Skip to content

Latest commit

 

History

History
156 lines (120 loc) · 9.21 KB

README.md

File metadata and controls

156 lines (120 loc) · 9.21 KB
Raw

Training module

Load the converted PyTables files and train DNNs with MXNet.

Setup

Install miniconda if you don't have it:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
# Follow the insturctions to finish the installation

Verify the installation is successful by running conda info.

If you cannot run conda command, check if the you added the conda path to your PATH variable in your bashrc/zshrc file, e.g.,

export PATH="$HOME/miniconda3/bin:$PATH"

Set up the environment for training

The following instruction is only for training with Nvidia GPU. CUDA 8.0 and cuDNN (>=5) is required.

# create a new conda environment
conda create -n mxnet python=3.7

# set up ROOT
# (below assumes centos7, for other systems please modify the ROOT installation path accordingly)
mkdir -p $HOME/miniconda3/envs/mxnet/etc/conda/
cd $HOME/miniconda3/envs/mxnet/etc/conda/
mkdir activate.d  deactivate.d
cd activate.d
# create the env_vars.sh file to get ROOT environment
cat << EOF > env_vars.sh
#!/bin/sh
# $HOME/miniconda3/envs/prep/etc/conda/activate.d/env_vars.sh
echo "Source root environment..."
# ROOT
source /cvmfs/sft.cern.ch/lcg/app/releases/ROOT/6.14.06/x86_64-centos7-gcc48-opt/bin/thisroot.sh
EOF

# activate the environment
source activate mxnet

# install the necessary python packages
conda install -c anaconda hdf5
pip install numpy numexpr pandas scikit-learn scipy tables matplotlib
pip install root-numpy

# install mxnet -- this depends on the CUDA version (the current recommendation is CUDA 10.1)
pip install mxnet-cu101==1.5.1.post0
pip install gluonnlp==0.8.3


# for other CUDA versions, please check https://mxnet.incubator.apache.org/install/

Instructions

Start a new training

python train_pfcands_simple.py --data-config data_ak8_parts_sv --network resnet_simple --model-prefix /path/to/model/checkpoints/model-name-without-suffix --batch-size 512 --optimizer adam --lr 0.001 --lr-step-epochs "10,20,30,50" --num-epochs 80 --data-train '/path/to/data/train_file_*.h5' --dataloader-nworkers 2 --dataloader-qsize 32 --gpus 0 &> /path/to/logfile.log &
  • --data-config: which configuration of the inputs to use. data_ak8_parts_sv -> data/data_ak8_parts_sv.py.
  • --network: the DNN model to use. resnet_simple -> symbols/resnet_simple.py.
  • --model-prefix: path for saving training checkpoints at the end of each epoch. The saved model can be used restarting a interrupted training, as well as running predictions to evaluate the performance.
  • --batch-size: minibatch size for training. Adjust this according the model complexity to fit the GPU memory. This can also be tuned as a hyperparameter.
  • --optimizer: training optimizer. Currently support adam and sgd.
  • --lr: learning rate.
  • --lr-step-epochs: the epochs to reduce the lr by --lr-factor (defaults to 0.1), e.g., "10,20,30,50" means the 10th, 20th, 30th, and 50th epoch
  • --num-epochs: max number of epochs to run
  • --data-train: path for the training files; support Unix style pathname pattern expansion (i.e., * and ?) using glob in python, but make sure you wrap it with single quote (').
  • --dataloader-nworkers: number of parallel threads for loading the dataset.
  • --dataloader-qsize: queue size of the dataloader (adjust according to the RAM size and --dataloader-nworkers).
  • --gpus: set which GPU to use. Multiple GPUs can be specified as a comma seperated string, e.g., "0,1,2,3". Set to an empty string "" if you want to use CPU.
  • More options can be found by running python train_pfcands_simple.py -h or checking the source code.
  • &> /path/to/logfile.log & will redirect both stdout/stderr to the file /path/to/logfile.log, and the training & will run this process in the background. You can view the log file with less (e.g., type F to follow the tail of the file).

Resume an interrupted training

python train_pfcands_simple.py --data-config data_ak8_parts_sv --network resnet_simple --model-prefix /path/to/model/checkpoints/model-name-without-suffix --batch-size 512 --optimizer adam --lr 0.001 --lr-step-epochs "10,20,30,50" --num-epochs 80 --data-train '/path/to/data/train_file_*.h5' --dataloader-nworkers 2 --dataloader-qsize 32 --gpus 0 --load-epoch 20 &>> /path/to/logfile.log &
  • Use --load-epoch option to load the checkpoint and resume the training (e.g., --load-epoch 20 will resume the training from the Epoch 20).
  • &>> allows you to append to the log file instead of overwriting it.
  • Note that although this is possible, it is not recommended in general as some optimzers have weight decay which depends on the number of epoch.

Run prediction with trained model

python train_pfcands_simple.py --data-config data_ak8_parts_sv --network resnet_simple --model-prefix /path/to/model/checkpoints/model-name-without-suffix --load-epoch 60 --batch-size 32 --data-train '/path/to/data/train_file_*.h5' --dataloader-nworkers 2 --dataloader-qsize 32 --gpus 0 --predict --data-test '/path/to/test-data/JMAR/Top/train_file_*.h5' --predict-output /path/to/output/mx-pred_Top.h5
  • --predict: run in prediction mode instead of training.
  • --load-epoch: load the model parameter from which epoch (e.g., --load-epoch 5 will load model-0005.params).
  • --batch-size 32: a smaller batch size is preferred in prediction mode to avoid losing events.
  • --data-test: path for the testing files; support Unix style pathname pattern expansion (i.e., * and ?) using glob in python, but make sure you wrap it with single quote (').
  • --predict-output: output file. Both PyTables (.h5) and root file will be created.

Reference training/prediction command

Nominal version (94X, V1)

Training:

python train_pfcands_simple.py --data-config data_ak8_pfcand_sv --network sym_ak8_pfcand_sv_resnet_v1 --model-prefix /data/hqu/training/mxnet/models/20190326_ak8_classrewgt/pfcand_sv_resnet_v1/resnet --batch-size 1024 --optimizer adam --lr 0.001 --lr-step-epochs "15,30,40" --num-epochs 50 --data-train '/data/hqu/ntuples/20190326_ak8/ak8puppi_parts_classrewgt/train_file_*.h5' --train-val-split 0.9 --dataloader-nworkers 3 --dataloader-qsize 48 --disp-batches 1000 --gpus 0 &> logs/train_ak8puppi_20190326_classrewgt_pfcand_sv_ref_resnet_v1.log &

Prediction:

python train_pfcands_simple.py --data-config data_ak8_pfcand_sv --network sym_ak8_pfcand_sv_resnet_v1 --model-prefix /data/hqu/training/mxnet/models/20190326_ak8_classrewgt/pfcand_sv_resnet_v1/resnet --load-epoch 39 --batch-size 128 --data-train '/data/hqu/ntuples/20190326_ak8/ak8puppi_parts_classrewgt/train_file_*.h5' --data-test '/data/hqu/ntuples/20190326_ak8/test_samples/JMAR/QCD/train_file_*.h5' --predict-output /data/hqu/training/mxnet/predict/20190326_ak8_classrewgt/pfcand_sv_resnet_v1/epoch39/JMAR/mx-pred_QCD.h5 --dataloader-nworkers 2 --dataloader-qsize 16 --gpus 0 --predict --predict-all &> logs/preds/pred_ak8puppi_20190326_classrewgt_pfcand_sv_ref_resnet_simple_epoch39.log &

Decorrelated version (94X, V1)

Training:

python train_features_adv.py \
 --data-config data_ak8_adv_pfcand_sv \
 --network block_ak8_adv_resnet_features_r_3x256_pfcand_sv_dropout \
 --model-prefix /data/hqu/training/mxnet/models/20190326_ak8_adv/pfcand_sv_resnet_features_r_3x256_dropout_mass30to250_22bins_advwgt5_advfreq10_lr_1e-2_decay0p1_30_60_90_advlr_1e-4_batch8k/resnet \
 --data-train '/data/hqu/ntuples/20190326_ak8/ak8puppi_parts_ptmasswgt/train_file_*.h5' \
 --dataloader-weight-scale 1 --dataloader-max-resample 100 --dataloader-nworkers 2 --dataloader-qsize 16 \
 --batch-size 8192 --num-epochs 120 --train-val-split 0.9 \
 --optimizer adam --lr 1e-2 --lr-factor 0.1 --lr-step-epochs "30,60,90" \
 --adv-lr 1e-4 --adv-lr-factor 0.1 --adv-lr-step-epochs "1000" \
 --adv-lambda 5 --adv-mass-min 30 --adv-mass-max 250 --adv-mass-nbins 22 --adv-train-freq 10 \
 --gpus 0 --disp-batches 100 \
 &> logs/dev-adv-ak8puppi_20190326_ptmasswgt_pfcand_sv_resnet_features_r_3x256_dropout_mass30to250_22bins_advwgt5_advfreq10_lr_1e-2_decay0p1_30_60_90_advlr_1e-4_batch8k.log &

Prediction:

python train_features_adv.py \
 --data-config data_ak8_adv_pfcand_sv \                                     
 --network block_ak8_adv_resnet_features_r_3x256_pfcand_sv_dropout \
 --model-prefix /data/hqu/training/mxnet/models/20190326_ak8_adv/pfcand_sv_resnet_features_r_3x256_dropout_mass30to250_22bins_advwgt5_advfreq10_lr_1e-2_decay0p1_30_60_90_advlr_1e-4_batch8k/resnet \
 --data-train '/data/hqu/ntuples/20190326_ak8/ak8puppi_parts_ptmasswgt/train_file_*.h5' \
 --dataloader-nworkers 2 --dataloader-qsize 16 \
 --batch-size 128 --data-test '/data/hqu/ntuples/20190326_ak8/test_samples/JMAR/QCD/train_file_*.h5' \
 --load-epoch 50 --predict-output /data/hqu/training/mxnet/predict/20190326_ak8_adv/pfcand_sv_resnet_features_r_3x256_dropout_mass30to250_22bins_advwgt5_advfreq10_lr_1e-2_decay0p1_30_60_90_advlr_1e-4_batch8k/JMAR/mx-pred_QCD.h5 \
 --predict --predict-all --predict-epochs "70,99,119" \
 --gpus 1 --disp-batches 100 \
 &> logs/preds/preds-adv-ak8puppi_20190326_ptmasswgt_pfcand_sv_resnet_features_r_3x256_dropout_mass30to250_22bins_advwgt5_advfreq10_lr_1e-2_decay0p1_30_60_90_advlr_1e-4_batch8k_epoch_70_99_119.log &