install tensorlfow on stjude HPC¶
Keras2+Tensorflow2¶
module load conda3/202011
conda create -n keras python=3.9
module load cuda11/toolkit/11.3
module load cudnn/8.2.0.53
nvidia-smi # check GPU usage
echo $CONDA_PREFIX # check conda dir
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/' > $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/
pip install tensorflow
python3 -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"
python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
conda install -c conda-forge jupyterlab
conda install -c conda-forge swifter
conda install -c anaconda seaborn
# run jupyter lab
export XDG_RUNTIME_DIR=""
mkdir /scratch_space/yli11/68985
export JUPYTER_RUNTIME_DIR=/scratch_space/yli11/68985
jupyter lab --ip='*' --NotebookApp.token='' --NotebookApp.password='' --no-browser --port 44444
# install janggu
# pysam install has an error in my system, so I used conda to install it
pip install --no-deps janggu[tf2_gpu]
# install those dependences manually
pip install biopython==1.77
vim ~/.conda/envs/keras/lib/python3.9/site-packages/janggu/layers.py
# change from keras.layers.wrappers import Wrapper to
# from tensorflow.keras.layers import Wrapper
Submit GPU job: bsub -gpu "num=1/host:mode=exclusive_process" < keras.lsf
check usage: ``nvidia-smi ``
#BSUB -P run_jupyte
#BSUB -o keras_%J_%I.out -e keras_%J_%I.err
#BSUB -n 1
#BSUB -q gpu
#BSUB -R "span[hosts=1] rusage[mem=60000]"
#BSUB -J "keras"
module purge
ncore=1
mem=60000
q=gpu
echo $PATH
PATH=/home/yli11/HemTools/bin:/hpcf/lsf/lsf_prod/10.1/linux3.10-glibc2.17-x86_64/etc:/hpcf/lsf/lsf_prod/10.1/linux3.10-glibc2.17-x86_64/bin:/usr/lpp/mmfs/bin:/usr/lpp/mmfs/lib:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/opt/ibutils/bin:/sbin:/cm/local/apps/environment-modules/3.2.10/bin:/opt/puppetlabs/bin:/home/yli11/Programs/HiC-Pro-2.11.1/bin/utils:/research/dept/hem/common/sequencing/chenggrp/pipelines/bin
echo $PATH
module load conda3/202011
source activate /home/yli11/.conda/envs/keras
module load cuda11/toolkit/11.3
module load cudnn/8.2.0.53
module load texlive/20190410
echo $PATH
echo `which python`
export XDG_RUNTIME_DIR=""
mkdir /scratch_space/yli11/$$
export JUPYTER_RUNTIME_DIR=/scratch_space/yli11/$$
cp /home/yli11/HemTools/share/misc/Introduction_6_21_2021_v2.ipynb run_jupyterlab_gpu_yli11_2022-07-07/
cp /home/yli11/HemTools/share/tutorial/* run_jupyterlab_gpu_yli11_2022-07-07/
yes | cp -rf /research/rgs01/home/clusterHome/yli11/.local/share/jupyter/kernels/ ~/.local/share/jupyter/
jupyterlab.py run_jupyterlab_gpu_yli11_2022-07-07/Introduction_6_21_2021_v2.ipynb
ref¶
https://github.com/kheyer/Genomic-ULMFiT https://github.com/kheyer/Genomic-ULMFiT/blob/master/Methods/Methods%20Long%20Form.ipynb https://github.com/rschwess/tutorial_dl_for_genomics https://www.tensorflow.org/api_docs/python/tf/keras/applications/efficientnet_v2/EfficientNetV2S https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/ https://towardsdatascience.com/an-in-depth-efficientnet-tutorial-using-tensorflow-how-to-use-efficientnet-on-a-custom-dataset-1cab0997f65c
Efficient net is probably good enough, but I want to try adding attention and some forget gate units, GELU, bidirectional, BERT. I haven’t try deep learning methods based on MSA, I think the reason protein people like to use MSA instead of one-hot encoding is because amino acid is 20 and DNA is only 4.
https://www.biorxiv.org/content/10.1101/2020.11.27.401232v2.full.pdf https://github.com/facebookresearch/esm https://www.biorxiv.org/content/10.1101/622803v4.full.pdf https://github.com/shashwattrivedi/Simple-TransformerEncoder-keras
Tensorflow¶
Currently, our GPU nodes only have cuda8.0, and for this, the latest TF version you can use is 1.3
ssh nodegpu121
module load conda3
conda create -n janggu3
conda activate janggu3
conda install -c anaconda tensorflow-gpu=1.3
To test GPU, see: https://www.tensorflow.org/beta/guide/using_gpu
Janggu¶
Not sure about the TF version this package used. Need to test.
conda install -c bioconda pybigwig
python -m pip install janggu
Test your code, should be finished within 5 min, since we are using GPU:
git clone https://github.com/BIMSBbioinfo/janggu
cd janggu
python ./src/examples/classify_fasta.py single