dream-challenge-gene-expression

Description

• train
  • Train with training data
  • Validate with validation data (validation data is different according to the 'fold' argument)
  • Log to wandb
  • Save best checkpoint
  • When training ends, test with HQ_testdata & log to wandb
• test
  • Test with HQ_testdata
  • Log to wandb
• predict
  • Make submission.txt file with challenge test data

How to run

View help to identify all the available parameters

python train.py --help

Train model with chosen model configuration from configs/model/ (You should write config(.yaml) file first)

python train.py model=deepfamq_conjoined_adamw

You can override any parameter from command line like this (or you can just fix config files)

python train.py trainer.max_epochs=20 datamodule.batch_size=64 model.net.conv_kernel_size=15

Set 'name' argument to represent the settings you used (ckpt directory & wandb group name are automatically set)

python train.py model=deepfamq_conjoined_adamw model.net.conv_kernel_size=15 name=deepfamq_conjoined_adamw_conv15

Train models with 5-fold CV scheme (You can use snakemake to wrap these runs)

python train.py model=deepfamq_conjoined_adamw trainer.gpus=[0] fold=0

python train.py model=deepfamq_conjoined_adamw trainer.gpus=[0] fold=1

                              ...

Train model with whole training set & validate with HQ_testdata (just set fold=None)

python train.py model=deepfamq_conjoined_adamw trainer.gpus=[0] fold=None

Test model with HQ_testdata (ckpt_path is automatically set according to args 'name' & 'fold')

python test.py model=deepfamq_conjoined_adamw name=deepfamq_conjoined_adamw_conv15 fold=0

Make prediction file to be submitted (ckpt_path is automatically set according to args 'name' & 'fold')

python predict.py model=deepfamq_conjoined_adamw name=deepfamq_conjoined_adamw_conv15 fold=0