def convert_concat(onnx_graph, node):
xs = oc.ONNXValue(onnx_graph,node.args.keywords['xs'])
axis = oc.try_get_attribute(node.args.keywords['axis'])
if isinstance(node.args.inputs[0], values.TupleValue) and node.args.inputs[0].has_constant_value():
vs = []
for v in xs.value.get_constant_value():
v_ = oc.ONNXValue(onnx_graph, v)
vs.append(v_)
onnx_graph.add_node(
"Concat",
vs,
node.outputs,
str(node.lineprop),
axis=axis,
)
elif isinstance(node.args.inputs[0], values.ListValue):
onnx_graph.add_node(
"ChainerSequenceConcat",
[xs.create_sequence()],
node.outputs,
str(node.lineprop),
axis=axis,
)
def convert_matmul(onnx_graph, node):
a = oc.ONNXValue(onnx_graph,node.args.keywords['a'])
b = oc.ONNXValue(onnx_graph,node.args.keywords['b'])
transa = oc.try_get_attribute(node.args.keywords['transa'])
transb = oc.try_get_attribute(node.args.keywords['transb'])
assert not transa # TODO(hamaji): Not supported yet.
assert not transb # TODO(hamaji): Not supported yet.
onnx_graph.add_node(
"MatMul",
[a.create_tensor(), b.create_tensor()],
node.outputs,
str(node.lineprop),
)
def convert_reshape(onnx_graph, node):
onnx_graph.add_node(
"Reshape",
[node.inputs[0],oc.ONNXValue(onnx_graph,node.inputs[1]).create_tensor()],
node.outputs,
str(node.lineprop))
def convert_split_axis(onnx_graph, node):
force_tuple = oc.try_get_attribute(node.args.keywords['force_tuple'])
assert(force_tuple is True) # TODO(hamaji): Not supported yet.
onnx_graph.add_node(
"ChainerSequenceSplitAxis",
[node.inputs[0],oc.ONNXValue(onnx_graph,node.args.keywords['indices_or_sections']).create_tensor()],
node.outputs,
str(node.lineprop),
axis = oc.try_get_attribute(node.args.keywords['axis']))
def convert_broadcast_to(onnx_graph, node):
node_ = node
shape = oc.ONNXValue(onnx_graph, node_.args.keywords['shape'])
onnx_graph.add_node(
"Expand",
[node_.inputs[0], shape.create_tensor()],
node_.outputs,
str(node.lineprop))
return
def convert_dropout(onnx_graph, node):
x = oc.ONNXValue(onnx_graph,node.args.keywords['x'])
ratio = oc.try_get_attribute(node.args.keywords['ratio'])
onnx_graph.add_node(
"Dropout",
[x],
node.outputs,
str(node.lineprop),
ratio=ratio,
)
def convert_roi_average_pooling_2d(onnx_graph, node):
x = oc.ONNXValue(onnx_graph,node.args.keywords['x'])
rois = oc.ONNXValue(onnx_graph,node.args.keywords['rois'])
roi_indices = oc.ONNXValue(onnx_graph,node.args.keywords['roi_indices'])
outsize = oc.ONNXValue(onnx_graph,node.args.keywords['outsize'])
spatial_scale = oc.ONNXValue(onnx_graph,node.args.keywords['spatial_scale'])
def _pair(x):
if isinstance(x, collections.Iterable):
return x
return (x, x)
onnx_graph.add_node(
"ChainerROIAveragePool2D",
[x.create_tensor(), rois.create_tensor(), roi_indices.create_tensor()],
node.outputs,
str(node.lineprop),
output_shape=_pair(oc.try_get_attribute(outsize.value)),
spatial_scale=oc.try_get_attribute(spatial_scale.value))
return
def convert_concat(onnx_graph, node):
xs = oc.ONNXValue(onnx_graph,node.args.keywords['xs'])
axis = oc.try_get_attribute(node.attribute_args.keywords['axis'])
onnx_graph.add_node(
"ChainerSequenceConcat",
[xs.create_sequence()],
node.outputs,
str(node.lineprop),
axis=axis,
)
def convert_onnx_chainer_convolution2d(onnx_graph: 'ONNXGraph',
node: 'nodes.NodeCall'):
chainer_inst = node.func.owner.inst # type: chainer.links.Convolution2D
ksize = oc.size2d(chainer_inst.ksize)
stride = oc.size2d(chainer_inst.stride)
ps = oc.size2d(chainer_inst.pad)
pads = ps + ps
x = oc.ONNXValue(onnx_graph, node.args.get_value('x'))
o = oc.ONNXValue(onnx_graph, node.outputs[0])
w = oc.ONNXValue(onnx_graph, chainer_inst.W)
b = None
if chainer_inst.b is not None:
b = oc.ONNXValue(onnx_graph, chainer_inst.b)
onnx_graph.add_node('Conv', [x, w] + ([] if b is None else [b]), [o],
str(node.lineprop),
kernel_shape=ksize,
pads=pads,
strides=stride)
def convert_onnx_chainer_linear(onnx_graph: 'ONNXGraph',
node: 'nodes.NodeCall'):
chainer_inst = node.func.owner.inst # type: chainer.links.Linear
onnx_name = oc.node2onnx_parameter[node].onnx_name
x = oc.ONNXValue(onnx_graph, node.args.get_value('x'))
o = oc.ONNXValue(onnx_graph, node.outputs[0])
if chainer_inst.W.data is None:
print("W is unknown. Please infer this model.")
w = oc.ONNXValue(onnx_graph, chainer_inst.W)
(x_shape, ) = onnx_graph.add_node('Shape', [x], [None], str(node.lineprop))
(batch_size_1, ) = onnx_graph.add_node('Gather', [
x_shape,
oc.ONNXValue(onnx_graph, np.array(0, dtype=np.int64),
[onnx_name, '/Zero'])
], [None], str(node.lineprop))
(batch_size_2, ) = onnx_graph.add_node('Unsqueeze', [batch_size_1], [None],
str(node.lineprop),
axes=[0])
(mat_shape, ) = onnx_graph.add_node('Concat', [
batch_size_2,
oc.ONNXValue(onnx_graph, np.array([-1], dtype=np.int64),
[onnx_name, '/Minus1'])
], [None],
str(node.lineprop),
axis=0)
(x_reshape, ) = onnx_graph.add_node('Reshape', [x, mat_shape], [None],
str(node.lineprop))
if chainer_inst.b is not None:
b = oc.ONNXValue(onnx_graph, chainer_inst.b)
onnx_graph.add_node('Gemm', [x_reshape, w, b], [o],
str(node.lineprop),
transA=0,
transB=1)
else:
temp = oc.ONNXValue(onnx_graph, np.float32, [onnx_name, '/Temp'])
onnx_graph.add_node('Transpose', [w], [temp],
str(node.lineprop),
perm=[1, 0])
onnx_graph.add_node('MatMul', [x_reshape, temp], [o],
str(node.lineprop))
def convert_unpooling_2d(onnx_graph, node : 'nodes.NodeCall'):
ksize = oc.try_get_attribute(node.args.keywords['ksize'])
stride = oc.try_get_attribute(node.args.keywords['stride'])
pad = oc.try_get_attribute(node.args.keywords['pad'])
outsize = oc.try_get_attribute(node.args.keywords['outsize'])
cover_all = oc.try_get_attribute(node.args.keywords['cover_all'])
assert(stride is None) # TODO(hamaji): Not supported yet.
assert(pad == 0) # TODO(hamaji): Not supported yet.
assert(outsize is None) # TODO(hamaji): Not supported yet.
assert(cover_all is False) # TODO(hamaji): Not supported yet.
scales = np.array([1, 1] + list(_pair(ksize)), dtype=np.float32)
scales_ = oc.ONNXValue(onnx_graph, scales, [node, '/Scale'], is_constant = True)
onnx_graph.add_node(
"Upsample",
[node.inputs[0], scales_],
[node.outputs[0]],
name=str(node.lineprop))
def convert_onnx_chainer_batch_normalization(onnx_graph: 'ONNXGraph',
node: 'nodes.NodeCall'):
chainer_inst = node.func.owner.inst # type: chainer.links.BatchNormalization
assert (chainer_inst.axis is None) # not support yet
x = oc.ONNXValue(onnx_graph, node.args.get_value('x'))
o = oc.ONNXValue(onnx_graph, node.outputs[0])
gamma = oc.ONNXValue(onnx_graph, chainer_inst.gamma)
beta = oc.ONNXValue(onnx_graph, chainer_inst.beta)
avg_mean = oc.ONNXValue(onnx_graph, chainer_inst.avg_mean, [node, 'mean'])
avg_var = oc.ONNXValue(onnx_graph, chainer_inst.avg_var, [node, 'var'])
eps = chainer_inst.eps
momentum = chainer_inst.decay
onnx_graph.add_node('BatchNormalization',
[x, gamma, beta, avg_mean, avg_var], [o],
str(node.lineprop),
epsilon=eps,
momentum=momentum)