• references

  • https://www.rapidtables.com/code/linux/gcc.html
  • https://www.mauriciopoppe.com/notes/computer-science/operating-systems/bin/make/

• Important!!

  • The job number should be manually defined in util/config.h !
  • This is due to the usage of for performance enhancement
  • The bad news is that bitset for STL do not support dynamically change on its size
  • For the vanilla Lu algorithm, prune__OneRjSumCj__LU_AND_SAL__Theorem1 must be on

• Simple Usage

  • train the network first: make -j5 net ./net -d directory name
  • prepare binary for inference change to "tester" branch, and compile binary for different labeler move the binaries to current "build" directory and rename them with
  • do valiation after training and pick the model that performs best ./track -mulval 100000 ../saved_model ../case/case-small/validation > case-small.valid
  • copy the best model for testing: cp ../saved_model/piNet_41000.pt ../saved_model/inf/piNet.pt
  • do testing: make -j5 main move the resulting .pt files to /inf directory ./main -d test directory name
  • draw: change the filename for evaluation in eval.ipynb run it

• Reproducibility

  • The torch and c++ share the same random seed, which is set in util/config.h with name "RANDOM_SEED"
  • The torch::manual_seed() is called in three places, the initialization of weight of both pi and q network and the main function that powers the training.
  • The randomness of c++ comes from the order of choosing the instances in the training process.

• Usage

  • for inference, use dry_submit = true at submit() to decrease memory use

• Grid Search

  • make sure the corresponding variable is defined in config.h
  • make sure the corresponding variable is defined in cmakelist.txt
  • rerun the cmake command and assign the value to the variable, ex:
    • cmake -D V_HIDDEN_DIM='128' -D V_MAX_NUM_CNTR='9' -D V_LR_PI='4e-6' -D V_LR_Q='4e-5'
  • recompile the code

• use "-f" to provide the training set path

• by default the program requires a /validation subdirectory in the training set directory for validation purpose, user can provide a different directory by "-v"