# テストデータの選択と生成
test_dt = pre.load(dbpath=TEST_DATASET_TABLE_PATH)
if args.base_result_ver is None:
test_dt_indexes = dataselector.init_dataset_table_indexes()
else:
test_dt_indexes = dataselector.load_table_indexes(path=BASE_TEST_DT_INDEXES_PATH)
selector = dataselector.DataSelector(test_dt, test_dt_indexes)
test_dt_indexes, _ = selector.randomly_add(dataN=args.add_test, seed=args.seed)
test = selector.out_selected_dataset()
dataselector.save_dataset_table_indexes(test_dt_indexes, savepath=ADDED_TEST_DT_INDEXES_PATH)
print("Number of test data: {0}".format(len(test)))
testloader = torch.utils.data.DataLoader(test, batch_size=args.test_batch_size, shuffle=False, num_workers=2)
# 学習と評価と保存
model = training.process(trainloader, testloader, MODEL, args.epochs, args.lr, log_savepath=TRAIN_LOG_PATH)
torch.save(model.state_dict(), MODEL_SAVE_PATH)
# 学習結果の可視化
train_losses, test_losses, train_accs, test_accs = tools.load_train_log(path=TRAIN_LOG_PATH)
plot.plot_loss(train_losses, test_losses, savepath=LOSS_PLOT_RESULT_PATH)
plot.plot_acc(train_accs, test_accs, savepath=ACC_PLOT_RESULT_PATH)
# 訓練誤差収束速度の評価
train_err_speed = eval_loss.eval_err_speed(train_losses)
test_err_speed = eval_loss.eval_err_speed(test_losses)
eval_loss.save(savepath=ERR_SPEED_SAVE_PATH, train_err_speed=train_err_speed, test_err_speed=test_err_speed)
print("train loss error speed: {0}".format(train_err_speed))
print("test loss error speed: {0}".format(test_err_speed))
# 学習の設定値の記録
test_dt_indexes, _ = selector.randomly_add(dataN=args.add_test,
seed=args.seed)
test = selector.out_selected_dataset()
dataselector.save_dataset_table_indexes(
test_dt_indexes, savepath=ADDED_TEST_DT_INDEXES_PATH)
print("Number of test data: {0}".format(len(test)))
testloader = torch.utils.data.DataLoader(test,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=2)
# 学習と評価
model = training.process(trainloader,
testloader,
MODEL,
args.epochs,
args.lr,
lr_scheduling=LR_SCHEDULING,
log_savepath=TRAIN_LOG_PATH)
# 学習結果の可視化
train_losses, test_losses, train_accs, test_accs = tools.load_train_log(
path=TRAIN_LOG_PATH)
plot.plot_loss(train_losses, test_losses, savepath=LOSS_PLOT_RESULT_PATH)
plot.plot_acc(train_accs, test_accs, savepath=ACC_PLOT_RESULT_PATH)
# 訓練誤差収束速度の評価
train_err_speed = eval_loss.eval_err_speed(train_losses)
test_err_speed = eval_loss.eval_err_speed(test_losses)
eval_loss.save(savepath=ERR_SPEED_SAVE_PATH,
train_err_speed=train_err_speed,
batch_size=args.batch_size,
shuffle=True,
num_workers=2)
print("number of train data: {0}".format(len(train)))
# テストデータの生成
test = preprocess.load_dataset(dataset_dir=TEST_DIR,
dataN=args.dataN // 4,
imsize=(32, 32))
testloader = torch.utils.data.DataLoader(test,
batch_size=args.batch_size,
shuffle=True,
num_workers=2)
print("number of train data: {0}".format(len(test)))
# 学習と評価
model = net.LeNet(2)
model = training.process(trainloader,
testloader,
model,
args.epochs,
args.lr,
log_savepath=LOG_SAVEPATH)
torch.save(model.state_dict(), MODEL_SAVEPATH)
# 学習結果の可視化
train_losses, test_losses, train_accs, test_accs = log_loader.load_log(
path=LOG_SAVEPATH)
plot.plot_loss(train_losses, test_losses, savepath=LOSS_SAVEPATH)
plot.plot_acc(train_accs, test_accs, savepath=ACC_SAVEPATH)