signal first before scaling up

[havu73]

Thank you so much @lucidrains I really appreciate it! This is super helpful. I have been a bit busy to catch up on this, but once I tested this out I will close the issue. Thanks again very much!