drop pytorch1.5 support

ppwwyyxx · facebook-github-bot · commit ab7055004d1b · 2021-02-25T15:13:21.000-08:00
Summary: pytorch 1.8 is coming

Reviewed By: theschnitz

Differential Revision: D25129974

fbshipit-source-id: d461766def65e3bd8f510fdfe469a575e102fb28
diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -217,10 +217,6 @@ workflows:
           torchvision_version: '0.8.1+cpu'
           context:
             - DOCKERHUB_TOKEN
-      - linux_gpu_tests:
-          name: linux_gpu_tests_pytorch1.5
-          pytorch_version: '1.5+cu101'
-          torchvision_version: '0.6+cu101'
       - linux_gpu_tests:
           name: linux_gpu_tests_pytorch1.6
           pytorch_version: '1.6+cu101'
diff --git a/.github/workflows/workflow.yml b/.github/workflows/workflow.yml
@@ -36,10 +36,8 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        torch: [1.5, 1.6, 1.7]
+        torch: [1.6, 1.7]
         include:
-          - torch: 1.5
-            torchvision: 0.6
           - torch: 1.6
             torchvision: 0.7
           - torch: 1.7
diff --git a/INSTALL.md b/INSTALL.md
@@ -2,7 +2,7 @@
 
 ### Requirements
 - Linux or macOS with Python ≥ 3.6
-- PyTorch ≥ 1.5 and [torchvision](https://github.com/pytorch/vision/) that matches the PyTorch installation.
+- PyTorch ≥ 1.6 and [torchvision](https://github.com/pytorch/vision/) that matches the PyTorch installation.
   Install them together at [pytorch.org](https://pytorch.org) to make sure of this
 - OpenCV is optional but needed by demo and visualization
 
diff --git a/detectron2/checkpoint/detection_checkpoint.py b/detectron2/checkpoint/detection_checkpoint.py
@@ -22,18 +22,7 @@ def __init__(self, model, save_dir="", *, save_to_disk=None, **checkpointables):
             save_to_disk=is_main_process if save_to_disk is None else save_to_disk,
             **checkpointables,
         )
-        if hasattr(self, "path_manager"):
-            self.path_manager = PathManager
-        else:
-            # This could only happen for open source
-            # TODO remove after upgrading fvcore version
-            from iopath.common.file_io import g_pathmgr
-
-            for handler in PathManager._path_handlers.values():
-                try:
-                    g_pathmgr.register_handler(handler)
-                except KeyError:
-                    pass
+        self.path_manager = PathManager
 
     def _load_file(self, filename):
         if filename.endswith(".pkl"):
diff --git a/detectron2/layers/wrappers.py b/detectron2/layers/wrappers.py
@@ -12,8 +12,6 @@
 import torch
 from torch.nn import functional as F
 
-from detectron2.utils.env import TORCH_VERSION
-
 
 def cat(tensors: List[torch.Tensor], dim: int = 0):
     """
@@ -96,33 +94,7 @@ def forward(self, x):
 ConvTranspose2d = torch.nn.ConvTranspose2d
 BatchNorm2d = torch.nn.BatchNorm2d
 interpolate = F.interpolate
-
-
-if TORCH_VERSION > (1, 5):
-    Linear = torch.nn.Linear
-else:
-
-    class Linear(torch.nn.Linear):
-        """
-        A wrapper around :class:`torch.nn.Linear` to support empty inputs and more features.
-        Because of https://github.com/pytorch/pytorch/issues/34202
-        """
-
-        def forward(self, x):
-            if x.numel() == 0:
-                output_shape = [x.shape[0], self.weight.shape[0]]
-
-                empty = _NewEmptyTensorOp.apply(x, output_shape)
-                if self.training:
-                    # This is to make DDP happy.
-                    # DDP expects all workers to have gradient w.r.t the same set of parameters.
-                    _dummy = sum(x.view(-1)[0] for x in self.parameters()) * 0.0
-                    return empty + _dummy
-                else:
-                    return empty
-
-            x = super().forward(x)
-            return x
+Linear = torch.nn.Linear
 
 
 def nonzero_tuple(x):
diff --git a/detectron2/modeling/roi_heads/box_head.py b/detectron2/modeling/roi_heads/box_head.py
@@ -6,7 +6,7 @@
 from torch import nn
 
 from detectron2.config import configurable
-from detectron2.layers import Conv2d, Linear, ShapeSpec, get_norm
+from detectron2.layers import Conv2d, ShapeSpec, get_norm
 from detectron2.utils.registry import Registry
 
 __all__ = ["FastRCNNConvFCHead", "build_box_head", "ROI_BOX_HEAD_REGISTRY"]
@@ -67,7 +67,7 @@ def __init__(
         for k, fc_dim in enumerate(fc_dims):
             if k == 0:
                 self.add_module("flatten", nn.Flatten())
-            fc = Linear(int(np.prod(self._output_size)), fc_dim)
+            fc = nn.Linear(int(np.prod(self._output_size)), fc_dim)
             self.add_module("fc{}".format(k + 1), fc)
             self.add_module("fc_relu{}".format(k + 1), nn.ReLU())
             self.fcs.append(fc)
diff --git a/detectron2/modeling/roi_heads/fast_rcnn.py b/detectron2/modeling/roi_heads/fast_rcnn.py
@@ -7,7 +7,7 @@
 from torch.nn import functional as F
 
 from detectron2.config import configurable
-from detectron2.layers import Linear, ShapeSpec, batched_nms, cat, cross_entropy, nonzero_tuple
+from detectron2.layers import ShapeSpec, batched_nms, cat, cross_entropy, nonzero_tuple
 from detectron2.modeling.box_regression import Box2BoxTransform
 from detectron2.structures import Boxes, Instances
 from detectron2.utils.events import get_event_storage
@@ -375,10 +375,10 @@ def __init__(
         self.num_classes = num_classes
         input_size = input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1)
         # prediction layer for num_classes foreground classes and one background class (hence + 1)
-        self.cls_score = Linear(input_size, num_classes + 1)
+        self.cls_score = nn.Linear(input_size, num_classes + 1)
         num_bbox_reg_classes = 1 if cls_agnostic_bbox_reg else num_classes
         box_dim = len(box2box_transform.weights)
-        self.bbox_pred = Linear(input_size, num_bbox_reg_classes * box_dim)
+        self.bbox_pred = nn.Linear(input_size, num_bbox_reg_classes * box_dim)
 
         nn.init.normal_(self.cls_score.weight, std=0.01)
         nn.init.normal_(self.bbox_pred.weight, std=0.001)
diff --git a/dev/packaging/build_all_wheels.sh b/dev/packaging/build_all_wheels.sh
@@ -56,9 +56,4 @@ else
   build_one cu101 1.6
   build_one cu92 1.6
   build_one cpu 1.6
-
-  build_one cu102 1.5
-  build_one cu101 1.5
-  build_one cu92 1.5
-  build_one cpu 1.5
 fi
diff --git a/docs/conf.py b/docs/conf.py
@@ -77,7 +77,7 @@ def resolve_any_xref(self, env, fromdocname, builder, target, node, contnode):
         "torchvision", "torchvision.ops",
     ]:
         sys.modules[m] = mock.Mock(name=m)
-    sys.modules['torch'].__version__ = "1.5"  # fake version
+    sys.modules['torch'].__version__ = "1.7"  # fake version
 else:
     try:
         torch.ops.detectron2 = mock.Mock(name="torch.ops.detectron2")
diff --git a/docs/requirements.txt b/docs/requirements.txt
@@ -16,5 +16,5 @@ future
 requests
 six
 git+git://github.com/facebookresearch/fvcore.git
-https://download.pytorch.org/whl/cpu/torch-1.5.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
-https://download.pytorch.org/whl/cpu/torchvision-0.6.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
+https://download.pytorch.org/whl/cpu/torch-1.7.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
+https://download.pytorch.org/whl/cpu/torchvision-0.8.1%2Bcpu-cp37-cp37m-linux_x86_64.whl
diff --git a/setup.py b/setup.py
@@ -12,7 +12,7 @@
 from torch.utils.hipify import hipify_python
 
 torch_ver = [int(x) for x in torch.__version__.split(".")[:2]]
-assert torch_ver >= [1, 5], "Requires PyTorch >= 1.5"
+assert torch_ver >= [1, 6], "Requires PyTorch >= 1.6"
 
 
 def get_version():
diff --git a/tests/layers/test_blocks.py b/tests/layers/test_blocks.py
@@ -5,7 +5,6 @@
 from torch import nn
 
 from detectron2.layers import ASPP, DepthwiseSeparableConv2d, FrozenBatchNorm2d
-from detectron2.utils.env import TORCH_VERSION
 
 
 """
@@ -22,7 +21,7 @@ def test_aspp(self):
         self.assertIsNot(m.convs[0].activation.weight, m.convs[1].activation.weight)
         self.assertIsNot(m.convs[0].activation.weight, m.project.activation.weight)
 
-    @unittest.skipIf(TORCH_VERSION < (1, 6) or not torch.cuda.is_available(), "CUDA not available")
+    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available")
     def test_frozen_batchnorm_fp16(self):
         from torch.cuda.amp import autocast
 
diff --git a/tests/layers/test_nms.py b/tests/layers/test_nms.py
@@ -4,7 +4,6 @@
 import torch
 
 from detectron2.layers import batched_nms
-from detectron2.utils.env import TORCH_VERSION
 from detectron2.utils.testing import random_boxes
 
 
@@ -14,7 +13,6 @@ def _create_tensors(self, N):
         scores = torch.rand(N)
         return boxes, scores
 
-    @unittest.skipIf(TORCH_VERSION < (1, 6), "Insufficient pytorch version")
     def test_nms_scriptability(self):
         N = 2000
         num_classes = 50
diff --git a/tests/modeling/test_anchor_generator.py b/tests/modeling/test_anchor_generator.py
@@ -6,7 +6,6 @@
 from detectron2.config import get_cfg
 from detectron2.layers import ShapeSpec
 from detectron2.modeling.anchor_generator import DefaultAnchorGenerator, RotatedAnchorGenerator
-from detectron2.utils.env import TORCH_VERSION
 
 logger = logging.getLogger(__name__)
 
@@ -40,7 +39,7 @@ def test_default_anchor_generator(self):
             ]
         )
 
-        assert torch.allclose(anchors[0].tensor, expected_anchor_tensor)
+        self.assertTrue(torch.allclose(anchors[0].tensor, expected_anchor_tensor))
 
     def test_default_anchor_generator_centered(self):
         # test explicit args
@@ -69,11 +68,10 @@ def test_default_anchor_generator_centered(self):
         )
 
         anchors = anchor_generator([features["stage3"]])
-        assert torch.allclose(anchors[0].tensor, expected_anchor_tensor)
+        self.assertTrue(torch.allclose(anchors[0].tensor, expected_anchor_tensor))
 
-        if TORCH_VERSION >= (1, 6):
-            anchors = torch.jit.script(anchor_generator)([features["stage3"]])
-            assert torch.allclose(anchors[0].tensor, expected_anchor_tensor)
+        anchors = torch.jit.script(anchor_generator)([features["stage3"]])
+        self.assertTrue(torch.allclose(anchors[0].tensor, expected_anchor_tensor))
 
     def test_rrpn_anchor_generator(self):
         cfg = get_cfg()
@@ -115,7 +113,7 @@ def test_rrpn_anchor_generator(self):
             ]
         )
 
-        assert torch.allclose(anchors[0].tensor, expected_anchor_tensor)
+        self.assertTrue(torch.allclose(anchors[0].tensor, expected_anchor_tensor))
 
 
 if __name__ == "__main__":
diff --git a/tests/modeling/test_fast_rcnn.py b/tests/modeling/test_fast_rcnn.py
@@ -136,7 +136,6 @@ def forward(self, proposal_deltas, proposal_boxes):
             o = func(torch.randn(20, 20), torch.randn(20, 4))
             self.assertEqual(o[0].shape, (20, 20))
 
-    @unittest.skipIf(TORCH_VERSION < (1, 6), "Insufficient pytorch version")
     def test_predict_probs_tracing(self):
         class Model(torch.nn.Module):
             def __init__(self, output_layer):
diff --git a/tests/modeling/test_matcher.py b/tests/modeling/test_matcher.py
@@ -5,11 +5,9 @@
 
 from detectron2.config import get_cfg
 from detectron2.modeling.matcher import Matcher
-from detectron2.utils.env import TORCH_VERSION
 
 
 class TestMatcher(unittest.TestCase):
-    @unittest.skipIf(TORCH_VERSION < (1, 6), "Insufficient pytorch version")
     def test_scriptability(self):
         cfg = get_cfg()
         anchor_matcher = Matcher(
diff --git a/tests/modeling/test_model_e2e.py b/tests/modeling/test_model_e2e.py
@@ -9,7 +9,6 @@
 import torch
 
 from detectron2.structures import BitMasks, Boxes, ImageList, Instances
-from detectron2.utils.env import TORCH_VERSION
 from detectron2.utils.events import EventStorage
 from detectron2.utils.testing import get_model_no_weights
 
@@ -155,9 +154,7 @@ def test_roiheads_inf_nan_data(self):
             det, _ = self.model.roi_heads(images, features, props)
             self.assertEqual(len(det[0]), 0)
 
-    @unittest.skipIf(
-        TORCH_VERSION < (1, 6) or not torch.cuda.is_available(), "Insufficient pytorch version"
-    )
+    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available")
     def test_autocast(self):
         from torch.cuda.amp import autocast
 
@@ -200,9 +197,7 @@ def test_inf_nan_data(self):
             if len(det[0]):
                 self.assertTrue(torch.isfinite(det[0].pred_boxes.tensor).sum() == 0)
 
-    @unittest.skipIf(
-        TORCH_VERSION < (1, 6) or not torch.cuda.is_available(), "Insufficient pytorch version"
-    )
+    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available")
     def test_autocast(self):
         from torch.cuda.amp import autocast
 
diff --git a/tests/modeling/test_roi_pooler.py b/tests/modeling/test_roi_pooler.py
@@ -117,7 +117,6 @@ def test_no_images(self):
         output = pooler.forward(features, [])
         self.assertEqual(output.shape, (0, C, 14, 14))
 
-    @unittest.skipIf(TORCH_VERSION < (1, 6), "Insufficient pytorch version")
     def test_fmt_box_list_tracing(self):
         class Model(torch.nn.Module):
             def forward(self, box_tensor):
@@ -130,7 +129,6 @@ def forward(self, box_tensor):
             self.assertEqual(func(torch.ones(5, 4)).shape, (5, 5))
             self.assertEqual(func(torch.ones(20, 4)).shape, (20, 5))
 
-    @unittest.skipIf(TORCH_VERSION < (1, 6), "Insufficient pytorch version")
     def test_roi_pooler_tracing(self):
         class Model(torch.nn.Module):
             def __init__(self, roi):
