LouisFaure

#!/bin/bash
# --- Configuration ---
PARTITION_NAME="componc_gpu" # Define the partition name here
# --- Function to convert memory string (e.g., "972G", "1031308M") to MB ---
convert_mem_to_mb() {
    local mem_str="$1"
    local value
    local unit
    if [[ "$mem_str" =~ ([0-9]+)([MG]) ]]; then
        value="${BASH_REMATCH[1]}"

LouisFaure

#!/bin/bash

# This script parses the output of 'scontrol show node <node_name>'
# to determine the available CPU, memory, and GPUs for a new task.

# Check if a node name is provided as an argument
if [ -z "$1" ]; then
    echo "Usage: $0 <node_name>"
    echo "Example: $0 iscg009"
    exit 1

LouisFaure

Negative Binomial Modeling for scRNA-seq

The negative binomial (NB) distribution is the canonical statistical model for single-cell RNA-seq (scRNA-seq) gene expression integer counts. It’s favored over the Poisson distribution because real scRNA-seq data are overdispersed—their variance greatly exceeds their mean–due to both biological and technical factors (cell heterogeneity, bursty transcription, varied sequencing depth, dropouts, etc.).

The Poisson’s simple assumption (variance equals mean) fails in this context. In contrast, the NB augments the Poisson distribution with a dispersion parameter ($\theta$), so its variance is $\sigma^2 = \mu + \mu^2 / \theta$ (for mean $\mu$). This flexibility means two genes with equal mean expression but different variances (as commonly seen) can be modeled, simply by adjusting $\theta$.

From a generative perspective, the NB arises as a Poisson–Gamma mixture: each cell/gene’s unknown expression rate $\lambda$ is drawn from a Gamma distribution

LouisFaure

import subprocess as subp
import re
if jax.default_backend()=='gpu':
    # Step 2: Call nvidia-smi and get the output
    result = subp.run([
        'nvidia-smi', 
        '--query-compute-apps=pid,used_memory', 
        '--format=csv,noheader,nounits'
    ], stdout=subp.PIPE, text=True)
    

LouisFaure

import scipy.sparse as sp
import numpy as np

def get_raw(adata):
    X = adata.X
    if not sp.isspmatrix_csr(X):
        raise Exception("matrix needs to be in csr format")
    if is_integer_vector(adata.X[0,:].data):
        raise Exception("matrix is already in integer format")
    row_mins = np.minimum.reduceat(X.data, X.indptr[:-1])

LouisFaure

import os, sys
os.environ['R_HOME'] = sys.exec_prefix+"/lib/R/"
import rpy2

LouisFaure

import mudatasets
import numpy as np
import pandas as pd
import scanpy as sc
import anndata as ad
import muon as mu

mdata = mudatasets.load("brain3k_multiome", full=True)
mdata.var_names_make_unique()

LouisFaure

import scanpy as sc
import cupy as cp
from scipy.sparse import csr_matrix, find, issparse
from scipy.sparse.linalg import eigs
import numpy as np
import pandas as pd
import cudf
import glob
from cupy.sparse import cupyx as cpx
from cupyx.scipy.sparse import coo_matrix as coo_matrix_gpu

LouisFaure

#!/usr/bin/env python
import sys
import warnings
warnings.filterwarnings("ignore")
import logging
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
from anndata2ri.scipy2ri._r2py import rmat_to_spmat
import anndata as ann
import pandas as pd