def version_10(cls, ctx, node, **kwargs):
# map to onnx as:
# not (isinf(x) or isnan(x))
shapes = node.output_shapes
dtypes = [onnx_pb.TensorProto.BOOL] * len(node.output_dtypes)
ctx.remove_node(node.name)
inf_node = ctx.make_node("IsInf",
inputs=node.input,
name=utils.make_name(node.name),
shapes=shapes,
dtypes=dtypes)
nan_node = ctx.make_node("IsNaN",
inputs=node.input,
name=utils.make_name(node.name),
shapes=shapes,
dtypes=dtypes)
or_node = ctx.make_node(
"Or",
inputs=[inf_node.output[0], nan_node.output[0]],
name=utils.make_name(node.name),
shapes=shapes,
dtypes=dtypes)
_ = ctx.make_node("Not",
inputs=or_node.output,
name=node.name,
shapes=shapes,
dtypes=dtypes)
def version_11(cls, ctx, node, **kwargs):
supported_dtypes = [
onnx_pb.TensorProto.INT32, onnx_pb.TensorProto.INT64
]
onnx_dtype = ctx.get_dtype(node.input[0])
utils.make_sure(onnx_dtype in supported_dtypes,
"InvertPermutation only applies on INT32, INT64.")
shape = ctx.get_shape(node.input[0])
shape_node = ctx.make_node("Shape",
inputs=node.input,
name=utils.make_name(node.name + '_shape'))
neg_node = ctx.make_node("Neg",
inputs=node.input,
name=utils.make_name(node.name + '_neg'))
topk_node = ctx.make_node(
"TopK",
inputs=[neg_node.output[0], shape_node.output[0]],
name=utils.make_name(node.name + '_topk'),
output_count=2)
ctx.remove_node(node.name)
last_node = ctx.make_node("Identity",
inputs=topk_node.output[1:],
name=utils.make_name(node.name + '_indices'),
shapes=[shape],
dtypes=[onnx_dtype])
ctx.replace_all_inputs(node.output[0],
last_node.output[0]) # ops=ctx.get_nodes()
def _create_loop_node(self, context, loop_props, init_cond_output, branches=None):
loop_outputs = []
loop_output_shapes = []
loop_output_dtypes = []
for tensor_value_info in loop_props.state_outputs_exits + loop_props.scan_outputs_exits:
if tensor_value_info.id:
loop_outputs.append(tensor_value_info.id)
loop_output_shapes.append(tensor_value_info.shape)
loop_output_dtypes.append(tensor_value_info.dtype)
n = self.g.get_node_by_output(tensor_value_info.id)
self.g.remove_node(n.name)
else:
loop_outputs.append(utils.make_name("unused_loop_output_"))
loop_output_shapes.append([-1])
loop_output_dtypes.append(None)
# trip count and cond are not used, giving them values just because bug
# (https://github.com/Microsoft/onnxruntime/issues/255) of onnxruntime.
trip_cnt = self.g.make_const(utils.make_name("trip_count"), np.array(sys.maxsize, dtype=np.int64))
loop_node = self.g.make_node("Loop", [trip_cnt.output[0]] + [init_cond_output] +
loop_props.state_inputs_initial_values, # ONNX Loop support state inputs only
outputs=loop_outputs, op_name_scope="generic_loop",
shapes=loop_output_shapes, dtypes=loop_output_dtypes,
skip_conversion=False, branches=branches)
return loop_node
def any_version(cls, opset, ctx, node, **kwargs):
node.domain = constants.CONTRIB_OPS_DOMAIN
separator = node.get_attr_value("separator")
if separator is None:
separator = b''
separator = separator.decode('UTF-8')
separator_node = ctx.make_const(utils.make_name("separator"),
np.array([separator], np.object))
axis_node = ctx.make_const(utils.make_name("axis"),
np.array([0], np.int64))
inps_with_shapes = [i for i in node.input if ctx.get_shape(i) != []]
shape_node = None
if 0 < len(inps_with_shapes) < len(node.input):
shape_node = ctx.make_node("Shape", [inps_with_shapes[0]])
unsqueezes = []
for inp in node.input:
if ctx.get_shape(inp) == [] and shape_node is not None:
expand_node = ctx.make_node("Expand",
[inp, shape_node.output[0]])
inp = expand_node.output[0]
unsqueeze_node = GraphBuilder(ctx).make_unsqueeze({
'data': inp,
'axes': [0]
})
unsqueezes.append(unsqueeze_node)
stack_node = ctx.make_node("Concat", unsqueezes, attr={'axis': 0})
ctx.replace_inputs(node, [
stack_node.output[0], separator_node.output[0], axis_node.output[0]
])
def version_1(cls, ctx, node, **kwargs):
if 'scale' not in node.attr:
# Somtimes tflite uses a Dequantize to go from fp16 to fp32
node.type = "Cast"
node.set_attr('to', ctx.get_dtype(node.output[0]))
return
scale = np.array(node.get_attr_value('scale'), dtype=np.float32)
zero_point = np.array(node.get_attr_value('zero_point'), dtype=np.float32)
axis = node.get_attr_value('quantized_dimension')
in_rank = ctx.get_rank(node.input[0])
def expand_tensor(t):
if t.shape == (1,):
return t[0]
utils.make_sure(in_rank is not None, "Cannot dequantize node %s with unknown input rank", node.name)
new_shape = [1] * in_rank
new_shape[axis] = t.shape[0]
return t.reshape(new_shape)
scale = expand_tensor(scale)
zero_point = expand_tensor(zero_point)
if node.inputs[0].is_const():
x_val = node.inputs[0].get_tensor_value(as_list=False).astype(np.float32)
new_val = (x_val - zero_point) * scale
dequant_const = ctx.make_const(utils.make_name(node.name), new_val)
ctx.replace_all_inputs(node.output[0], dequant_const.output[0])
ctx.remove_node(node.name)
else:
scale_const = ctx.make_const(utils.make_name(node.name + "_scale"), scale).output[0]
zero_point_const = ctx.make_const(utils.make_name(node.name + "_zero_point"), zero_point).output[0]
cast_node = ctx.make_node("Cast", [node.input[0]], attr={'to': TensorProto.FLOAT},
op_name_scope=node.name).output[0]
sub_node = ctx.make_node("Sub", [cast_node, zero_point_const], op_name_scope=node.name).output[0]
mul_node = ctx.make_node("Mul", [sub_node, scale_const], op_name_scope=node.name).output[0]
ctx.replace_all_inputs(node.output[0], mul_node)
ctx.remove_node(node.name)
def process_weights_and_bias_per_layer(self, context, i):
weights = context.weights[i]
w_r_icfo = weights["weight"]
w_dtype = weights["weight"].dtype
b_r_icfo = weights["bias"]
b_dtype = weights["bias"].dtype
ft_bias_scalar = weights["ft_bias"]
# split bias for each hidden unit
# b_r_icfo: (4 * num_units,)
bias_dim = b_r_icfo.shape[0]
hidden_size = int(bias_dim / 4)
b_r_icfo = np.reshape(b_r_icfo, (1, bias_dim))
bias_gates = np.split(b_r_icfo, 4, axis=1)
ft_bias = np.add(bias_gates[2], ft_bias_scalar)
wb_bias_iofc = np.concatenate(
(bias_gates[0], bias_gates[3], ft_bias, bias_gates[1]), axis=1)
# fill Rb with empty since in TF, we have only one bias.
rb_bias_iofc = np.zeros((1, bias_dim), dtype=b_dtype)
B = np.concatenate((wb_bias_iofc, rb_bias_iofc), axis=1)
assert B.shape == (1, 2 * bias_dim)
[wx, wh] = np.split(w_r_icfo, [-1 * hidden_size])
input_size = wx.shape[0]
assert wx.shape[0] == input_size
assert int(wx.shape[1] / 4) == hidden_size
# split weight for gates
w_gates = np.split(wx, 4, axis=1)
new_wx = np.concatenate(
(w_gates[0], w_gates[3], w_gates[2], w_gates[1]), axis=1)
h_gates = np.split(wh, 4, axis=1)
new_wh = np.concatenate(
(h_gates[0], h_gates[3], h_gates[2], h_gates[1]), axis=1)
W_iofc = np.transpose(new_wx)
R_iofc = np.transpose(new_wh)
W = np.array([W_iofc], w_dtype)
R = np.array([R_iofc], w_dtype)
# create node
w_name = utils.make_name("W" + str(i))
w_node = self.g.make_const(w_name, W, skip_conversion=True)
r_name = utils.make_name("R" + str(i))
r_node = self.g.make_const(r_name, R, skip_conversion=True)
b_name = utils.make_name("B" + str(i))
b_node = self.g.make_const(b_name, B, skip_conversion=True)
context.input_size[i] = input_size
context.hidden_size[i] = hidden_size
context.onnx_input_ids[i]["W"] = w_node.output[0]
context.onnx_input_ids[i]["R"] = r_node.output[0]
context.onnx_input_ids[i]["B"] = b_node.output[0]
def wire_tfl_while_body(g, loop_node_inputs, output_shapes, output_dtypes,
cond_graph, scan_outputs):
"""Wire subgraph graph into main."""
g = copy.deepcopy(g)
graph_inputs = g.inputs.copy()
# onnx will pass in cond as argument
iter_node = g.make_node("Placeholder", [],
name=utils.make_name("iteration_num"),
output_count=1,
dtypes=[TensorProto.INT64],
shapes=[[]])
cond_node = g.make_node("Placeholder", [],
name=utils.make_name("cond"),
output_count=1,
dtypes=[TensorProto.BOOL],
shapes=[[]])
cond_binding = parameter_binding(cond_graph, g.outputs)
to_remove = set()
for idx, scan_output in scan_outputs:
inp = graph_inputs[idx]
# Remove consumers of scan input
stack = [inp]
while stack:
node = stack.pop()
if node not in to_remove:
to_remove.add(node)
for out in node.output:
stack += g.find_output_consumers(out)
# Remove scan input from cond graph
cond_binding = {
k: "@@ALLOC" if v == g.outputs[idx] else v
for k, v in cond_binding.items()
}
del g.inputs[idx]
del g.outputs[idx]
g.outputs.append(scan_output)
for node in to_remove:
g.remove_node(node.name)
# in onnx the body inputs are: index, cond, [loop_vars]
g.inputs = [iter_node, cond_node] + g.inputs
for p, c in zip(loop_node_inputs, g.input_names):
shape = p.output_shapes[0]
g.set_shape(c, shape)
cond_outputs = inline_subgraph(g, cond_graph, "cond__", cond_binding)
g.outputs = [cond_outputs[0]] + g.outputs
return g
def version_11(cls, ctx, node, **kwargs):
# add min and max as inputs
node.type = "Clip"
onnx_dtype = ctx.get_dtype(node.input[0])
np_dtype = utils.ONNX_TO_NUMPY_DTYPE[onnx_dtype]
clip_min = ctx.make_const(utils.make_name("{}_min".format(node.name)),
np.array(0.0, dtype=np_dtype))
clip_max = ctx.make_const(utils.make_name("{}_max".format(node.name)),
np.array(6.0, dtype=np_dtype))
node.input.append(clip_min.output[0])
node.input.append(clip_max.output[0])
def any_version(cls, opset, ctx, node, **kwargs):
"""
Computes the modules of a complex.
If the matrix dtype is not complex64 or complex128,
it assumes the first dimension means real part (0)
and imaginary part (1, :, :...).
"""
supported_dtypes = [
onnx_pb.TensorProto.FLOAT,
onnx_pb.TensorProto.FLOAT16,
onnx_pb.TensorProto.DOUBLE,
onnx_pb.TensorProto.COMPLEX64,
onnx_pb.TensorProto.COMPLEX128,
]
onnx_dtype = ctx.get_dtype(node.input[0])
utils.make_sure(onnx_dtype in supported_dtypes, "Unsupported input type.")
shape = ctx.get_shape(node.input[0])
np_dtype = utils.map_onnx_to_numpy_type(onnx_dtype)
utils.make_sure(shape[0] == 2, "ComplexAbs expected the first dimension to be 2 but shape is %r", shape)
ind0 = ctx.make_const(name=utils.make_name('cst0'), np_val=np.array([0], dtype=np.int64))
ind1 = ctx.make_const(name=utils.make_name('cst1'), np_val=np.array([1], dtype=np.int64))
p2 = ctx.make_const(name=utils.make_name('p2'), np_val=np.array([2], dtype=np_dtype))
real_part = ctx.make_node(
'Gather', inputs=[node.input[0], ind0.name], attr=dict(axis=0),
name=utils.make_name('Real_' + node.name))
imag_part = ctx.make_node(
'Gather', inputs=[node.input[0], ind1.name], attr=dict(axis=0),
name=utils.make_name('Imag_' + node.name))
real_part2 = ctx.make_node(
'Pow', inputs=[real_part.output[0], p2.name],
name=utils.make_name(real_part.name + 'p2p'))
imag_part2 = ctx.make_node(
'Pow', inputs=[imag_part.output[0], p2.name],
name=utils.make_name(imag_part.name + 'p2p'))
ctx.remove_node(node.name)
add = ctx.make_node(
"Add", inputs=[real_part2.output[0], imag_part2.output[0]],
name=utils.make_name('ComplexAbs_' + node.name))
squeezed = GraphBuilder(ctx).make_squeeze(
{'data': add.output[0], 'axes': [0]}, name=utils.make_name('ComplexAbs' + node.name), return_node=True)
last_node = ctx.make_node(
"Sqrt", inputs=squeezed.output[:1],
name=utils.make_name('ComplexAbs' + node.name),
shapes=[shape[1:]], dtypes=[onnx_dtype])
ctx.replace_all_inputs(node.output[0], last_node.output[0]) # ops=ctx.get_nodes()
def version_8(cls, ctx, node, **kwargs):
supported = [
onnx_pb.TensorProto.FLOAT16, onnx_pb.TensorProto.FLOAT,
onnx_pb.TensorProto.DOUBLE
]
# fetch those upfront since they are not accessible once we remove 'node'
shapes = node.output_shapes
dtypes = node.output_dtypes
input_dtype = ctx.get_dtype(node.input[0])
name = node.name
min_node = node.input[1]
if ctx.get_dtype(min_node) not in supported:
# cast min if needed
min_node = ctx.insert_new_node_on_input(
node, "Cast", min_node, to=onnx_pb.TensorProto.FLOAT).output[0]
max_node = node.input[2]
if ctx.get_dtype(max_node) not in supported:
# cast max if needed
max_node = ctx.insert_new_node_on_input(
node, "Cast", max_node, to=onnx_pb.TensorProto.FLOAT).output[0]
ctx.remove_node(name)
new_node = ctx.make_node("Max", [node.input[0], min_node],
outputs=[node.output[0]],
shapes=shapes,
dtypes=dtypes)
if input_dtype not in supported:
# cast the data tensor if needed
ctx.insert_new_node_on_input(new_node,
"Cast",
new_node.input[0],
to=onnx_pb.TensorProto.FLOAT)
new_node = ctx.insert_new_node_on_output("Min",
new_node.output[0],
name=utils.make_name(name))
new_node.input.append(max_node)
# copy shape and type
ctx.set_dtype(new_node.output[0], dtypes[0])
ctx.set_shape(new_node.output[0], shapes[0])
if dtypes[0] not in supported:
# cast output if needed
new_node = ctx.insert_new_node_on_output(
"Cast",
new_node.output[0],
name=utils.make_name(name),
to=dtypes[0])
# copy shape and type
ctx.set_dtype(new_node.output[0], dtypes[0])
ctx.set_shape(new_node.output[0], shapes[0])
def version_11(cls, ctx, node, **kwargs):
dir_map = {"LeftShift": "LEFT", "RightShift": "RIGHT"}
direction = dir_map[node.type]
supported = [
onnx_pb.TensorProto.UINT8, onnx_pb.TensorProto.UINT16,
onnx_pb.TensorProto.UINT32, onnx_pb.TensorProto.UINT64
]
type_map = {
onnx_pb.TensorProto.INT8: onnx_pb.TensorProto.UINT8,
onnx_pb.TensorProto.INT16: onnx_pb.TensorProto.UINT32,
onnx_pb.TensorProto.INT32: onnx_pb.TensorProto.UINT64
}
shapes = node.output_shapes
dtypes = node.output_dtypes
ctx.remove_node(node.name)
node = ctx.make_node("BitShift",
inputs=node.input,
outputs=node.output,
name=node.name,
shapes=shapes,
dtypes=dtypes,
domain=constants.ONNX_DOMAIN,
attr={'direction': direction})
if node.maybe_cast_input([supported, supported], type_map):
cast_back_node = ctx.insert_new_node_on_output(
"Cast",
node.output[0],
name=utils.make_name(node.name) + "_castback",
to=dtypes[0])
ctx.set_dtype(cast_back_node.output[0], dtypes[0])
ctx.copy_shape(node.name, cast_back_node.output[0])
def version_6(cls, ctx, node, **kwargs):
# T output = All(T x, list(int) reduce_indices, @bool keepdims)
# T output = Any(T x, list(int) reduce_indices, @bool keepdims)
reduce_dim = node.inputs[1].get_tensor_value()
# for Any, the reduce_indices can be scalar as observed.
if np.isscalar(reduce_dim):
reduce_dim = [reduce_dim]
if ctx.opset < 11:
utils.make_sure(all(i >= 0 for i in reduce_dim), "negative reduce axis is not supported in onnx for now")
cast = ctx.make_node(op_type="Cast", inputs=[node.input[0]], attr={"to": onnx_pb.TensorProto.FLOAT})
keepdims = helper.get_attribute_value(node.get_attr("keep_dims"))
op_type = "ReduceMin" if node.type == "All" else "ReduceSum"
if op_type == "ReduceSum":
reduce_node_output = GraphBuilder(ctx).make_reduce_sum(
{"data": cast.output[0], "axes": reduce_dim, "keepdims": keepdims, "noop_with_empty_axes": 1})
else:
reduce_node_output = ctx.make_node(op_type=op_type, inputs=cast.output,
attr={"axes": reduce_dim, "keepdims": keepdims}).output[0]
zero_node = ctx.make_const(utils.make_name("zero_reduce"), np.array(0, dtype=np.float32))
shapes = node.output_shapes
dtypes = node.output_dtypes
ctx.remove_node(node.name)
ctx.make_node(op_type="Greater", inputs=[reduce_node_output, zero_node.output[0]],
name=node.name, outputs=node.output, shapes=shapes, dtypes=dtypes)
def test_rewrite_subgraph(self):
graph_proto = self.sample_net()
g = GraphUtil.create_graph_from_onnx_graph(graph_proto)
pattern = \
OpTypePattern('Abs', name='output', inputs=[
OpTypePattern('Add', name='input')
])
ops = g.get_nodes()
matcher = GraphMatcher(pattern)
match_results = list(matcher.match_ops(ops))
for match in match_results:
input_node = match.get_op('input')
output_node = match.get_op('output')
op_name = utils.make_name("ReplacedOp")
out_name = utils.port_name(op_name)
new_node = g.make_node("Sub",
inputs=input_node.input,
outputs=[out_name],
name=op_name)
g.replace_all_inputs(output_node.output[0],
new_node.output[0]) # ops=ops
for n in set(match.get_nodes()):
g.remove_node(n.name)
g.topological_sort(ops)
result = onnx_to_graphviz(g)
expected = 'digraph { Placeholder__5 [op_type=Placeholder] n1 [op_type=Abs] ' \
'n3 [op_type=Abs] n2 [op_type=Abs] ReplacedOp__6 [op_type=Sub] ' \
'n6 [op_type=Identity] n5_graph_outputs_Identity__4 [op_type=Identity] ' \
'input -> n1 n1:0 -> n3 n1:0 -> n2 n2:0 -> ReplacedOp__6 n3:0 -> ReplacedOp__6 ' \
'ReplacedOp__6:0 -> n6 ReplacedOp__6:0 -> n5_graph_outputs_Identity__4 }'
self.assertEqual(expected, result)
def version_9(cls, ctx, node, **kwargs):
# T output = Select(bool condition, T x, T y)
# T1 output = Where(bool condition, T1 x, T1 y)
# NOTE: condition can be 1-dimension in tensorflow, while in onnx,
# it should be broadcastable with other two inputs
if ctx.get_dtype(node.output[0]) != TensorProto.STRING:
# Due to bad ORT implementation, Mul/Add ops are faster than Where op
cls.version_7(ctx, node, **kwargs)
return
cond_shape = ctx.get_shape(node.input[0])
input_shape = ctx.get_shape(node.input[1])
if input_shape is None:
input_shape = ctx.get_shape(node.input[2])
input_rank = len(input_shape) if input_shape is not None else None
cond_rank = len(cond_shape) if cond_shape is not None else None
# if cond shape is 1-dimensional while input has higher rank, need to be reshaped to broadcast
if node.type == "Select" and cond_rank == 1 and input_rank != 1:
utils.make_sure(input_rank is not None,
"input_rank unknown and cond_rank == 1")
broadcast_shape = [cond_shape[0]] + [1] * (input_rank - 1)
shape_const = ctx.make_const(
utils.make_name(node.name),
np.array(broadcast_shape, dtype=np.int64))
reshape = ctx.make_node("Reshape",
[node.input[0], shape_const.output[0]])
ctx.replace_input(node, node.input[0], reshape.output[0], 0)
node.type = "Where"
def version_1(cls, ctx, node, **kwargs):
# output_type output = ArgMin(T input, Tidx dimension, @type Tidx, @type output_type)
# tensor(int32) reduced = ArgMin(T data, @INT axis, @INT keepdims)
axis_node = node.inputs[1]
axis = axis_node.get_tensor_value()
if axis < 0:
# ArgMax|ArgMin in onnx don't necessary support negative axis(not in doc explicitly)
input_shape = ctx.get_shape(node.input[0])
dim_count = len(input_shape) if input_shape else 0
axis = dim_count + axis
# TF ArgMin/ArgMax may return int32 or int64
# Onnx ArgMin/ArgMax only supports int64 output, add cast if needed
if node.get_attr_int("output_type") == onnx_pb.TensorProto.INT32:
# current node will return int64 after conversion, which differs from previous dtype got from tf
ctx.set_dtype(node.output[0], onnx_pb.TensorProto.INT64)
op_name = utils.make_name("Cast")
cast_node = ctx.insert_new_node_on_output("Cast", node.output[0], name=op_name,
to=onnx_pb.TensorProto.INT32)
ctx.set_dtype(cast_node.output[0], onnx_pb.TensorProto.INT32)
ctx.copy_shape(node.output[0], cast_node.output[0])
node.set_attr("axis", axis)
node.set_attr("keepdims", 0)
ctx.remove_input(node, node.input[1], 1)
def version_1(cls, ctx, node, **kwargs):
node.type = "Sqrt"
op_name = utils.make_name(node.name)
reciprocal = ctx.insert_new_node_on_output("Reciprocal",
node.output[0],
name=op_name)
ctx.copy_shape(node.output[0], reciprocal.output[0])
def version_1(cls, ctx, node, **kwargs):
# ONNX: Each input value is divided by (bias+(alpha/size)*sum(xi^2 for every xi in the local region))^beta
# TF: sqr_sum[a, b, c, d] = sum(input[a, b, c, d - depth_radius : d + depth_radius + 1] ** 2)
# output = input / (bias + alpha * sqr_sum) ** beta
# by default, depth_radius is 5 in tensorflow
size = node.get_attr_value("depth_radius", 5) * 2 + 1
node.set_attr("size", size)
node.set_attr("alpha", size * node.get_attr("alpha").f)
shapes = node.output_shapes[0]
dtypes = node.output_dtypes[0]
ctx.insert_new_node_on_input(node,
"Transpose",
node.input[0],
perm=constants.NHWC_TO_NCHW)
ctx.update_node_shape_dtype(node, override=True)
op_name = utils.make_name(node.name)
ctx.insert_new_node_on_output("Transpose",
node.output[0],
perm=constants.NCHW_TO_NHWC,
name=op_name,
shapes=shapes,
dtypes=dtypes)
def version_1(cls, ctx, node, **kwargs):
node.type = "Sub"
op_name = utils.make_name(node.name)
mul = ctx.insert_new_node_on_output("Mul",
node.output[0],
name=op_name)
mul.input.append(node.output[0])
def version_1(cls, ctx, node, **kwargs):
if ctx.is_target(constants.TARGET_CAFFE2):
# workaround a bug in caffe2 pre Feb2018, pow(a, b) becomes np.exp(np.log(a) * b)
node.type = "Log"
b = node.input[1]
ctx.remove_input(node, node.input[1], 1)
op_name = utils.make_name(node.name)
mul_op = ctx.insert_new_node_on_output("Mul",
node.output[0],
name=op_name)
mul_op.input.append(b)
op_name = utils.make_name(node.name)
exp_op = ctx.insert_new_node_on_output("Exp",
mul_op.output[0],
name=op_name)
ctx.copy_shape(node.output[0], exp_op.output[0])
BroadcastOp.version_1(ctx, mul_op, **kwargs)
def version_1(cls, ctx, node, **kwargs):
node.domain = constants.CONTRIB_OPS_DOMAIN
num_buckets = node.get_attr_int('num_buckets')
num_buckets_const = ctx.make_const(
utils.make_name('num_buckets'),
np.array([num_buckets], dtype=np.int64))
ctx.replace_inputs(node, [node.input[0], num_buckets_const.output[0]])
del node.attr['num_buckets']
def version_1(cls, ctx, node, **kwargs):
node.domain = constants.CONTRIB_OPS_DOMAIN
node.type = "StringRegexReplace"
pattern = node.get_attr_str("pattern")
rewrite = node.get_attr_str("rewrite")
utils.make_sure(
node.get_attr_value("replace_global") != 0,
"Can not convert StaticRegexReplace if replace_global is False")
pattern_node = ctx.make_const(utils.make_name("pattern"),
np.array([pattern], np.object))
rewrite_node = ctx.make_const(utils.make_name("rewrite"),
np.array([rewrite], np.object))
del node.attr["pattern"]
del node.attr["rewrite"]
del node.attr["replace_global"]
ctx.replace_inputs(
node,
[node.input[0], pattern_node.output[0], rewrite_node.output[0]])
def version_10(cls, ctx, node, **kwargs):
scale = node.get_attr_value('scale')
zero_point = node.get_attr_value('zero_point')
axis = node.get_attr_value('quantized_dimension')
np_q_type = utils.map_onnx_to_numpy_type(ctx.get_dtype(node.output[0]))
if len(scale) > 1 or len(zero_point) > 1:
utils.make_sure(ctx.opset >= 13, "Opset 13 is required for per-axis quantization for node %s", node.name)
node.set_attr("axis", axis)
scale_node = ctx.make_const(utils.make_name("scale"), np.array(scale[0], dtype=np.float32))
zero_point_node = ctx.make_const(utils.make_name("zero_point"), np.array(zero_point[0], dtype=np_q_type))
ctx.replace_inputs(node, [node.input[0], scale_node.output[0], zero_point_node.output[0]])
del node.attr["scale"]
del node.attr["zero_point"]
del node.attr["quantized_dimension"]
if "min" in node.attr:
del node.attr["min"]
if "max" in node.attr:
del node.attr["max"]
def version_7(cls, ctx, node, **kwargs):
GreaterLess.version_7(ctx, node, **kwargs)
output_name = node.output[0]
node.op.op_type = "Less" if node.op.op_type == "GreaterEqual" else "Greater"
new_node = ctx.insert_new_node_on_output("Not",
output_name,
name=utils.make_name(
node.name))
ctx.copy_shape(output_name, new_node.output[0])
ctx.set_dtype(new_node.output[0], ctx.get_dtype(output_name))
def version_1(cls, ctx, node, **kwargs):
need_not = node.type == "NotEqual"
common.BroadcastOp.version_1(ctx, node, **kwargs)
if need_not:
node.type = "Equal"
output_name = node.output[0]
not_node = ctx.insert_new_node_on_output("Not",
output_name,
name=utils.make_name(
node.name))
ctx.copy_shape(output_name, not_node.output[0])
ctx.copy_dtype(output_name, not_node.output[0])
def to_tf(cls, ctx, node, **kwargs):
if len(node.input) == 1 or ctx.get_rank(node.input[1]) != 1:
new_shape = node.get_attr_value('new_shape')
if new_shape == [0]:
# Legacy tflite models use a shape parameter of [0] to indicate scalars
new_shape = []
new_shape_const = ctx.make_const(utils.make_name("new_shape"),
np.array(new_shape, np.int64))
ctx.replace_inputs(node,
[node.input[0], new_shape_const.output[0]])
if 'new_shape' in node.attr:
del node.attr['new_shape']
def version_9(cls, ctx, node, **kwargs):
# T_y output = Where(T_x condition), return indices of elements whose value are True
node.type = "NonZero"
# in onnx, indices are returned in this way [[ind_a_0, ind_b_0, ...], [ind_a_1, ind_b_1,...]];
# while in tf, the result will be [[ind_a_0, ind_a_1, ...], [ind_b_0, ind_b_1, ...], ...]
# this is the reason a transpose node inserted here.
transpose_node = ctx.insert_new_node_on_output(
"Transpose",
node.output[0],
name=utils.make_name("where_op_added"))
ctx.copy_shape(node.output[0], transpose_node.output[0])
ctx.copy_dtype(node.output[0], transpose_node.output[0])
def process_single_init_node(g, fw_init_input_id, bw_init_input_id, to_append):
fw_init_is_const, init_fw_val = check_const(g, fw_init_input_id)
bw_init_is_const, init_bw_val = check_const(g, bw_init_input_id)
if fw_init_is_const and bw_init_is_const:
initial_val = np.concatenate((init_fw_val, init_bw_val), axis=0)
init_name = utils.make_name("initial")
init_node = g.make_const(init_name, initial_val, skip_conversion=True)
else:
init_node = g.make_node("Concat", [fw_init_input_id, bw_init_input_id],
attr={"axis": 0})
to_append.append(init_node)
return init_node
def version_11(cls, ctx, node, **kwargs):
# starting with opset-11, equal supports all types (but both operands must be of the same type)
_add_cast_to_same_type_to_inputs(ctx, node)
need_not = node.type == "NotEqual"
if need_not:
node.type = "Equal"
output_name = node.output[0]
not_node = ctx.insert_new_node_on_output("Not",
output_name,
name=utils.make_name(
node.name))
ctx.copy_shape(output_name, not_node.output[0])
ctx.copy_dtype(output_name, not_node.output[0])
def make_range_const(ctx, start, limit, delta, output, scope_name, shape,
dtype):
"""make Range subgraph if all inputs are const."""
# T range = Range(T start, T limit, T delta)
# V v_final_and_scan_outputs = Loop(int64 M, B cond, V v_initial)
base_name = utils.make_name(scope_name)
start = ctx.get_node_by_output(start).get_tensor_value(as_list=False)
limit = ctx.get_node_by_output(limit).get_tensor_value(as_list=False)
delta = ctx.get_node_by_output(delta).get_tensor_value(as_list=False)
val = np.arange(start, limit, delta, dtype=start.dtype)
const_range = ctx.make_const(base_name, val)
ctx.make_node("Identity", [const_range.output[0]],
shapes=[shape],
dtypes=[dtype],
outputs=[output])
def version_6(cls, ctx, node, **kwargs):
# T output = FloorDiv(T x, T y)
node.type = "Div"
dtype = ctx.get_dtype(node.input[0])
if dtype in [
onnx_pb.TensorProto.FLOAT, onnx_pb.TensorProto.FLOAT16,
onnx_pb.TensorProto.DOUBLE
]:
new_node_name = utils.make_name("floor_div_res")
floor_res = ctx.insert_new_node_on_output(
op_type="Floor",
output_name=node.output[0],
name=new_node_name)
ctx.copy_dtype(node.output[0], floor_res.output[0])
ctx.copy_shape(node.output[0], floor_res.output[0])