import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, random_split
from torchtext.data.functional import to_map_style_dataset
from hyperopt import fmin, tpe, hp, Trials, STATUS_OK
defobjective(params):
EPOCHS = int(params['epochs'])
TASA_APRENDIZAJE = params['lr']
BATCH_TAMANO = int(params['batch_size'])
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.SGD(model.parameters(), lr=TASA_APRENDIZAJE)
train_iter, test_iter = DBpedia()
train_dataset = to_map_style_dataset(train_iter)
test_dataset = to_map_style_dataset(test_iter)
num_train = int(len(train_dataset) * 0.95)
split_train_, split_valid_ = random_split(train_dataset, [num_train, len(train_dataset) - num_train])
train_dataloader = DataLoader(split_train_, batch_size=BATCH_TAMANO, shuffle=True, collate_fn=collate_batch)
valid_dataloader = DataLoader(split_valid_, batch_size=BATCH_TAMANO, shuffle=True, collate_fn=collate_batch)
best_valid_loss = float('inf')
for epoch in range(1, EPOCHS + 1):
train_loss, train_acc = train(train_dataloader, model, optimizer, criterion)
validation_loss, validation_acc = evaluate(valid_dataloader, model, criterion)
if validation_loss < best_valid_loss:
best_valid_loss = validation_loss
torch.save(model.state_dict(), 'best-model.pt')
return {'loss': best_valid_loss, 'status': STATUS_OK}
# Define the search space
space = {
'epochs': hp.quniform('epochs', 1, 10, 1),
'lr': hp.loguniform('lr', -5, 0),
'batch_size': hp.quniform('batch_size', 32, 256, 16),
}
trials = Trials()
best = fmin(
fn=objective,
space=space,
algo=tpe.suggest,
max_evals=50,
trials=trials
)
print("Best hyperparameters found were:", best)
Con este sencillo codigo podemos entrenar la red optimizando los hiperparametros con hyperOpt para el mejor resultado posible.
