def test_integration_median_absolute_percentage_error_with_output_transform(): np.random.seed(1) size = 105 np_y_pred = np.random.rand(size, 1) np_y = np.random.rand(size, 1) np.random.shuffle(np_y) np_median_absolute_percentage_error = 100.0 * np.median( np.abs(np_y - np_y_pred) / np.abs(np_y)) batch_size = 15 def update_fn(engine, batch): idx = (engine.state.iteration - 1) * batch_size y_true_batch = np_y[idx:idx + batch_size] y_pred_batch = np_y_pred[idx:idx + batch_size] return idx, torch.from_numpy(y_pred_batch), torch.from_numpy( y_true_batch) engine = Engine(update_fn) m = MedianAbsolutePercentageError(output_transform=lambda x: (x[1], x[2])) m.attach(engine, "median_absolute_percentage_error") data = list(range(size // batch_size)) median_absolute_percentage_error = engine.run( data, max_epochs=1).metrics["median_absolute_percentage_error"] assert np_median_absolute_percentage_error == pytest.approx( median_absolute_percentage_error)
def _test(n_epochs, metric_device): metric_device = torch.device(metric_device) n_iters = 80 size = 105 y_true = torch.rand(size=(size, )).to(device) y_preds = torch.rand(size=(size, )).to(device) def update(engine, i): return ( y_preds[i * size:(i + 1) * size], y_true[i * size:(i + 1) * size], ) engine = Engine(update) m = MedianAbsolutePercentageError(device=metric_device) m.attach(engine, "mape") data = list(range(n_iters)) engine.run(data=data, max_epochs=n_epochs) assert "mape" in engine.state.metrics res = engine.state.metrics["mape"] np_y_true = y_true.cpu().numpy().ravel() np_y_preds = y_preds.cpu().numpy().ravel() e = np.abs(np_y_true - np_y_preds) / np.abs(np_y_true) np_res = 100.0 * np.median(e) e_prepend = np.insert(e, 0, e[0], axis=0) np_res_prepend = 100.0 * np.median(e_prepend) # The results between numpy.median() and torch.median() are Inconsistant # when the length of the array/tensor is even. So this is a hack to avoid that. # issue: https://github.com/pytorch/pytorch/issues/1837 if np_y_preds.shape[0] % 2 == 0: assert pytest.approx(res) == np_res_prepend else: assert pytest.approx(res) == np_res