rnn_model/using_torch/optimise.py

from typing import Type

import numpy as np
import optuna
import torch
from torch import nn

from utils import tts_last_n
from .early_stopping import EarlyStopping
from .loaders import get_dataloader
from .train_eval import train_model, validate_model, predict
from sklearn.metrics import roc_auc_score
import math


def objective_tune_reg_a(trial: optuna.Trial,
                         model_class: Type[nn.Module],
                         inputs: np.ndarray,
                         targets: np.ndarray
                         ) -> float:
    criterion = nn.MSELoss()
    hidden_size = trial.suggest_categorical('hidden_size', [4, 8, 16, 32, 64, 128])
    num_layers = trial.suggest_int('num_layers', 1, 4)
    dropout = trial.suggest_float('dropout', 0.05, 0.5, log=True)
    learning_rate = trial.suggest_float('learning_rate', 1e-4, 1e-1, log=True)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    model = model_class(
        input_size=inputs.shape[2],
        hidden_size=hidden_size,
        num_layers=num_layers,
        dropout=dropout,
        output_size=1
    )

    train_seq, test_seq, train_tar, test_tar = tts_last_n(inputs, targets, 252)
    tloader = get_dataloader(train_seq, train_tar, device)
    vloader = get_dataloader(test_seq, test_tar, device)
    train_model(
        model=model,
        criterion=criterion,
        train_loader=tloader,
        val_loader=vloader,
        early_stopping=EarlyStopping(5, 1e-6),
        epochs=100,
        lr=learning_rate,
        verbose=0
    )

    return math.sqrt(validate_model(model, vloader, criterion, device))


def objective_tune_clas_a(trial: optuna.Trial,
                          model_class: Type[nn.Module],
                          inputs: np.ndarray,
                          targets: np.ndarray
                          ) -> float:
    criterion = nn.BCELoss()
    hidden_size = trial.suggest_categorical('hidden_size', [4, 8, 16, 32, 64, 128])
    num_layers = trial.suggest_int('num_layers', 1, 4)
    dropout = trial.suggest_float('dropout', 0.05, 0.5, log=True)
    learning_rate = trial.suggest_float('learning_rate', 1e-4, 1e-1, log=True)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    model = model_class(
        input_size=inputs.shape[2],
        hidden_size=hidden_size,
        num_layers=num_layers,
        dropout=dropout,
        output_size=1
    )

    train_seq, test_seq, train_tar, test_tar = tts_last_n(inputs, targets, 252)
    tloader = get_dataloader(train_seq, train_tar, device)
    vloader = get_dataloader(test_seq, test_tar, device)
    train_model(
        model=model,
        criterion=criterion,
        train_loader=tloader,
        val_loader=vloader,
        early_stopping=EarlyStopping(5, 1e-6),
        epochs=100,
        lr=learning_rate,
        verbose=0
    )

    labels, preds = predict(model, vloader, device)
    return roc_auc_score(labels, preds)