Commit 0d40c194 authored by Zhang, Chen's avatar Zhang, Chen
Browse files

new default training params

parent 4d15f4b1

scripts/train_gpt.py

+11 −11
Original line number Diff line number Diff line
@@ -3,28 +3,28 @@ 
from tgreft.train.train_gpt import train



if __name__ == "__main__":

    model_params = {

        "d_model": 1024,

        "n_head": 8,

        "num_encoder_layers": 4,

        "input_dim": 150,

        "output_dim": 17,

        "to_log": True,

    }

        

    train_params = {

        "cuda_id": 0,

        "n_epochs": 50,  # seems like 50 is the point where training and validation loss diverge

        "n_training": 1_500_000,

        "error": 0.07,

        "batch_size": 200,

        "batch_size": 250,

        "learning_rate": 0.0057929655918116715,

        "weight_decay": 7.198921885462489e-07,

        "optimizer": "SGD",

        "loss": "huber",

        "cache_dir": "data",

        "experiment_name": "Train_REFL_GPT",

        "run_name": "gpt_d1024_h4_l4",

    }



    model_params = {

        "d_model": 1024,

        "n_head": 4,

        "num_encoder_layers": 4,

        "input_dim": 150,

        "output_dim": 17,

        "to_log": True,

        "run_name": "gpt_d1024_h8_l4_newlog",

    }



    train(

src/tgreft/utils/data/data_loader.py

+1 −1
Original line number Diff line number Diff line
@@ -84,7 +84,7 @@ def generate_data( 
    # generate the reference parameters

    parameters_ref = np.column_stack(

        [

            np.random.uniform(3.5, 6.5, n_dataset),  # electolyte_sld,

            np.random.uniform(5.0, 7.0, n_dataset),  # electolyte_sld,

            np.random.uniform(5, 120, n_dataset),  # electolyte_roughness,



            np.random.uniform(-5.0, 6.5, n_dataset),  # sei_sld,