Support Exporting RPN to ONNX (#1329)

* Support Exporting RPN to ONNX

* address PR comments

* fix cat

* add flatten

* replace cat by stack

* update test to run only on rpn module

* use tolerate_small_mismatch

Author: lara-hdr

test/test_onnx.py

@@ -1,7 +1,10 @@
 import io
 import torch
 from torchvision import ops
+from torchvision.models.detection.image_list import ImageList
 from torchvision.models.detection.transform import GeneralizedRCNNTransform
+from torchvision.models.detection.rpn import AnchorGenerator, RPNHead, RegionProposalNetwork
+from torchvision.models.detection.backbone_utils import resnet_fpn_backbone
 
 from collections import OrderedDict
 
@@ -20,7 +23,7 @@ class ONNXExporterTester(unittest.TestCase):
     def setUpClass(cls):
         torch.manual_seed(123)
 
-    def run_model(self, model, inputs_list):
+    def run_model(self, model, inputs_list, tolerate_small_mismatch=False):
         model.eval()
 
         onnx_io = io.BytesIO()
@@ -36,9 +39,9 @@ def run_model(self, model, inputs_list):
                 test_ouputs = model(*test_inputs)
                 if isinstance(test_ouputs, torch.Tensor):
                     test_ouputs = (test_ouputs,)
-            self.ort_validate(onnx_io, test_inputs, test_ouputs)
+            self.ort_validate(onnx_io, test_inputs, test_ouputs, tolerate_small_mismatch)
 
-    def ort_validate(self, onnx_io, inputs, outputs):
+    def ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismatch=False):
 
         inputs, _ = torch.jit._flatten(inputs)
         outputs, _ = torch.jit._flatten(outputs)
@@ -58,7 +61,13 @@ def to_numpy(tensor):
         ort_outs = ort_session.run(None, ort_inputs)
 
         for i in range(0, len(outputs)):
-            torch.testing.assert_allclose(outputs[i], ort_outs[i], rtol=1e-03, atol=1e-05)
+            try:
+                torch.testing.assert_allclose(outputs[i], ort_outs[i], rtol=1e-03, atol=1e-05)
+            except AssertionError as error:
+                if tolerate_small_mismatch:
+                    assert ("(0.00%)" in str(error)), str(error)
+                else:
+                    assert False, str(error)
 
     def test_nms(self):
         boxes = torch.rand(5, 4)
@@ -91,11 +100,7 @@ def test_transform_images(self):
         class TransformModule(torch.nn.Module):
             def __init__(self_module):
                 super(TransformModule, self_module).__init__()
-                min_size = 800
-                max_size = 1333
-                image_mean = [0.485, 0.456, 0.406]
-                image_std = [0.229, 0.224, 0.225]
-                self_module.transform = GeneralizedRCNNTransform(min_size, max_size, image_mean, image_std)
+                self_module.transform = self._init_test_generalized_rcnn_transform()
 
             def forward(self_module, images):
                 return self_module.transform(images)[0].tensors
@@ -104,6 +109,66 @@ def forward(self_module, images):
         input_test = [torch.rand(3, 800, 1280), torch.rand(3, 800, 800)]
         self.run_model(TransformModule(), [input, input_test])
 
+    def _init_test_generalized_rcnn_transform(self):
+        min_size = 800
+        max_size = 1333
+        image_mean = [0.485, 0.456, 0.406]
+        image_std = [0.229, 0.224, 0.225]
+        transform = GeneralizedRCNNTransform(min_size, max_size, image_mean, image_std)
+        return transform
+
+    def _init_test_rpn(self):
+        anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
+        aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
+        rpn_anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)
+        out_channels = 256
+        rpn_head = RPNHead(out_channels, rpn_anchor_generator.num_anchors_per_location()[0])
+        rpn_fg_iou_thresh = 0.7
+        rpn_bg_iou_thresh = 0.3
+        rpn_batch_size_per_image = 256
+        rpn_positive_fraction = 0.5
+        rpn_pre_nms_top_n = dict(training=2000, testing=1000)
+        rpn_post_nms_top_n = dict(training=2000, testing=1000)
+        rpn_nms_thresh = 0.7
+
+        rpn = RegionProposalNetwork(
+            rpn_anchor_generator, rpn_head,
+            rpn_fg_iou_thresh, rpn_bg_iou_thresh,
+            rpn_batch_size_per_image, rpn_positive_fraction,
+            rpn_pre_nms_top_n, rpn_post_nms_top_n, rpn_nms_thresh)
+        return rpn
+
+    def test_rpn(self):
+        class RPNModule(torch.nn.Module):
+            def __init__(self_module, images):
+                super(RPNModule, self_module).__init__()
+                self_module.rpn = self._init_test_rpn()
+                self_module.images = ImageList(images, [i.shape[-2:] for i in images])
+
+            def forward(self_module, features):
+                return self_module.rpn(self_module.images, features)
+
+        def get_features(images):
+            s0, s1 = images.shape[-2:]
+            features = [
+                ('0', torch.rand(2, 256, s0 // 4, s1 // 4)),
+                ('1', torch.rand(2, 256, s0 // 8, s1 // 8)),
+                ('2', torch.rand(2, 256, s0 // 16, s1 // 16)),
+                ('3', torch.rand(2, 256, s0 // 32, s1 // 32)),
+                ('4', torch.rand(2, 256, s0 // 64, s1 // 64)),
+            ]
+            features = OrderedDict(features)
+            return features
+
+        images = torch.rand(2, 3, 600, 600)
+        features = get_features(images)
+        test_features = get_features(images)
+
+        model = RPNModule(images)
+        model.eval()
+        model(features)
+        self.run_model(model, [(features,), (test_features,)], tolerate_small_mismatch=True)
+
     def test_multi_scale_roi_align(self):
 
         class TransformModule(torch.nn.Module):

torchvision/models/detection/_utils.py

@@ -3,6 +3,7 @@
 import math
 
 import torch
+import torchvision
 
 
 class BalancedPositiveNegativeSampler(object):
@@ -162,7 +163,7 @@ def decode(self, rel_codes, boxes):
         if isinstance(rel_codes, (list, tuple)):
             rel_codes = torch.cat(rel_codes, dim=0)
         assert isinstance(rel_codes, torch.Tensor)
-        boxes_per_image = [len(b) for b in boxes]
+        boxes_per_image = [b.size(0) for b in boxes]
         concat_boxes = torch.cat(boxes, dim=0)
         pred_boxes = self.decode_single(
             rel_codes.reshape(sum(boxes_per_image), -1), concat_boxes
@@ -201,16 +202,11 @@ def decode_single(self, rel_codes, boxes):
         pred_w = torch.exp(dw) * widths[:, None]
         pred_h = torch.exp(dh) * heights[:, None]
 
-        pred_boxes = torch.zeros_like(rel_codes)
-        # x1
-        pred_boxes[:, 0::4] = pred_ctr_x - 0.5 * pred_w
-        # y1
-        pred_boxes[:, 1::4] = pred_ctr_y - 0.5 * pred_h
-        # x2
-        pred_boxes[:, 2::4] = pred_ctr_x + 0.5 * pred_w
-        # y2
-        pred_boxes[:, 3::4] = pred_ctr_y + 0.5 * pred_h
-
+        pred_boxes1 = pred_ctr_x - torch.tensor(0.5, dtype=pred_ctr_x.dtype) * pred_w
+        pred_boxes2 = pred_ctr_y - torch.tensor(0.5, dtype=pred_ctr_y.dtype) * pred_h
+        pred_boxes3 = pred_ctr_x + torch.tensor(0.5, dtype=pred_ctr_x.dtype) * pred_w
+        pred_boxes4 = pred_ctr_y + torch.tensor(0.5, dtype=pred_ctr_y.dtype) * pred_h
+        pred_boxes = torch.stack((pred_boxes1, pred_boxes2, pred_boxes3, pred_boxes4), dim=2).flatten(1)
         return pred_boxes
 
 

torchvision/models/detection/rpn.py

@@ -3,11 +3,25 @@
 from torch.nn import functional as F
 from torch import nn
 
+import torchvision
 from torchvision.ops import boxes as box_ops
 
 from . import _utils as det_utils
 
 
+@torch.jit.unused
+def _onnx_get_num_anchors_and_pre_nms_top_n(ob, orig_pre_nms_top_n):
+    from torch.onnx import operators
+    num_anchors = operators.shape_as_tensor(ob)[1].unsqueeze(0)
+    # TODO : remove cast to IntTensor/num_anchors.dtype when
+    #        ONNX Runtime version is updated with ReduceMin int64 support
+    pre_nms_top_n = torch.min(torch.cat(
+        (torch.tensor([orig_pre_nms_top_n], dtype=num_anchors.dtype),
+         num_anchors), 0).to(torch.int32)).to(num_anchors.dtype)
+
+    return num_anchors, pre_nms_top_n
+
+
 class AnchorGenerator(nn.Module):
     """
     Module that generates anchors for a set of feature maps and
@@ -85,14 +99,24 @@ def grid_anchors(self, grid_sizes, strides):
         ):
             grid_height, grid_width = size
             stride_height, stride_width = stride
+            if torchvision._is_tracing():
+                # required in ONNX export for mult operation with float32
+                stride_width = torch.tensor(stride_width, dtype=torch.float32)
+                stride_height = torch.tensor(stride_height, dtype=torch.float32)
             device = base_anchors.device
             shifts_x = torch.arange(
                 0, grid_width, dtype=torch.float32, device=device
             ) * stride_width
             shifts_y = torch.arange(
                 0, grid_height, dtype=torch.float32, device=device
             ) * stride_height
-            shift_y, shift_x = torch.meshgrid(shifts_y, shifts_x)
+            # TODO: remove tracing pass when exporting torch.meshgrid()
+            #       is suported in ONNX
+            if torchvision._is_tracing():
+                shift_y = shifts_y.view(-1, 1).expand(grid_height, grid_width)
+                shift_x = shifts_x.view(1, -1).expand(grid_height, grid_width)
+            else:
+                shift_y, shift_x = torch.meshgrid(shifts_y, shifts_x)
             shift_x = shift_x.reshape(-1)
             shift_y = shift_y.reshape(-1)
             shifts = torch.stack((shift_x, shift_y, shift_x, shift_y), dim=1)
@@ -114,7 +138,9 @@ def cached_grid_anchors(self, grid_sizes, strides):
     def forward(self, image_list, feature_maps):
         grid_sizes = tuple([feature_map.shape[-2:] for feature_map in feature_maps])
         image_size = image_list.tensors.shape[-2:]
-        strides = tuple((image_size[0] / g[0], image_size[1] / g[1]) for g in grid_sizes)
+        strides = tuple((float(image_size[0]) / float(g[0]),
+                         float(image_size[1]) / float(g[1]))
+                        for g in grid_sizes)
         dtype, device = feature_maps[0].dtype, feature_maps[0].device
         self.set_cell_anchors(dtype, device)
         anchors_over_all_feature_maps = self.cached_grid_anchors(grid_sizes, strides)
@@ -300,8 +326,11 @@ def _get_top_n_idx(self, objectness, num_anchors_per_level):
         r = []
         offset = 0
         for ob in objectness.split(num_anchors_per_level, 1):
-            num_anchors = ob.shape[1]
-            pre_nms_top_n = min(self.pre_nms_top_n, num_anchors)
+            if torchvision._is_tracing():
+                num_anchors, pre_nms_top_n = _onnx_get_num_anchors_and_pre_nms_top_n(ob, self.pre_nms_top_n)
+            else:
+                num_anchors = ob.shape[1]
+                pre_nms_top_n = min(self.pre_nms_top_n, num_anchors)
             _, top_n_idx = ob.topk(pre_nms_top_n, dim=1)
             r.append(top_n_idx + offset)
             offset += num_anchors

torchvision/ops/misc.py

@@ -132,7 +132,7 @@ def _output_size(dim):
 
 
 # This is not in nn
-class FrozenBatchNorm2d(torch.jit.ScriptModule):
+class FrozenBatchNorm2d(torch.nn.Module):
     """
     BatchNorm2d where the batch statistics and the affine parameters
     are fixed
@@ -145,7 +145,6 @@ def __init__(self, n):
         self.register_buffer("running_mean", torch.zeros(n))
         self.register_buffer("running_var", torch.ones(n))
 
-    @torch.jit.script_method
     def forward(self, x):
         # move reshapes to the beginning
         # to make it fuser-friendly