def run_best(trainset, testset):
train_loader = data.DataLoader(trainset,
batch_size=bsz,
shuffle=False,
collate_fn=collate)
test_loader = data.DataLoader(testset,
batch_size=bsz,
shuffle=True,
collate_fn=collate)
model = SeqLSTM(device=device, input_size=best_params['input_size'],
embedding_size=best_params['embedding_size'],
hidden_size=best_params['hidden_size'],
output_size=best_params['output_size'],
n_layers=best_params['n_layers'],
bidir=best_params['bidir'],
embedding=None).to(device)
opt = optim.Adam(model.parameters(), lr=best_params['lr'], weight_decay=0.0005)
for i in range(1, best_params['num_epochs'] + 1):
train_epoch(model, opt, train_loader)
test_loss, test_acc, test_tpr, test_tnr, test_auc = evaluate(model, test_loader)
result = "loss = {}, acc = {}, ".format(test_loss, test_acc)
result += "tpr/sensitvity = {}, tnr/specificity = {}, ".format(test_tpr, test_tnr)
result += "AUROC = {}".format(test_auc)
with open(output_file, 'a+') as f:
f.write("\n\nFinal model: " + str(best_params) + '\n' + result + '\n')
return
def run(params, trainset):
global highest_auc, best_params
print(params)
trainset.split()
total_acc = 0
total_auc = 0
num_epochs = 0
for i in range(num_folds):
train, vali = trainset.get_fold(i)
train_loader = data.DataLoader(train,
batch_size=bsz,
shuffle=False,
collate_fn=collate)
vali_loader = data.DataLoader(vali,
batch_size=bsz,
shuffle=False,
collate_fn=collate)
model = SeqLSTM(
device=device,
input_size=params["input_size"],
embedding_size=params["embedding_size"],
hidden_size=params["hidden_size"],
output_size=params["output_size"],
n_layers=params["n_layers"],
bidir=params["bidir"],
embedding=None,
).to(device)
opt = optim.Adam(model.parameters(),
lr=params["lr"],
weight_decay=0.0005)
for i in range(1, epochs + 1):
num_epochs = i
train_loss = train_epoch(model, opt, train_loader)
vali_loss, acc, tpr, tnr, auc = evaluate(model, vali_loader)
# early stopping criterion
if vali_loss > train_loss:
break
vali_loss, acc, tpr, tnr, auc = evaluate(model, vali_loader)
result = "train_loss = {}, vali_loss = {}, ".format(
train_loss, vali_loss)
result += "acc = {}, tpr/sensitvity = {}, ".format(acc, tpr)
result += "tnr/specificity = {}, AUROC = {}".format(tnr, auc)
total_acc += acc
total_auc += auc
acc = total_acc / num_folds
auc = total_auc / num_folds
if auc > highest_auc:
highest_auc = auc
best_params = params
best_params["num_epochs"] = num_epochs
with open(output_file, "a+") as f:
f.write("\n\n" + str(params) + "num_epochs: " + str(num_epochs) +
"\n" + result + "\n")