config¶
Baseline for CKODE project
usage: python trainModel.py [-h] [--seed SEED] [--mech_path MECH_PATH] [--input_path INPUT_PATH]
[--label_path LABEL_PATH] [--shuffle] [--batch_size BATCH_SIZE]
[--valid_batch_size VALID_BATCH_SIZE] [--train_size TRAIN_SIZE]
[--valid_size VALID_SIZE] [--valid_ratio VALID_RATIO] [--prefetch PREFETCH]
[--num_workers NUM_WORKERS] [--pin_memory]
[--power_transform POWER_TRANSFORM] [--delta_t DELTA_T] [--dim DIM]
[-l LAYERS [LAYERS ...]] [--net_type NET_TYPE] [--actfun ACTFUN]
[--max_epoch MAX_EPOCH] [--epoch_decay EPOCH_DECAY] [-lr LEARNRATE]
[--lr_decay_rate LR_DECAY_RATE] [--batch_grow_rate BATCH_GROW_RATE]
[--lossfun LOSSFUN] [--optim OPTIM] [--modelname MODELNAME]
[--model_path MODEL_PATH] [--device DEVICE] [--n_gpu N_GPU] [-DP] [-DDP]
[--local_rank LOCAL_RANK] [-ws WORLD_SIZE] [--backend BACKEND]
[-note DESCRIPTION]
Named Arguments¶
- --seed
random seed
Default: 2022
- --mech_path
chemical mechnism file dir
- --input_path
input dataset dir
- --label_path
label dataset dir
- --shuffle
shuffle the training dataset
Default: True
- --batch_size
use for training duration per worker
Default: 1024
- --valid_batch_size
use for validation duration per worker
Default: 8192
- --train_size
training dataset size
- --valid_size
validation dataset size
- --valid_ratio
split percentages of training data as validation
Default: 0.1
- --prefetch
use for training duration per worker
Default: 10
- --num_workers
num_workers for dataloaders
Default: 16
- --pin_memory
pin_memory for DataLoader
Default: True
- --power_transform
power_transform for BCT
Default: 0.1
- --delta_t
dnn time step
Default: 1e-06
- --dim
feature dimention of data
Default: -1
- -l, --layers
dnn hidden layers
Default: [1600, 800, 400]
- --net_type
dnn type
Default: “fc”
- --actfun
activation function
Default: “gelu”
- --max_epoch
max epochs in training
Default: 5000
- --epoch_decay
epoch interval for lr decay
Default: 2500
- -lr, --learnrate
Default: 0.0001
- --lr_decay_rate
learning rate decay
Default: 0.1
- --batch_grow_rate
every [epoch_decay] batch_size grows
Default: 128
- --lossfun
loss funtion: MSE,L1/MAE,CEl
Default: “L1”
- --optim
optimizer for training
Default: “Adam”
- --modelname
the dnn model name
- --model_path
the dnn model dir
- --device
device setup
Default: “cuda:0”
- --n_gpu
gpu numbers of the machine
Default: 0
- -DP, --use_DP
whether use torch.nn.DataParallel
Default: False
- -DDP, --use_DDP
whether use DistributedDataParallel
Default: False
- --local_rank
local rank
Default: -1
- -ws, --world_size
world_size for DDP
Default: 1
- --backend
current process backend for DDP, gloo,nccl,mpi
Default: “nccl”
- -note, --description
description of the experiment(purpose/target/motivation)
Default: “test”