adds available_device to test_precision_recall_curve #3335 (#3368)

* adds available_device to test_precision_recall_curve #3335

* forces float32 when converting to tensor on mps

* creates the data directly with torch tensors instead of numpy arrays

* ensures compatibility with MPS by converting to float32

* comments on float32 conversion

* makes sure that sklearn does not convert float32 to float64
(otherwize type error on on MPS)

* another attempt of avoiding float64

* avoiding float64 for MPS

* avoiding float64 for MPS

* another attempt at avoiding float64 on MPS

* moves conversion to float32 before assertions

* conversion to float32

* more conversion to float32

* more conversion to float32

* more conversion to float32

* more conversion to float32

* in precision_recall_curve.py add dtype when creating tensors for precision, recall and thresholds

* removes unnecessary conversions

* move tensors to CPU before passing them to precision_recall_curve

* move tensors to CPU before passing them to precision_recall_curve

* move tensors to CPU before passing them to precision_recall_curve

* replace np.testing.assert_array_almost_equal with pytest.approx

* removes manual_seed

---------

Co-authored-by: vfdev <[email protected]>

Author: BanzaiTokyo

ignite/metrics/precision_recall_curve.py

@@ -110,11 +110,11 @@ def compute(self) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:  # type: i
             if idist.get_rank() == 0:
                 # Run compute_fn on zero rank only
                 precision, recall, thresholds = cast(Tuple, self.compute_fn(_prediction_tensor, _target_tensor))
-                precision = torch.tensor(precision, device=_prediction_tensor.device)
-                recall = torch.tensor(recall, device=_prediction_tensor.device)
+                precision = torch.tensor(precision, device=_prediction_tensor.device, dtype=self._double_dtype)
+                recall = torch.tensor(recall, device=_prediction_tensor.device, dtype=self._double_dtype)
                 # thresholds can have negative strides, not compatible with torch tensors
                 # https://discuss.pytorch.org/t/negative-strides-in-tensor-error/134287/2
-                thresholds = torch.tensor(thresholds.copy(), device=_prediction_tensor.device)
+                thresholds = torch.tensor(thresholds.copy(), device=_prediction_tensor.device, dtype=self._double_dtype)
             else:
                 precision, recall, thresholds = None, None, None
 

tests/ignite/metrics/test_precision_recall_curve.py

@@ -2,7 +2,6 @@
 from typing import Tuple
 from unittest.mock import patch
 
-import numpy as np
 import pytest
 import sklearn
 import torch
@@ -28,85 +27,97 @@ def test_no_sklearn(mock_no_sklearn):
         pr_curve.compute()
 
 
-def test_precision_recall_curve():
+def test_precision_recall_curve(available_device):
     size = 100
-    np_y_pred = np.random.rand(size, 1)
-    np_y = np.zeros((size,))
-    np_y[size // 2 :] = 1
-    sk_precision, sk_recall, sk_thresholds = precision_recall_curve(np_y, np_y_pred)
+    y_pred = torch.rand(size, 1, dtype=torch.float32, device=available_device)
+    y_true = torch.zeros(size, dtype=torch.float32, device=available_device)
+    y_true[size // 2 :] = 1.0
+    expected_precision, expected_recall, expected_thresholds = precision_recall_curve(
+        y_true.cpu().numpy(), y_pred.cpu().numpy()
+    )
 
-    precision_recall_curve_metric = PrecisionRecallCurve()
-    y_pred = torch.from_numpy(np_y_pred)
-    y = torch.from_numpy(np_y)
+    precision_recall_curve_metric = PrecisionRecallCurve(device=available_device)
+    assert precision_recall_curve_metric._device == torch.device(available_device)
 
-    precision_recall_curve_metric.update((y_pred, y))
+    precision_recall_curve_metric.update((y_pred, y_true))
     precision, recall, thresholds = precision_recall_curve_metric.compute()
-    precision = precision.numpy()
-    recall = recall.numpy()
-    thresholds = thresholds.numpy()
 
-    assert pytest.approx(precision) == sk_precision
-    assert pytest.approx(recall) == sk_recall
-    # assert thresholds almost equal, due to numpy->torch->numpy conversion
-    np.testing.assert_array_almost_equal(thresholds, sk_thresholds)
+    precision = precision.cpu().numpy()
+    recall = recall.cpu().numpy()
+    thresholds = thresholds.cpu().numpy()
+
+    assert pytest.approx(precision) == expected_precision
+    assert pytest.approx(recall) == expected_recall
+    assert thresholds == pytest.approx(expected_thresholds, rel=1e-6)
 
 
-def test_integration_precision_recall_curve_with_output_transform():
-    np.random.seed(1)
+def test_integration_precision_recall_curve_with_output_transform(available_device):
     size = 100
-    np_y_pred = np.random.rand(size, 1)
-    np_y = np.zeros((size,))
-    np_y[size // 2 :] = 1
-    np.random.shuffle(np_y)
+    y_pred = torch.rand(size, 1, dtype=torch.float32, device=available_device)
+    y_true = torch.zeros(size, dtype=torch.float32, device=available_device)
+    y_true[size // 2 :] = 1.0
+    perm = torch.randperm(size)
+    y_pred = y_pred[perm]
+    y_true = y_true[perm]
 
-    sk_precision, sk_recall, sk_thresholds = precision_recall_curve(np_y, np_y_pred)
+    expected_precision, expected_recall, expected_thresholds = precision_recall_curve(
+        y_true.cpu().numpy(), y_pred.cpu().numpy()
+    )
 
     batch_size = 10
 
     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)
+        y_true_batch = y_true[idx : idx + batch_size]
+        y_pred_batch = y_pred[idx : idx + batch_size]
+        return idx, y_pred_batch, y_true_batch
 
     engine = Engine(update_fn)
 
-    precision_recall_curve_metric = PrecisionRecallCurve(output_transform=lambda x: (x[1], x[2]))
+    precision_recall_curve_metric = PrecisionRecallCurve(
+        output_transform=lambda x: (x[1], x[2]), device=available_device
+    )
+    assert precision_recall_curve_metric._device == torch.device(available_device)
     precision_recall_curve_metric.attach(engine, "precision_recall_curve")
 
     data = list(range(size // batch_size))
     precision, recall, thresholds = engine.run(data, max_epochs=1).metrics["precision_recall_curve"]
-    precision = precision.numpy()
-    recall = recall.numpy()
-    thresholds = thresholds.numpy()
-    assert pytest.approx(precision) == sk_precision
-    assert pytest.approx(recall) == sk_recall
-    # assert thresholds almost equal, due to numpy->torch->numpy conversion
-    np.testing.assert_array_almost_equal(thresholds, sk_thresholds)
+    precision = precision.cpu().numpy()
+    recall = recall.cpu().numpy()
+    thresholds = thresholds.cpu().numpy()
+    assert pytest.approx(precision) == expected_precision
+    assert pytest.approx(recall) == expected_recall
+    assert thresholds == pytest.approx(expected_thresholds, rel=1e-6)
 
 
-def test_integration_precision_recall_curve_with_activated_output_transform():
-    np.random.seed(1)
+def test_integration_precision_recall_curve_with_activated_output_transform(available_device):
     size = 100
-    np_y_pred = np.random.rand(size, 1)
-    np_y_pred_sigmoid = torch.sigmoid(torch.from_numpy(np_y_pred)).numpy()
-    np_y = np.zeros((size,))
-    np_y[size // 2 :] = 1
-    np.random.shuffle(np_y)
-
-    sk_precision, sk_recall, sk_thresholds = precision_recall_curve(np_y, np_y_pred_sigmoid)
+    y_pred = torch.rand(size, 1, dtype=torch.float32, device=available_device)
+    y_true = torch.zeros(size, dtype=torch.float32, device=available_device)
+    y_true[size // 2 :] = 1.0
+    perm = torch.randperm(size)
+    y_pred = y_pred[perm]
+    y_true = y_true[perm]
+
+    sigmoid_y_pred = torch.sigmoid(y_pred).cpu().numpy()
+    expected_precision, expected_recall, expected_thresholds = precision_recall_curve(
+        y_true.cpu().numpy(), sigmoid_y_pred
+    )
 
     batch_size = 10
 
     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)
+        y_true_batch = y_true[idx : idx + batch_size]
+        y_pred_batch = y_pred[idx : idx + batch_size]
+        return idx, y_pred_batch, y_true_batch
 
     engine = Engine(update_fn)
 
-    precision_recall_curve_metric = PrecisionRecallCurve(output_transform=lambda x: (torch.sigmoid(x[1]), x[2]))
+    precision_recall_curve_metric = PrecisionRecallCurve(
+        output_transform=lambda x: (torch.sigmoid(x[1]), x[2]), device=available_device
+    )
+    assert precision_recall_curve_metric._device == torch.device(available_device)
     precision_recall_curve_metric.attach(engine, "precision_recall_curve")
 
     data = list(range(size // batch_size))
@@ -115,25 +126,26 @@ def update_fn(engine, batch):
     recall = recall.cpu().numpy()
     thresholds = thresholds.cpu().numpy()
 
-    assert pytest.approx(precision) == sk_precision
-    assert pytest.approx(recall) == sk_recall
-    # assert thresholds almost equal, due to numpy->torch->numpy conversion
-    np.testing.assert_array_almost_equal(thresholds, sk_thresholds)
+    assert pytest.approx(precision) == expected_precision
+    assert pytest.approx(recall) == expected_recall
+    assert thresholds == pytest.approx(expected_thresholds, rel=1e-6)
 
 
-def test_check_compute_fn():
+def test_check_compute_fn(available_device):
     y_pred = torch.zeros((8, 13))
     y_pred[:, 1] = 1
     y_true = torch.zeros_like(y_pred)
     output = (y_pred, y_true)
 
-    em = PrecisionRecallCurve(check_compute_fn=True)
+    em = PrecisionRecallCurve(check_compute_fn=True, device=available_device)
+    assert em._device == torch.device(available_device)
 
     em.reset()
     with pytest.warns(EpochMetricWarning, match=r"Probably, there can be a problem with `compute_fn`"):
         em.update(output)
 
-    em = PrecisionRecallCurve(check_compute_fn=False)
+    em = PrecisionRecallCurve(check_compute_fn=False, device=available_device)
+    assert em._device == torch.device(available_device)
     em.update(output)
 
 
@@ -225,14 +237,14 @@ def update(engine, i):
         np_y_true = y_true.cpu().numpy().ravel()
         np_y_preds = y_preds.cpu().numpy().ravel()
 
-        sk_precision, sk_recall, sk_thresholds = precision_recall_curve(np_y_true, np_y_preds)
+        expected_precision, expected_recall, expected_thresholds = precision_recall_curve(np_y_true, np_y_preds)
 
-        assert precision.shape == sk_precision.shape
-        assert recall.shape == sk_recall.shape
-        assert thresholds.shape == sk_thresholds.shape
-        assert pytest.approx(precision.cpu().numpy()) == sk_precision
-        assert pytest.approx(recall.cpu().numpy()) == sk_recall
-        assert pytest.approx(thresholds.cpu().numpy()) == sk_thresholds
+        assert precision.shape == expected_precision.shape
+        assert recall.shape == expected_recall.shape
+        assert thresholds.shape == expected_thresholds.shape
+        assert pytest.approx(precision.cpu().numpy()) == expected_precision
+        assert pytest.approx(recall.cpu().numpy()) == expected_recall
+        assert pytest.approx(thresholds.cpu().numpy()) == expected_thresholds
 
     metric_devices = ["cpu"]
     if device.type != "xla":