fix some bugs

tests/run_pretrained_models.py

@@ -15,6 +15,7 @@
 import time
 import traceback
 import zipfile
+import logging
 
 import PIL.Image
 import numpy as np
@@ -34,6 +35,9 @@
 
 # pylint: disable=broad-except,logging-not-lazy,unused-argument,unnecessary-lambda
 
+logging.basicConfig(level=logging.INFO)
+log = logging.getLogger("tf2onnx")
+
 TEMP_DIR = os.path.join(utils.get_temp_directory(), "run_pretrained")
 PERFITER = 1000
 
@@ -246,9 +250,10 @@ def run_test(self, name, backend="caffe2", debug=False, onnx_file=None, opset=No
             for k in inputs.keys():  # pylint: disable=consider-iterating-dictionary
                 t = sess.graph.get_tensor_by_name(k)
                 dtype = tf.as_dtype(t.dtype).name
-                if type != "float32":
-                    v = inputs[k]
-                    inputs[k] = v.astype(dtype)
+                v = inputs[k]
+                if dtype != v.dtype:
+                    log.warning("input dtype doesn't match tensorflow's")
+                    inputs[k] = np.array(v, dtype=dtype)
             if self.force_input_shape:
                 for k, v in inputs.items():
                     shape_override[k] = list(v.shape)

tests/test_backend.py

@@ -596,6 +596,15 @@ def test_min(self):
         _ = tf.identity(mi, name=_TFOUTPUT)
         self._run_test_case([_OUTPUT], {_INPUT: x_val1, _INPUT1: x_val2})
 
+        tf.reset_default_graph()
+        x_val1 = np.array([4.0, 16.0, 4.0, 1.6], dtype=np.int32).reshape((2, 2))
+        x_val2 = np.array([4.0, 4.0, 4.0, 4.0], dtype=np.int32).reshape((2, 2))
+        x1 = tf.placeholder(tf.int32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.placeholder(tf.int32, x_val2.shape, name=_TFINPUT1)
+        mi = tf.minimum(x1, x2)
+        _ = tf.identity(mi, name=_TFOUTPUT)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val1, _INPUT1: x_val2})
+
     @skip_caffe2_backend("issue with broadcasting scalar")
     @check_onnxruntime_incompatibility("Sub")
     def test_min_broadcast(self):
@@ -788,6 +797,35 @@ def test_concat(self):
         _ = tf.identity(x_, name=_TFOUTPUT)
         self._run_test_case([_OUTPUT], {_INPUT: x_val1, _INPUT1: x_val2, "input3:0": x_val3})
 
+    def test_concat_empty_const_input(self):
+        x_val1 = np.array([1, 2, 3], dtype=np.float32)
+        x_val2 = np.array([], dtype=np.float32)
+        x1 = tf.placeholder(tf.float32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.constant(x_val2, dtype=tf.float32)
+        x_ = tf.concat([x1, x2], 0)
+        _ = tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val1})
+
+        tf.reset_default_graph()
+        x_val1 = np.array([[1, 2, 3]], dtype=np.float32)
+        x_val2 = np.array([[]], dtype=np.float32)
+        x1 = tf.placeholder(tf.float32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.constant(x_val2, dtype=tf.float32)
+        x_ = tf.concat([x1, x2], 1)
+        _ = tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val1})
+
+        tf.reset_default_graph()
+        x_val1 = np.array([1, 2, 3], dtype=np.float32)
+        x_val2 = np.array([], dtype=np.float32)
+        x_val3 = np.array([13, 14, 15], dtype=np.float32)
+        x1 = tf.placeholder(tf.float32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.constant(x_val2, dtype=tf.float32)
+        x3 = tf.placeholder(tf.float32, x_val3.shape, name=_TFINPUT1)
+        x_ = tf.concat([x1, x2, x3], 0)
+        _ = tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val1, _INPUT1: x_val3})
+
     @check_opset_min_version(6, "cast")
     def test_concat_int64(self):
         x_val1 = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int64)

tests/test_internals.py

@@ -12,8 +12,9 @@
 from collections import namedtuple
 
 import graphviz as gv
+import numpy as np
 from onnx import TensorProto
-from onnx import helper
+from onnx import helper, numpy_helper
 
 import tensorflow as tf
 from tf2onnx import utils
@@ -247,6 +248,37 @@ def test_node_attr_onnx(self):
         self.assertTrue("my_attr" in n1.attr)
         self.assertTrue("my_attr" in n1.attr_onnx)
 
+    def test_tensor_data(self):
+        tensors = {
+            "empty_tensor": np.array([], dtype=np.float32),
+            "multi_dim_empty_tensor": np.array([[], []], dtype=np.float32),
+            "scalar": np.array(1., dtype=np.float32),
+            "one_item_array": np.array([1.], dtype=np.float32),
+            "normal_array": np.array([[1., 2.], [2., 3.]], dtype=np.float32)
+        }
+        tf.reset_default_graph()
+        with tf.Session() as sess:
+            for n, data in tensors.items():
+                tf.constant(data, dtype=tf.float32, name=n)
+
+        for tf_node in sess.graph.get_operations():
+            name = tf_node.name
+            self.assertTrue(name in tensors.keys())
+
+            self.assertTrue("value" in tf_node.node_def.attr)
+            # convert to onnx tensor value
+            tensor_value = utils.tf_to_onnx_tensor(
+                utils.get_tf_node_attr(tf_node, "value"),
+                name=utils.port_name(tf_node.name)
+            )
+            attr = helper.make_attribute("value", tensor_value)
+            # same as node.get_tensor_value(is_list=False)
+            actual = numpy_helper.to_array(helper.get_attribute_value(attr))
+
+            expected = tensors[name]
+
+            self.assertTrue(np.array_equal(expected, actual))
+
 
 if __name__ == '__main__':
     unittest_main()

tf2onnx/graph.py

@@ -908,7 +908,7 @@ def insert_new_node_on_input(self, node, op_type, input_name, name=None, domain=
                 break
         return new_node
 
-    def insert_new_node_on_output(self, op_type, output_name, name=None, domain=None, **kwargs):
+    def insert_new_node_on_output(self, op_type, output_name, name, domain=None, **kwargs):
         """Create and insert a new node into the graph.
         Args:
             op_type: type for new operation

tf2onnx/tfonnx.py

@@ -766,6 +766,15 @@ def concat_op(ctx, node, name, args):
 def concatv2_op(ctx, node, name, args):
     # T output = ConcatV2(T values, Tidx axis, @int N, @type Tidx)
     # T concat_result = Concat(T inputs, @INT axis)
+    # if any input is empty, remove the input and concat the others
+    # NOTE: workaround for https://github.com/Microsoft/onnxruntime/issues/681
+    for i, inp in enumerate(node.inputs):
+        if inp.is_const() and inp.get_tensor_value(as_list=False).size == 0:
+            ctx.remove_input(node, node.input[i])
+    # all inputs are deleted
+    if not node.input:
+        raise RuntimeError("all inputs of {} are empty".format(name))
+
     axis_node = node.inputs[-1]
     axis_val = axis_node.get_tensor_value()
     ctx.remove_input(node, node.input[-1])
@@ -1154,6 +1163,28 @@ def minmax_op(ctx, node, name, args):
     # handle this by doing something like:
     # y = min(x1, add(x2, sub(x1, x1))), where x1, x2 are the inputs and x2 is a scalar
     # this will create a tensor of zeros of the shape of x1, adds x2 to it (which broadcasts) and use that for min.
+    # support more dtype
+    supported_dtypes = [
+        onnx_pb.TensorProto.FLOAT,
+        onnx_pb.TensorProto.FLOAT16,
+        onnx_pb.TensorProto.DOUBLE
+    ]
+    target_dtype = onnx_pb.TensorProto.FLOAT
+    for inp in node.input:
+        dtype = ctx.get_dtype(inp)
+        utils.make_sure(dtype, "dtype of {} is None".format(inp))
+        if dtype not in supported_dtypes:
+            inp_cast = ctx.insert_new_node_on_input(node, "Cast", inp, to=target_dtype)
+            ctx.copy_shape(inp, inp_cast.output[0])
+            ctx.set_dtype(inp_cast.output[0], target_dtype)
+    origin_dtype = ctx.get_dtype(node.output[0])
+    utils.make_sure(origin_dtype is not None, "dtype of {} is None".format(node.output[0]))
+    ctx.set_dtype(node.output[0], target_dtype)
+    cast_name = utils.make_name(name)
+    cast_node = ctx.insert_new_node_on_output("Cast", node.output[0], name=cast_name, to=origin_dtype)
+    to_replace = [n for n in ctx.get_nodes() if n != cast_node]
+    ctx.replace_all_inputs(to_replace, node.output[0], cast_node.output[0])
+
     shapeo = ctx.get_shape(node.output[0])
     needs_broadcast_op = []
     has_correct_shape = []
@@ -1676,6 +1707,7 @@ def where_op(ctx, node, name, args):
     "BiasAdd": (biasadd_op, []),
     "BiasAddV1": (biasadd_op, []),
     "Cast": (cast_op, []),
+    "CheckNumerics": (identity_op, ["Identity"]),
     "Concat": (concat_op, ["Concat"]),
     "ConcatV2": (concatv2_op, ["Concat"]),
     "Const": (direct_op, []),

tf2onnx/utils.py

@@ -120,21 +120,39 @@ def split_nodename_and_shape(name):
 
 
 def tf_to_onnx_tensor(tensor, name=""):
-    """Convert tensorflow tensor to onnx tensor."""
+    """
+    Convert tensorflow tensor to onnx tensor.
+    Here deal with three types of tensor:
+    1. normal tensor, e.g., np.array([1,2,3], dtype=DTYPE):
+        tensor_content: raw data of [1,2,3]
+        tensor_shape.dim: [3]
+        DTYPE_val: empty
+    2. scalar tensor, e.g., np.array(1, dtype=DTYPE):
+        tensor_content: empty
+        tensor_shape.dim: [0]
+        DTYPE_val: 1
+    3. empty tensor, e.g., np.array([], dtype=DTYPE) and np.array([[]], dtype=DTYPE):
+        tensor_content: empty
+        tensor_shape.dim: [0] and [1, 0]
+        DTYPE_val: empty
+    """
     new_type = TF_TO_ONNX_DTYPE[tensor.dtype]
     tdim = tensor.tensor_shape.dim
     dims = [d.size for d in tdim]
-    # FIXME: something is fishy here
-    if dims == [0]:
-        dims = [1]
     is_raw, data = get_tf_tensor_data(tensor)
+    # empty tensor
+    if not is_raw and data is None:
+        np_data = np.array([], dtype=map_onnx_to_numpy_type(new_type)).reshape(dims)
+        return numpy_helper.from_array(np_data, name=name)
+    make_sure(data, "tensor data isn't expected to be None or empty")
+    # scalar tensor
+    if dims == [0] and not is_raw and len(data) == 1:
+        return helper.make_tensor(name, new_type, [], data, False)
     if not is_raw and len(data) == 1 and np.prod(dims) > 1:
         batch_data = np.zeros(dims, dtype=map_onnx_to_numpy_type(new_type))
         batch_data.fill(data[0])
-        onnx_tensor = numpy_helper.from_array(batch_data, name=name)
-    else:
-        onnx_tensor = helper.make_tensor(name, new_type, dims, data, is_raw)
-    return onnx_tensor
+        return numpy_helper.from_array(batch_data, name=name)
+    return helper.make_tensor(name, new_type, dims, data, is_raw)
 
 
 def get_tf_tensor_data(tensor):
@@ -154,16 +172,15 @@ def get_tf_tensor_data(tensor):
         data = tensor.int64_val
     elif tensor.bool_val:
         data = tensor.bool_val
-    elif tensor.dtype == tf.int32:
-        data = [0]
-    elif tensor.dtype == tf.int64:
-        data = [0]
-    elif tensor.dtype == tf.float32:
-        data = [0.]
-    elif tensor.dtype == tf.float16:
-        data = [0]
     elif tensor.string_val:
         data = tensor.string_val
+    elif tensor.dtype in [
+            tf.int32,
+            tf.int64,
+            tf.float32,
+            tf.float16
+    ]:
+        data = None
     else:
         raise ValueError('tensor data not supported')
     return [is_raw, data]