Add tf custom op conversion example (#1878)

* add tf custom op example

Signed-off-by: hwangdeyu <[email protected]>

* fix name and remove unused code comment

Signed-off-by: hwangdeyu <[email protected]>
Co-authored-by: fatcat-z <[email protected]>

examples/tf_custom_op/add_one.cc

@@ -0,0 +1,46 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/shape_inference.h"
+
+using namespace tensorflow;
+
+
+// opregister
+REGISTER_OP("AddOne")
+  .Input("add_one: int32")
+  .Output("result: int32")
+  .SetShapeFn([](::tensorflow::shape_inference::InferenceContext *c) {
+    c->set_output(0, c->input(0));
+    return Status::OK();
+  });
+
+
+// keneldefinition
+#include "tensorflow/core/framework/op_kernel.h"
+
+class AddOneOp : public OpKernel {
+public:
+  explicit AddOneOp(OpKernelConstruction* context) : OpKernel(context) {}
+
+  void Compute(OpKernelContext* context) override {
+    // Tensor in input
+    const Tensor& input_tensor = context->input(0);
+    auto input = input_tensor.flat<int32>();
+
+    // Tensor in output
+    Tensor* output_tensor = NULL;
+    OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(), &output_tensor));
+    auto output = output_tensor->flat<int32>();
+
+    const int N = input.size();
+    for (int i = 0; i < N; i++) {
+      output(i) += 1;
+    }
+  }
+};
+
+
+REGISTER_KERNEL_BUILDER(Name("AddOne").Device(DEVICE_CPU), AddOneOp);

examples/tf_custom_op/addone_custom_op.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+
+
+import numpy as np
+import tensorflow as tf
+import tf2onnx
+import onnx
+import os
+from tf2onnx import utils
+from tf2onnx.handler import tf_op
+from tf2onnx.tf_loader import tf_placeholder
+
+
+DIR_PATH = os.path.realpath(os.path.dirname(__file__))
+saved_model_path = os.path.join(DIR_PATH, "model.onnx")
+tf_library_path = os.path.join(DIR_PATH, "add_one.so")
+
+
+@tf_op("AddOne", onnx_op="Add")
+class AddOne:
+    @classmethod
+    def version_1(cls, ctx, node, **kwargs):
+        node_shape = ctx.get_shape(node.input[0])
+        const_one = ctx.make_const(utils.make_name("const_one"), np.ones(node_shape, dtype = np.int32)).output[0]
+        node.input.append(const_one)
+
+
+with tf.compat.v1.Session() as sess:
+    x = tf_placeholder(tf.int32, [2, 3], name="input")
+    AddOne = tf.load_op_library(tf_library_path)
+    x_ = AddOne.add_one(x)
+    _ = tf.identity(x_, name="output")
+
+    onnx_graph = tf2onnx.tfonnx.process_tf_graph(sess.graph,
+                                                 input_names=["input:0"],
+                                                 output_names=["output:0"])
+    model_proto = onnx_graph.make_model("test")
+    with open(saved_model_path, "wb") as f:
+        f.write(model_proto.SerializeToString())
+
+onnx_model = onnx.load(saved_model_path)
+onnx.checker.check_model(onnx_model)
+
+
+
+## Run the model in ONNXRuntime to verify the result.
+import onnxruntime as ort
+input = np.arange(6).reshape(2,3).astype(np.int32)
+ort_session = ort.InferenceSession(saved_model_path)
+ort_inputs = {ort_session.get_inputs()[0].name: input}
+
+ort_outs = ort_session.run(None, ort_inputs)
+print("input:", input, "\nort_outs:", ort_outs)

tf2onnx/graph.py

@@ -676,7 +676,7 @@ def make_node(self, op_type, inputs, attr=None, output_count=1, outputs=None, sk
         return node
 
     def append_node(self, node):
-        "Add a node to the graph."
+        """Add a node to the graph."""
         output_shapes = node.output_shapes
         output_dtypes = node.output_dtypes
         node.graph = self

tf2onnx/tfonnx.py

@@ -518,7 +518,7 @@ def process_parsed_graph(g, custom_op_handlers, inputs_as_nchw, continue_on_erro
     # or override existing ops with a custom op.
     if custom_op_handlers is not None:
         # below is a bit tricky since there are a few api's:
-        # 1. the future way we want custom ops to be registered with the @tf_op decorator. THose handlers will be
+        # 1. the future way we want custom ops to be registered with the @tf_op decorator. Those handlers will be
         #     registered via the decorator on load of the module ... nothing is required here.
         # 2. the old custom op api: a dictionary of {name: (func, args[])
         #     We deal with this by using a compat_handler that wraps to old handler with a new style handler.