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))
def tensorflow_to_onnx(graph):
"""
Load tensorflow graph into an onnx graph with minimal rewrites so
we can use the onnx graph as intermediate graph.
"""
# ignore the following attributes
ignored_attr = ["unknown_rank", "_class", "Tidx", "Tshape", "use_cudnn_on_gpu", "Index",
"Tpaddings", "TI", "Tparams", "Tindices", "Tlen", "Tdim", "dynamic_size", "element_shape",
"Tmultiples", "output_dtype", "Tblock_shape", "Tcrops", "index_type"]
# some stats
op_cnt = collections.Counter()
attr_cnt = collections.Counter()
onnx_nodes = []
output_shapes = {}
dtypes = {}
# find outputs
ops = graph.get_operations()
# create dict with output to shape mappings
for node in ops:
for out in node.outputs:
try:
shape = out.get_shape().as_list()
except Exception as ex:
shape = []
output_shapes[out.name] = shape
# minimal conversion of attributes
for node in ops:
attr = {}
takeit = True
op_cnt[node.type] += 1
for a in node.node_def.attr:
attr_cnt[a] += 1
if a == "dtype":
attr[a] = utils.get_tf_dtype(node)
elif a == "T":
dtype = node.get_attr("T")
if dtype:
if not isinstance(dtype, list):
dtypes[node.name] = utils.TF_TO_ONNX_DTYPE.get(dtype)
elif a == "output_type":
out_type = node.get_attr("output_type")
out_type = utils.TF_TO_ONNX_DTYPE[out_type]
attr[a] = out_type
elif a == "out_type":
out_type = node.get_attr("out_type")
out_type = utils.TF_TO_ONNX_DTYPE[out_type]
attr[a] = out_type
elif a == "shape":
attr[a] = utils.get_shape(node)
elif a == "Tperm":
pass
elif a == "_output_shapes":
attr[a] = utils.get_shape(node)
elif a == "value":
onnx_tensor = utils.tf_to_onnx_tensor(node.get_attr(a), name=node.name + ":0")
attr[a] = onnx_tensor
elif a == "DstT":
dst = node.get_attr("DstT")
dst = tf2onnx.utils.TF_TO_ONNX_DTYPE[dst]
dst = tf2onnx.utils.ONNX_DTYPE_NAMES[dst]
attr["to"] = dst
elif a == "SrcT":
continue
elif a in ignored_attr:
continue
else:
attr[a] = node.get_attr(a)
if takeit:
try:
input_names = [i.name for i in node.inputs]
output_names = [i.name for i in node.outputs]
onnx_node = helper.make_node(node.type, input_names, output_names, name=node.name, **attr)
onnx_nodes.append(onnx_node)
except Exception as ex:
log.error("pass1 convert failed for %s, ex=%s", node, ex)
raise
return onnx_nodes, op_cnt, attr_cnt, output_shapes, dtypes
def tflist_to_onnx(node_list, shape_override):
"""
Convert the tf-node list into an onnx graph with minimal rewrites so
we can use the onnx graph as intermediate graph.
"""
# ignore the following attributes
ignored_attr = [
"unknown_rank", "_class", "Tshape", "use_cudnn_on_gpu", "Index",
"Tpaddings", "TI", "Tparams", "Tindices", "Tlen", "Tdim",
"dynamic_size", "Tmultiples", "Tblock_shape", "Tcrops", "index_type",
"Taxis", "U", "maxval", "Tout", "Tlabels", "Tindex", "element_shape"
]
# some stats
op_cnt = collections.Counter()
attr_cnt = collections.Counter()
onnx_nodes = []
output_shapes = {}
dtypes = {}
# find outputs
ops = node_list
# create dict with output to shape mappings
for node in ops:
for out in node.outputs:
shape = shape_override.get(out.name)
if shape is None:
try:
shape = out.get_shape().as_list()
except Exception as ex:
shape = None
dtypes[out.name] = utils.map_tf_dtype(out.dtype)
output_shapes[out.name] = shape
# minimal conversion of attributes
for node in ops:
attr = {}
takeit = True
op_cnt[node.type] += 1
for a in node.node_def.attr:
attr_cnt[a] += 1
if a == "dtype":
attr[a] = utils.map_tf_dtype(
utils.get_tf_node_attr(node, "dtype"))
elif a == "T":
dtype = utils.get_tf_node_attr(node, "T")
if dtype:
if not isinstance(dtype, list):
dtypes[node.name] = utils.map_tf_dtype(dtype)
elif a in [
"output_type", "output_dtype", "out_type", "Tidx",
"out_idx"
]:
# Tidx is used by Range
# out_idx is used by ListDiff
attr[a] = utils.map_tf_dtype(utils.get_tf_node_attr(node, a))
elif a == "shape":
attr[a] = utils.get_shape(node)
elif a == "Tperm":
pass
elif a == "_output_shapes":
attr[a] = utils.get_shape(node)
elif a == "value":
onnx_tensor = utils.tf_to_onnx_tensor(
utils.get_tf_node_attr(node, a), name=port_name(node.name))
attr[a] = onnx_tensor
elif a == "DstT":
attr["to"] = utils.map_tf_dtype(
utils.get_tf_node_attr(node, "DstT"))
elif a == "SrcT":
continue
elif a in ignored_attr:
continue
else:
attr[a] = utils.get_tf_node_attr(node, a)
if takeit:
try:
input_names = [i.name for i in node.inputs]
output_names = [i.name for i in node.outputs]
onnx_node = helper.make_node(node.type,
input_names,
output_names,
name=node.name,
**attr)
onnx_nodes.append(onnx_node)
except Exception as ex:
log.error("pass1 convert failed for %s, ex=%s", node, ex)
raise
return onnx_nodes, op_cnt, attr_cnt, output_shapes, dtypes