Merge branch 'master' of github.com:onnx/tensorflow-onnx into add_missing_import

Author: andhus

tests/test_backend.py

@@ -1,3 +1,4 @@
+
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # Licensed under the MIT license.
 
@@ -2488,6 +2489,19 @@ def func(x, z):
                     return space_to_batch_nd(x, y, z, name=_TFOUTPUT)
                 self._run_test_case(func, [_OUTPUT], {_INPUT: x_val, _INPUT2: z_val})
 
+    @check_opset_min_version(10, "CropAndResize")
+    def test_crop_and_resize(self):
+        boxes_val = [[0.5, 0.7, 0.7, 0.9], [0.2, 0.4, 0.4, 0.6]]
+        def func(input_x, box_ind):
+            boxes = tf.constant(boxes_val, dtype=tf.float32)
+            corp_size = tf.constant(np.array([20, 20]).astype(np.int32))
+            return tf.image.crop_and_resize(input_x, boxes, box_ind, corp_size, name=_TFOUTPUT, method='bilinear')
+
+        input_x_val = np.random.randint(low=0, high=256, size=[2, 36, 36, 3]).astype(np.float32)  # NHWC
+        box_ind_val = np.array([1, 0]).astype(np.int32)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: input_x_val, _INPUT2: box_ind_val},
+                            rtol=1e-04, atol=1e-03)
+
     @check_opset_min_version(11, "CropAndResize")
     def test_crop_and_resize_linear(self):
         def func(input_x, boxes, box_ind, corp_size):
@@ -2966,6 +2980,16 @@ def func(input_holder):
         for input_val in input_vals:
             self._run_test_case(func, [_OUTPUT], {_INPUT: input_val})
 
+    @check_opset_min_version(8)
+    def test_broadcast(self):
+        input_tensor_val = np.random.randint(low=0, high=256, size=[2, 3]).astype(np.float32)
+        new_shape_val = np.array([3, 2, 3]).astype(np.int64)
+
+        def func(input_tensor, new_shape):
+            return tf.broadcast_to(input_tensor, new_shape, _TFOUTPUT)
+
+        self._run_test_case(func, [_OUTPUT], {_INPUT: input_tensor_val, _INPUT1: new_shape_val})
+
 
 if __name__ == '__main__':
     unittest_main()

tf2onnx/graph.py

@@ -135,7 +135,8 @@ def domain(self, val):
     @property
     def data_format(self):
         """Return data_format."""
-        return self.get_attr_str("data_format")
+        attr_str = self.get_attr_value("data_format")
+        return "unkown" if attr_str is None else attr_str.decode("utf-8")
 
     @data_format.setter
     def data_format(self, val):

tf2onnx/onnx_opset/nn.py

@@ -589,6 +589,48 @@ def version_11(cls, ctx, node, **kwargs):
 
 @tf_op(["CropAndResize"])
 class CropAndResize:
+    @classmethod
+    def version_10(cls, ctx, node, **kwargs):
+        utils.make_sure(node.inputs[1].type == "Const", "boxes input must be a Const")
+        utils.make_sure(node.inputs[3].type == "Const", "boxes input must be a Const")
+        name = node.name
+        output_height = node.inputs[3].get_tensor_value()[0]
+        output_width = node.inputs[3].get_tensor_value()[1]
+        rois = node.inputs[1].get_tensor_value()
+        rois_shape = ctx.get_shape(node.input[1])
+        img_shape = ctx.get_shape(node.input[0])
+        transform_rois = np.zeros(list(rois_shape), dtype=np.float32)
+        for i in range(rois_shape[0]):
+            y1, x1, y2, x2 = rois[i]
+            y1 = y1 * (img_shape[1] - 1)
+            y2 = y2 * (img_shape[1] - 1)
+            x1 = x1 * (img_shape[2] - 1)
+            x2 = x2 * (img_shape[2] - 1)
+            spacing_h = (y2 - y1)
+            spacing_w = (x2 - x1)
+            b1 = y1 - 0.5 * spacing_h / (output_height - 1)
+            a1 = x1 - 0.5 * spacing_w / (output_width - 1)
+            b2 = y2 + 0.5 * spacing_h / (output_height - 1)
+            a2 = x2 + 0.5 * spacing_w / (output_width - 1)
+            transform_rois[i][0] = a1
+            transform_rois[i][1] = b1
+            transform_rois[i][2] = a2
+            transform_rois[i][3] = b2
+        cast_node = ctx.make_node("Cast", [node.input[2]], attr={"to": onnx_pb.TensorProto.INT64})
+        bbox_node = ctx.make_const(utils.make_name("bbox"), transform_rois)
+        dtypes = [ctx.get_dtype(node.output[0])]
+        shapes = [ctx.get_shape(node.output[0])]
+        input_nchw = ctx.make_node("Transpose", [node.input[0]], {"perm": [0, 3, 1, 2]},
+                                   name=utils.make_name(node.name))
+        crop_and_resize = ctx.make_node("RoiAlign", inputs=[input_nchw.output[0], bbox_node.output[0],
+                                                            cast_node.output[0]],
+                                        attr={"output_height": output_height, "output_width": output_width,
+                                              "spatial_scale": 1.0, "sampling_ratio": 1},
+                                        name=utils.make_name(node.name), dtypes=dtypes, shapes=shapes)
+        ctx.remove_node(name)
+        res = ctx.make_node("Transpose", crop_and_resize.output, {"perm": [0, 2, 3, 1]},
+                            name=name, outputs=node.output, shapes=shapes, dtypes=dtypes)
+
     @classmethod
     def version_11(cls, ctx, node, **kwargs):
         # create loop of resize to cater to tensorflow CropAndResize, one box one iteration

tf2onnx/onnx_opset/tensor.py

@@ -1806,3 +1806,12 @@ def version_11(cls, ctx, node, **kwargs):
         ctx.remove_node(node.name)
         squeezed_result = ctx.make_node('Squeeze', [gathered_result.output[0]], attr={"axes": [-1]},
                                         name=node.name, outputs=node.output, shapes=shapes, dtypes=dtypes)
+
+
+@tf_op("BroadcastTo")
+class BroadcastTo:
+    @classmethod
+    def version_8(cls, ctx, node, **kwargs):
+        # broadcast by expanding
+        node.type = "Expand"
+        ctx.insert_new_node_on_input(node, "Cast", node.input[1], to=TensorProto.INT64)

tf2onnx/optimizer/transpose_optimizer.py

@@ -557,11 +557,13 @@ def _pad_handler(self, trans, node):
             node.set_attr("pads", new_pads)
             return self._switch_transpose_and_node(node, trans)
         if node.inputs[1].is_const():
-            pads = node.inputs[1].get_tensor_value()
-            # NHWC->NCHW
-            new_pads = np.array([pads[0], pads[3], pads[1], pads[2], pads[4], pads[7], pads[5], pads[6]],
-                                dtype=np.int64)
-            node.inputs[1].set_tensor_value(new_pads)
+            if node.inputs[1].data_format in ["NHWC", "unkown"]:
+                pads = node.inputs[1].get_tensor_value()
+                # NHWC->NCHW
+                new_pads = np.array([pads[0], pads[3], pads[1], pads[2], pads[4], pads[7], pads[5], pads[6]],
+                                    dtype=np.int64)
+                node.inputs[1].set_tensor_value(new_pads)
+                node.inputs[1].data_format = "NCHW"
             return self._switch_transpose_and_node(node, trans)
         return False