Explore fixing test issues for older pytorch versions

.github/workflows/pytorch-version-tests.yml

@@ -15,14 +15,13 @@ jobs:
       max-parallel: 5
       fail-fast: false
       matrix:
-        python-version: [3.9, "3.10", "3.11"]
+        python-version: ["3.9", "3.10", "3.11"]
         pytorch-version: [2.5.1, 2.4.1, 2.3.1, 2.2.2, 1.13.1, 1.12.1, 1.10.0]
         exclude:
           - pytorch-version: 1.10.0
             python-version: "3.10"
           - pytorch-version: 1.10.0
             python-version: "3.11"
-
           - pytorch-version: 1.11.0
             python-version: "3.10"
           - pytorch-version: 1.11.0
@@ -68,12 +67,13 @@ jobs:
       - name: Install dependencies
         shell: bash -l {0}
         run: |
-          conda install pytorch=${{ matrix.pytorch-version }} torchvision cpuonly python=${{ matrix.python-version }} -c pytorch
+
+          conda install pytorch=${{ matrix.pytorch-version }} torchvision cpuonly python=${{ matrix.python-version }} -c pytorch -y
 
           # We should install numpy<2.0 for pytorch<2.3
           numpy_one_pth_version=$(python -c "import torch; print(float('.'.join(torch.__version__.split('.')[:2])) < 2.3)")
           if [ "${numpy_one_pth_version}" == "True" ]; then
-            pip install -U "numpy<2.0"
+            pip install "numpy<2.0"
           fi
 
           pip install -r requirements-dev.txt
@@ -83,7 +83,7 @@ jobs:
           # which raises the error: AttributeError: module 'distutils' has no attribute 'version' for setuptools>59
           bad_pth_version=$(python -c "import torch; print('.'.join(torch.__version__.split('.')[:2]) in ['1.9', '1.10'])")
           if [ "${bad_pth_version}" == "True" ]; then
-            pip install --upgrade "setuptools<59"
+            pip install "setuptools<59"
             python -c "from setuptools import distutils; distutils.version.LooseVersion"
           fi
 

ignite/metrics/gan/fid.py

@@ -37,7 +37,7 @@ def fid_score(
     diff = mu1 - mu2
 
     # Product might be almost singular
-    covmean, _ = scipy.linalg.sqrtm(sigma1.mm(sigma2), disp=False)
+    covmean, _ = scipy.linalg.sqrtm(sigma1.mm(sigma2).numpy(), disp=False)
     # Numerical error might give slight imaginary component
     if np.iscomplexobj(covmean):
         if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
@@ -48,7 +48,7 @@ def fid_score(
     tr_covmean = np.trace(covmean)
 
     if not np.isfinite(covmean).all():
-        tr_covmean = np.sum(np.sqrt(((np.diag(sigma1) * eps) * (np.diag(sigma2) * eps)) / (eps * eps)))
+        tr_covmean = np.sum(np.sqrt(((np.diag(sigma1.numpy()) * eps) * (np.diag(sigma2.numpy()) * eps)) / (eps * eps)))
 
     return float(diff.dot(diff).item() + torch.trace(sigma1) + torch.trace(sigma2) - 2 * tr_covmean)
 

ignite/metrics/mean_average_precision.py

@@ -2,6 +2,7 @@
 from typing import Callable, cast, List, Optional, Sequence, Tuple, Union
 
 import torch
+from packaging.version import Version
 from typing_extensions import Literal
 
 import ignite.distributed as idist
@@ -11,6 +12,9 @@
 from ignite.utils import to_onehot
 
 
+_torch_version_lt_113 = Version(torch.__version__) < Version("1.13.0")
+
+
 class _BaseAveragePrecision:
     def __init__(
         self,
@@ -97,9 +101,12 @@ def _compute_average_precision(self, recall: torch.Tensor, precision: torch.Tens
         if self.rec_thresholds is not None:
             rec_thresholds = self.rec_thresholds.repeat((*recall.shape[:-1], 1))
             rec_thresh_indices = torch.searchsorted(recall, rec_thresholds)
-            precision = precision.take_along_dim(
-                rec_thresh_indices.where(rec_thresh_indices != recall.size(-1), 0), dim=-1
-            ).where(rec_thresh_indices != recall.size(-1), 0)
+            rec_mask = rec_thresh_indices != recall.size(-1)
+            precision = torch.where(
+                rec_mask,
+                precision.take_along_dim(torch.where(rec_mask, rec_thresh_indices, 0), dim=-1),
+                0.0,
+            )
             recall = rec_thresholds
         recall_differential = recall.diff(
             dim=-1, prepend=torch.zeros((*recall.shape[:-1], 1), device=recall.device, dtype=recall.dtype)
@@ -335,9 +342,10 @@ def _compute_recall_and_precision(
         Returns:
             `(recall, precision)`
         """
-        indices = torch.argsort(y_pred, stable=True, descending=True)
+        kwargs = {} if _torch_version_lt_113 else {"stable": True}
+        indices = torch.argsort(y_pred, descending=True, **kwargs)
         tp_summation = y_true[indices].cumsum(dim=0)
-        if tp_summation.device != torch.device("mps"):
+        if tp_summation.device.type != "mps":
             tp_summation = tp_summation.double()
 
         # Adopted from Scikit-learn's implementation
@@ -354,7 +362,7 @@ def _compute_recall_and_precision(
             recall = tp_summation / y_true_positive_count
 
         predicted_positive = tp_summation + fp_summation
-        precision = tp_summation / torch.where(predicted_positive == 0, 1, predicted_positive)
+        precision = tp_summation / torch.where(predicted_positive == 0, 1.0, predicted_positive)
         return recall, precision
 
     def compute(self) -> Union[torch.Tensor, float]:
@@ -371,7 +379,7 @@ def compute(self) -> Union[torch.Tensor, float]:
             torch.long if self._type == "multiclass" else torch.uint8,
             self._device,
         )
-        fp_precision = torch.double if self._device != torch.device("mps") else torch.float32
+        fp_precision = torch.double if self._device.type != "mps" else torch.float32
         y_pred = _cat_and_agg_tensors(self._y_pred, (num_classes,), fp_precision, self._device)
 
         if self._type == "multiclass":

ignite/metrics/vision/object_detection_average_precision_recall.py

@@ -1,6 +1,7 @@
 from typing import Callable, cast, Dict, List, Optional, Sequence, Tuple, Union
 
 import torch
+from packaging.version import Version
 from typing_extensions import Literal
 
 from ignite.metrics import MetricGroup
@@ -9,6 +10,9 @@
 from ignite.metrics.metric import Metric, reinit__is_reduced, sync_all_reduce
 
 
+_torch_version_lt_113 = Version(torch.__version__) < Version("1.13.0")
+
+
 def coco_tensor_list_to_dict_list(
     output: Tuple[
         Union[List[torch.Tensor], List[Dict[str, torch.Tensor]]],
@@ -213,7 +217,8 @@ def _compute_recall_and_precision(
         Returns:
             `(recall, precision)`
         """
-        indices = torch.argsort(scores, dim=-1, stable=True, descending=True)
+        kwargs = {} if _torch_version_lt_113 else {"stable": True}
+        indices = torch.argsort(scores, descending=True, **kwargs)
         tp = TP[..., indices]
         tp_summation = tp.cumsum(dim=-1)
         if tp_summation.device.type != "mps":
@@ -226,7 +231,7 @@ def _compute_recall_and_precision(
 
         recall = tp_summation / y_true_count
         predicted_positive = tp_summation + fp_summation
-        precision = tp_summation / torch.where(predicted_positive == 0, 1, predicted_positive)
+        precision = tp_summation / torch.where(predicted_positive == 0, 1.0, predicted_positive)
 
         return recall, precision
 
@@ -258,9 +263,12 @@ def _compute_average_precision(self, recall: torch.Tensor, precision: torch.Tens
             if recall.size(-1) != 0
             else torch.LongTensor([], device=self._device)
         )
-        precision_integrand = precision_integrand.take_along_dim(
-            rec_thresh_indices.where(rec_thresh_indices != recall.size(-1), 0), dim=-1
-        ).where(rec_thresh_indices != recall.size(-1), 0)
+        recall_mask = rec_thresh_indices != recall.size(-1)
+        precision_integrand = torch.where(
+            recall_mask,
+            precision_integrand.take_along_dim(torch.where(recall_mask, rec_thresh_indices, 0), dim=-1),
+            0.0,
+        )
         return torch.sum(precision_integrand, dim=-1) / len(cast(torch.Tensor, self.rec_thresholds))
 
     @reinit__is_reduced
@@ -298,6 +306,7 @@ def update(self, output: Tuple[List[Dict[str, torch.Tensor]], List[Dict[str, tor
                                             This key is optional.
                 ========= ================= =================================================
         """
+        kwargs = {} if _torch_version_lt_113 else {"stable": True}
         self._check_matching_input(output)
         for pred, target in zip(*output):
             labels = target["labels"]
@@ -312,7 +321,7 @@ def update(self, output: Tuple[List[Dict[str, torch.Tensor]], List[Dict[str, tor
 
             # Matching logic of object detection mAP, according to COCO reference implementation.
             if len(pred["labels"]):
-                best_detections_index = torch.argsort(pred["scores"], stable=True, descending=True)
+                best_detections_index = torch.argsort(pred["scores"], descending=True, **kwargs)
                 max_best_detections_index = torch.cat(
                     [
                         best_detections_index[pred["labels"][best_detections_index] == c][

tests/ignite/distributed/utils/test_native.py

@@ -25,6 +25,9 @@
 )
 
 
+_torch_version_lt_1132 = Version(torch.__version__) < Version("1.13.2")
+
+
 def _test_native_distrib_single_node_launch_tool(backend, device, local_rank, world_size, init_method=None, **kwargs):
     import os
 
@@ -230,7 +233,9 @@ def test_idist_all_reduce_nccl(distributed_context_single_node_nccl):
 
 
 @pytest.mark.distributed
+@pytest.mark.order(-1)
 @pytest.mark.skipif(not has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.skipif(_torch_version_lt_1132, reason="Skip if older pytorch version")
 def test_idist_all_reduce_gloo(distributed_context_single_node_gloo):
     device = idist.device()
     _test_distrib_all_reduce(device)
@@ -252,6 +257,7 @@ def test_idist_all_gather_nccl(distributed_context_single_node_nccl):
 @pytest.mark.distributed
 @pytest.mark.skipif(not has_native_dist_support, reason="Skip if no native dist support")
 @pytest.mark.skipif(Version(torch.__version__) < Version("1.7.0"), reason="dist.all_gather_object is not implemented")
+@pytest.mark.order(-3)
 def test_idist_all_gather_gloo(distributed_context_single_node_gloo):
     device = idist.device()
     _test_distrib_all_gather(device)
@@ -271,6 +277,7 @@ def test_idist_all_gather_tensors_with_shapes_nccl(distributed_context_single_no
 
 @pytest.mark.distributed
 @pytest.mark.skipif(not has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.order(-2)
 def test_idist_all_gather_tensors_with_shapes_gloo(distributed_context_single_node_gloo):
     device = idist.device()
     _test_idist_all_gather_tensors_with_shapes(device)

tests/ignite/metrics/test_mean_average_precision.py

@@ -45,30 +45,30 @@ def test__prepare_output():
     metric = MeanAveragePrecision()
 
     metric._type = "binary"
-    scores, y = metric._prepare_output((torch.rand((5, 4, 3, 2)), torch.randint(0, 2, (5, 4, 3, 2)).bool()))
+    scores, y = metric._prepare_output((torch.rand((5, 4, 3, 2)), torch.randint(0, 2, (5, 4, 3, 2))))
     assert scores.shape == y.shape == (1, 120)
 
     metric._type = "multiclass"
     scores, y = metric._prepare_output((torch.rand((5, 4, 3, 2)), torch.randint(0, 4, (5, 3, 2))))
     assert scores.shape == (4, 30) and y.shape == (30,)
 
     metric._type = "multilabel"
-    scores, y = metric._prepare_output((torch.rand((5, 4, 3, 2)), torch.randint(0, 2, (5, 4, 3, 2)).bool()))
+    scores, y = metric._prepare_output((torch.rand((5, 4, 3, 2)), torch.randint(0, 2, (5, 4, 3, 2))))
     assert scores.shape == y.shape == (4, 30)
 
 
 def test_update():
     metric = MeanAveragePrecision()
     assert len(metric._y_pred) == len(metric._y_true) == 0
-    metric.update((torch.rand((5, 4)), torch.randint(0, 2, (5, 4)).bool()))
+    metric.update((torch.rand((5, 4)), torch.randint(0, 2, (5, 4))))
     assert len(metric._y_pred) == len(metric._y_true) == 1
 
 
 def test__compute_recall_and_precision():
     m = MeanAveragePrecision()
 
     scores = torch.rand((50,))
-    y_true = torch.randint(0, 2, (50,)).bool()
+    y_true = torch.randint(0, 2, (50,))
     precision, recall, _ = precision_recall_curve(y_true.numpy(), scores.numpy())
     P = y_true.sum(dim=-1)
     ignite_recall, ignite_precision = m._compute_recall_and_precision(y_true, scores, P)
@@ -77,7 +77,7 @@ def test__compute_recall_and_precision():
 
     # When there's no actual positive. Numpy expectedly raises warning.
     scores = torch.rand((50,))
-    y_true = torch.zeros((50,)).bool()
+    y_true = torch.zeros((50,))
     precision, recall, _ = precision_recall_curve(y_true.numpy(), scores.numpy())
     P = torch.tensor(0)
     ignite_recall, ignite_precision = m._compute_recall_and_precision(y_true, scores, P)
@@ -147,7 +147,7 @@ def test_compute_nonbinary_data(class_mean):
 
     # Multilabel
     m = MeanAveragePrecision(is_multilabel=True, class_mean=class_mean)
-    y_true = torch.randint(0, 2, (130, 5, 2, 2)).bool()
+    y_true = torch.randint(0, 2, (130, 5, 2, 2))
     m.update((scores[:50], y_true[:50]))
     m.update((scores[50:], y_true[50:]))
     ignite_map = m.compute().numpy()

tests/ignite/metrics/vision/test_object_detection_map.py

@@ -872,7 +872,7 @@ def test__compute_recall_and_precision(available_device):
 def test_compute(sample):
     device = idist.device()
 
-    if device == torch.device("mps"):
+    if device.type == "mps":
         pytest.skip("Due to MPS backend out of memory")
 
     # [email protected], [email protected], [email protected], AP-S, AP-M, AP-L, AR-1, AR-10, AR-100, AR-S, AR-M, AR-L
@@ -932,7 +932,7 @@ def test_integration(sample):
     bs = 3
 
     device = idist.device()
-    if device == torch.device("mps"):
+    if device.type == "mps":
         pytest.skip("Due to MPS backend out of memory")
 
     def update(engine, i):
@@ -1003,7 +1003,7 @@ def test_distrib_update_compute(distributed, sample):
 
     device = idist.device()
 
-    if device == torch.device("mps"):
+    if device.type == "mps":
         pytest.skip("Due to MPS backend out of memory")
 
     metric_device = "cpu" if device.type == "xla" else device