main.py

from dataset import get_data, get_iterators
from fitness import fit, evaluate
from utils import load_embedding
import pickle

if __name__ == "__main__":
    config = {
        "embedding_path": "./dataset/embeddings/w2v/w2v_title_300Epochs_1MinCount_9ContextWindow_100d.txt",
        "epochs": 30,
        "lr": 1e-03,
        "rnn_units": 100,
        "convs_filter_banks": 32,
        "convs_kernel_size": 2,
        "denses_depth1": 32,
        "denses_depth2": 16,
        "similarity_type": "dot",
        "automl_path": './exps/w2v_10Epochs_100d_CrossEntropy_BothDenses.pickle',
    }

    if config["automl_path"]:
        with open(config["automl_path"], "rb") as f:
            params = pickle.load(f)
        max_f1 = 0
        best_params = {}
        for param in params:
            if param[1] > max_f1:
                best_params = param[0]
                max_f1 = param[1]

        config["lr"] = best_params["lr"]
        config["rnn_units"] = int(best_params["rnn_units"])
        config["convs_filter_banks"] = int(best_params["convs_filter_banks"])
        config["denses_depth1"] = int(best_params["dense_depth1"])
        config["denses_depth2"] = int(best_params["dense_depth1"])
        config["similarity_type"] = (
            "dot" if best_params["similarity_type"] == 0 else "cosine"
        )
        print(config)

    train_ds, valid_ds, test_ds, TEXT = get_data(
        train_path="./dataset/computers/train/computers_train_splitted_medium.json",
        valid_path="./dataset/computers/valid/computers_splitted_valid_medium.json",
        test_path="./dataset/computers/test/computers_gs.json",
    )
    train_dl, valid_dl, test_dl = get_iterators(train_ds, valid_ds, test_ds)
    load_embedding(TEXT, config["embedding_path"])

    model = fit(
        TEXT,
        train_dl,
        valid_dl,
        config=config,
        hidden_dim=config["rnn_units"],
        conv_depth=config["convs_filter_banks"],
        kernel_size=config["convs_kernel_size"],
        dense_depth1=config["denses_depth1"],
        dense_depth2=config["denses_depth2"],
        lr=config["lr"],
        similarity="dot",
        loss="CrossEntropyLoss",
        validate_each_epoch=True,
        trainable=True
    )
    evaluate(model, test_dl, print_results=True)