Commit c997f706 authored by Václav Novák's avatar Václav Novák
Browse files

feat: add parallelization, modify suggestions

parent daac456c

main.py

+14 −1
Original line number Diff line number Diff line
@@ -15,6 +15,7 @@ from src.evaluation import evaluate 
NUM_WORKERS = 0 if platform.system() == "Windows" else 6

CUDA_CAPABLE = torch.cuda.is_available()

MAX_SECONDS_PER_TRIAL = 7200 if len(sys.argv) < 2 else int(sys.argv[1])

STORAGE_PATH = "results.log" if len(sys.argv) < 3 else sys.argv[2]



###############################################################################
@@ -54,7 +55,19 @@ def objective(trial): 
###############################################################################





if platform.system() != "Windows":

    storage = optuna.storages.JournalStorage(

        optuna.storages.JournalFileStorage(STORAGE_PATH),

    )

    

else:

    lock_obj = optuna.storages.JournalFileOpenLock(STORAGE_PATH)

    storage = optuna.storages.JournalStorage(

        optuna.storages.JournalFileStorage(STORAGE_PATH, lock_obj=lock_obj),

    )





if __name__ == '__main__':

    study = optuna.create_study(direction="maximize", sampler=optuna.samplers.TPESampler())

    study = optuna.create_study(direction="maximize", sampler=optuna.samplers.TPESampler(), storage=storage)



    study.optimize(objective)

src/builders.py

+1 −0
Original line number Diff line number Diff line
@@ -166,4 +166,5 @@ def build_l_layers(trial: Trial, first_input_size: int) -> nn.Sequential: 


            l_layers.append(dropout_layer)





    return nn.Sequential(*l_layers)

src/suggestions.py

+11 −11
Original line number Diff line number Diff line
@@ -12,20 +12,20 @@ def suggest_params_general(trial: Trial) -> None: 
    """



    # Parameters of input

    trial.suggest_categorical("input_size", [96, 112, 128])

    trial.suggest_int("input_size", 96, 128, step=16)



    # Parameters of learning

    trial.suggest_int("epochs", 6, 64)

    trial.suggest_categorical("batch_size", [4, 12, 16, 24, 32, 40, 48, 64])

    trial.suggest_int("epochs", 6, 72)

    trial.suggest_int("batch_size", 4, 80, step=4)



    # Parameters of loss function

    trial.suggest_float("margin", 1, 3)



    # Parameters of optimizer

    trial.suggest_float("lr", 1e-8, 0.0001)

    trial.suggest_float("weight_decay", 1e-7, 1e-2, log=True)

    trial.suggest_float("beta1", 0.85, 0.95)

    trial.suggest_float("beta2", 0.98, 0.999)

    trial.suggest_float("lr", 1e-5, 0.1)

    trial.suggest_float("weight_decay", 1e-6, 1e-2, log=True)

    trial.suggest_float("beta1", 0.65, 0.95)

    trial.suggest_float("beta2", 0.945, 0.999)



    # Parameters of the network itself

    trial.suggest_int("c_layers", 3, 5)
@@ -155,13 +155,13 @@ def suggest_l_layers(trial: Trial) -> None: 


    # Config for count of fully-connected layers

    if trial.params["l_layers"] == 3:

        outputs = [(256, 6144, 256), (128, 1536, 32), (2, 32, 2)]

        outputs = [(256, 4096, 256), (128, 1536, 32), (2, 32, 2)]



    elif trial.params["l_layers"] == 4:

        outputs = [(512, 8192, 256), (256, 2048, 64), (96, 768, 32), (2, 32, 2)]

        outputs = [(256, 6144, 256), (256, 2048, 64), (96, 768, 32), (2, 32, 2)]



    else:

        outputs = [(512, 8192, 256), (512, 3072, 128), (192, 896, 32), (96, 384, 16), (2, 32, 2)]

        outputs = [(256, 8192, 256), (256, 3072, 128), (128, 896, 32), (96, 384, 16), (2, 32, 2)]



    # Suggest each fully-connected layer

    for i in range(0, trial.params["l_layers"]):