def verify_pad(indata, pads, value=0.0):
indata = np.array(indata).astype(np.float32)
# numpy expect result
len_dim = len(pads) // 2
np_pads = [(pads[i], pads[i+len_dim]) for i in range(len_dim)]
outdata = np.pad(indata, pad_width=np_pads, mode='constant', constant_values=value)
# onnx graph
node = helper.make_node(
'Pad',
inputs=['input'],
outputs=['output'],
mode='constant',
pads=pads,
value=value
)
graph = helper.make_graph([node],
'pad_test',
inputs = [helper.make_tensor_value_info("input",
TensorProto.FLOAT, list(indata.shape))],
outputs = [helper.make_tensor_value_info("output",
TensorProto.FLOAT, list(outdata.shape))])
model = helper.make_model(graph, producer_name='pad_test')
# tvm result
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, indata, target, ctx, outdata.shape, 'float32')
tvm.testing.assert_allclose(outdata, tvm_out, rtol=1e-5, atol=1e-5)
def test_onnx_to_caffe2_loop(self):
body_nodes = [helper.make_node(
"MatMul", ["_X", "W"], ["_Y"])]
nodes = self._make_fake_loop_op(body_nodes,
[(TensorProto.FLOAT, (2, 2), "X")],
[(TensorProto.FLOAT, (2, 2), "Y")])
X = np.random.rand(2, 2).astype(np.float32)
W = np.random.rand(2, 2).flatten().astype(np.float32)
graph_def = helper.make_graph(
nodes,
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, (2, 2)),
helper.make_tensor_value_info("W", TensorProto.FLOAT, (2, 2))],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, (2, 2))],
initializer=[helper.make_tensor("W",
TensorProto.FLOAT,
[2, 2],
W.tolist())]
)
model_def = helper.make_model(graph_def, producer_name='onnx-to-caffe2-test')
Y = X
for _ in range(10):
Y = np.matmul(Y, W.reshape(2, 2))
p = c2.prepare(model_def)
out = p.run(X)
np.testing.assert_allclose(out.Y, Y)
def _test_upsample_bilinear_opset9():
scale = 2
in_shape = (1, 1, 3, 3)
out_shape = (1, 1, 3*scale, 3*scale)
y = helper.make_node("Upsample", ['in','scales'], ['out'], mode='linear')
scales=[1.0, 1.0, 2.0, 2.0]
in_array = np.random.uniform(size=in_shape).astype(np.float32)
out_array = topi.testing.bilinear_resize_python(in_array, (3*scale, 3*scale), "NCHW")
ref_array = np.array(scales)
ref_node = helper.make_node('Constant',
inputs=[],
outputs=['scales'],
value=onnx.helper.make_tensor(name = 'const_tensor',
data_type = TensorProto.FLOAT,
dims = ref_array.shape,
vals = ref_array.flatten().astype(float)))
graph = helper.make_graph([ref_node, y],
'upsample_bilinear_opset9_test',
inputs = [helper.make_tensor_value_info("in", TensorProto.FLOAT, list(in_shape))],
outputs = [helper.make_tensor_value_info("out", TensorProto.FLOAT, list(out_shape))])
model = helper.make_model(graph, producer_name='upsample_bilinear_opset9_test')
inputs = []
inputs.append(in_array)
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, inputs, target, ctx, out_shape, 'float32')
tvm.testing.assert_allclose(out_array, tvm_out, rtol=1e-5, atol=1e-5)
def _test_power_iteration(x_shape, y_shape):
if isinstance(y_shape, int):
y_shape = [y_shape]
x = np.random.uniform(size=x_shape).astype(np.float32)
y = np.random.uniform(size=y_shape).astype(np.float32)
np_res = np.power(x, y).astype(np.float32)
res = helper.make_node("Pow", ['x', 'y'], ['out'])
graph = helper.make_graph([res],
'power_test',
inputs = [helper.make_tensor_value_info("x",
TensorProto.FLOAT, list(x_shape)),
helper.make_tensor_value_info("y",
TensorProto.FLOAT, list(y_shape))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(np_res.shape))])
model = helper.make_model(graph, producer_name='power_test')
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, [x, y], target, ctx, np_res.shape)
tvm.testing.assert_allclose(np_res, tvm_out, rtol=1e-5, atol=1e-5)
def test_spacetodepth():
n, c, h, w = shape = (1, 1, 4, 6)
input1 = np.random.rand(n, c, h, w).astype("float32")
blocksize = 2
inputs = [helper.make_tensor_value_info("input1", TensorProto.FLOAT, shape=shape)]
outputs = [helper.make_tensor_value_info("output", TensorProto.FLOAT, shape=(1, 4, 2, 3))]
nodes = [helper.make_node("SpaceToDepth", ["input1"], ["output"], block_size=blocksize)]
graph = helper.make_graph(nodes,
"spacetodepth_test",
inputs,
outputs)
spacetodepth_model = helper.make_model(graph)
bkd_rep = backend.prepare(spacetodepth_model)
output = bkd_rep.run([input1])
tmp = np.reshape(input1, [n, c,
h // blocksize, blocksize,
w // blocksize, blocksize])
tmp = np.transpose(tmp, [0, 3, 5, 1, 2, 4])
numpy_op = np.reshape(tmp, [n, c * (blocksize**2),
h // blocksize,
w // blocksize])
npt.assert_almost_equal(output[0], numpy_op)
def verify_mean(input_dim):
dtype = 'float32'
a_np1 = np.random.uniform(size=input_dim).astype(dtype)
a_np2 = np.random.uniform(size=input_dim).astype(dtype)
a_np3 = np.random.uniform(size=input_dim).astype(dtype)
b_np = np.mean((a_np1, a_np2, a_np3), axis=0)
mean_node = helper.make_node("Mean", ["a_np1", "a_np2", "a_np3"], ["out"])
graph = helper.make_graph([mean_node],
"Mean_test",
inputs = [helper.make_tensor_value_info("a_np1",
TensorProto.FLOAT, list(input_dim)),
helper.make_tensor_value_info("a_np2",
TensorProto.FLOAT, list(input_dim)),
helper.make_tensor_value_info("a_np3",
TensorProto.FLOAT, list(input_dim))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(b_np.shape))])
model = helper.make_model(graph, producer_name='Mean_test')
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, [a_np1, a_np2, a_np3], target, ctx, b_np.shape)
tvm.testing.assert_allclose(b_np, tvm_out, rtol=1e-5, atol=1e-5)
def verify_reduce_x(name, indata, axis, keepdims):
indata = np.array(indata).astype(np.float32)
# numpy expect result
if name == 'ReduceMax':
outdata = np.maximum.reduce(indata, axis=axis, keepdims=keepdims == 1)
elif name == 'ReduceMin':
outdata = np.minimum.reduce(indata, axis=axis, keepdims=keepdims == 1)
elif name == 'ReduceSum':
outdata = np.sum(indata, axis=axis, keepdims=keepdims == 1)
elif name == 'ReduceMean':
outdata = np.mean(indata, axis=axis, keepdims=keepdims == 1)
else:
raise Exception('unsupport op: {}'.format(name))
if len(np.asarray(outdata).shape) == 0:
outdata = np.asarray([outdata])
# onnx graph
if axis is None:
node = helper.make_node(name, inputs=['input'], outputs=['output'],
keepdims=keepdims)
else:
node = helper.make_node(name, inputs=['input'], outputs=['output'],
axis=axis, keepdims=keepdims)
graph = helper.make_graph([node],
'{}_test'.format(name),
inputs = [helper.make_tensor_value_info("input",
TensorProto.FLOAT, list(indata.shape))],
outputs = [helper.make_tensor_value_info("output",
TensorProto.FLOAT, list(outdata.shape))])
model = helper.make_model(graph, producer_name='{}_test'.format(name))
# tvm result
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, indata, target, ctx, outdata.shape, 'float32')
tvm.testing.assert_allclose(outdata, tvm_out, rtol=1e-5, atol=1e-5)
def test_onnx_to_caffe2_zipfile(self):
buf = tempfile.NamedTemporaryFile()
onnx_model = zipfile.ZipFile(buf, 'w')
output = tempfile.NamedTemporaryFile()
init_net_output = tempfile.NamedTemporaryFile()
node_def = helper.make_node(
"MatMul", ["X", "W"], ["Y"])
X = np.random.rand(2, 3).astype(np.float32)
W = np.random.rand(3, 2).flatten().astype(np.float32)
graph_def = helper.make_graph(
[node_def],
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, (2, 3)),
helper.make_tensor_value_info("W", TensorProto.FLOAT, (3, 2))],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, (2, 2))],
initializer=[helper.make_tensor("W",
TensorProto.FLOAT,
[3, 2],
b'__EXTERNAL',
raw=True)])
model_def = helper.make_model(graph_def, producer_name='onnx-to-caffe2-test')
onnx_model.writestr('__MODEL_PROTO', model_def.SerializeToString())
onnx_model.writestr('W', W.tobytes())
onnx_model.close()
W = W.reshape((3, 2))
Y_expect = np.matmul(X, W)
c2_model = c2.prepare_zip_archive(buf)
Y = c2_model.run(X).Y
np.testing.assert_allclose(Y, Y_expect)
def test_model_docstring(self): # type: () -> None
graph = helper.make_graph([], "my graph", [], [])
model_def = helper.make_model(graph, doc_string='test')
# models may have their own documentation, but don't have a name
# their name is the domain-qualified name of the underlying graph.
self.assertFalse(hasattr(model_def, "name"))
self.assertEqual(model_def.doc_string, 'test')
def verify_split(indata, outdatas, split, axis=0):
indata = np.array(indata).astype(np.float32)
outdatas = [np.array(o).astype(np.float32) for o in outdatas]
node = helper.make_node(
'Split',
inputs=['input'],
outputs=['output_{}'.format(i) for i in range(len(split))],
axis=axis,
split=split
)
graph = helper.make_graph([node],
'split_test',
inputs = [helper.make_tensor_value_info("input",
TensorProto.FLOAT, list(indata.shape))],
outputs = [helper.make_tensor_value_info("output_{}".format(i),
TensorProto.FLOAT, list(outdatas[i].shape))
for i in range(len(split))
])
model = helper.make_model(graph, producer_name='split_test')
for target, ctx in ctx_list():
output_shape = [o.shape for o in outdatas]
output_type = ['float32', 'float32', 'float32']
tvm_out = get_tvm_output(model, indata, target, ctx, output_shape, output_type)
for o, t in zip(outdatas, tvm_out):
tvm.testing.assert_allclose(o, t)
def test_initializer(self):
X = np.array([[1, 2], [3, 4]]).astype(np.float32)
Y = np.array([[1, 2], [3, 4]]).astype(np.float32)
weight = np.array([[1, 0], [0, 1]])
graph_def = make_graph(
[make_node("Add", ["X", "Y"], ["Z0"]),
make_node("Cast", ["Z0"], ["Z"], to="float"),
make_node("Mul", ["Z", "weight"], ["W0"]),
make_node("Tanh", ["W0"], ["W1"]),
make_node("Sigmoid", ["W1"], ["W2"]),
make_node("Scale", ["W2"], ["W3"], scale=-1.0)],
name="test_initializer",
inputs=[
make_tensor_value_info("X", onnx.TensorProto.FLOAT, (2, 2)),
make_tensor_value_info("Y", onnx.TensorProto.FLOAT, (2, 2)),
make_tensor_value_info("weight", onnx.TensorProto.FLOAT, (2, 2)),
],
outputs=[
make_tensor_value_info("W3", onnx.TensorProto.FLOAT, (2, 2))
],
initializer=[make_tensor("weight",
onnx.TensorProto.FLOAT,
[2, 2],
weight.flatten().astype(float))]
)
def sigmoid(x):
return 1 / (1 + np.exp(-x))
W_ref = -sigmoid(np.tanh((X + Y) * weight))
c2_rep = c2.prepare(make_model(graph_def, producer_name='caffe2-ref-test'))
output = c2_rep.run({"X": X, "Y": Y})
np.testing.assert_almost_equal(output["W3"], W_ref)
def test_small_model(self):
# Create one input
X = helper.make_tensor_value_info('IN', TensorProto.FLOAT, [2, 3])
# Create one output
Y = helper.make_tensor_value_info('OUT', TensorProto.FLOAT, [2, 3])
# Create a node
node_def = helper.make_node('Abs', ['IN'], ['OUT'])
# Create the model
graph_def = helper.make_graph([node_def], "test-model", [X], [Y])
onnx_model = helper.make_model(graph_def,
producer_name='onnx-example')
model = Model()
model.BuildFromOnnxModel(onnx_model)
schedule = model.OptimizeSchedule()
schedule = schedule.replace('\n', ' ')
expected_schedule = r'// Target: .+// MachineParams: .+// Delete this line if not using Generator Pipeline pipeline = get_pipeline\(\);.+Func OUT = pipeline.get_func\(1\);.+{.+}.+'
self.assertRegex(schedule, expected_schedule)
input = np.random.rand(2, 3) - 0.5
outputs = model.run([input])
self.assertEqual(1, len(outputs))
output = outputs[0]
expected = np.abs(input)
np.testing.assert_allclose(expected, output)
def test_scalars(self):
# Create 2 inputs
X = helper.make_tensor_value_info('A', TensorProto.INT32, [])
Y = helper.make_tensor_value_info('B', TensorProto.INT32, [])
# Create one output
Z = helper.make_tensor_value_info('C', TensorProto.INT32, [])
# Create a node
node_def = helper.make_node('Add', ['A', 'B'], ['C'])
# Create the model
graph_def = helper.make_graph([node_def], "scalar-model", [X, Y], [Z])
onnx_model = helper.make_model(graph_def,
producer_name='onnx-example')
model = Model()
model.BuildFromOnnxModel(onnx_model)
schedule = model.OptimizeSchedule()
schedule = schedule.replace('\n', ' ')
expected_schedule = r'// Target: .+// MachineParams: .+// Delete this line if not using Generator Pipeline pipeline = get_pipeline\(\);.+Func C = pipeline.get_func\(2\);.+{.+}.+'
self.assertRegex(schedule, expected_schedule)
input1 = np.random.randint(-10, 10, size=())
input2 = np.random.randint(-10, 10, size=())
outputs = model.run([input1, input2])
self.assertEqual(1, len(outputs))
output = outputs[0]
expected = input1 + input2
np.testing.assert_allclose(expected, output)
def test_reshape_like():
in_shape = (4, 3, 3, 4)
ref_shape = (3, 4, 4, 3)
ref_array = np.random.uniform(size=ref_shape).astype('float32')
ref_node = onnx.helper.make_node('Constant',
inputs=[],
outputs=['ref_in'],
value=onnx.helper.make_tensor(name = 'const_tensor',
data_type = onnx.TensorProto.FLOAT,
dims = ref_array.shape,
vals = ref_array.flatten().astype(float)))
copy_node = helper.make_node("Identity", ["ref_in"], ["copy_in"])
reshape_node = helper.make_node("Reshape", ["in", "copy_in"], ["out"])
graph = helper.make_graph([ref_node, copy_node, reshape_node],
"reshape_like_test",
inputs = [helper.make_tensor_value_info("in",
TensorProto.FLOAT, list(in_shape))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(ref_shape))])
model = helper.make_model(graph, producer_name='reshape_like_test')
for target, ctx in ctx_list():
x = np.random.uniform(size=in_shape).astype('float32')
tvm_out = get_tvm_output(model, x, target, ctx, ref_shape, 'float32')
tvm.testing.assert_allclose(ref_shape, tvm_out.shape)
def verify_constantfill(is_shape, input_dim, out_dim, value, dtype, **kwargs):
input_a = np.random.uniform(size=input_dim).astype(dtype)
out = np.empty(shape=out_dim, dtype=dtype)
out.fill(value)
if is_shape == True:
fill_node = helper.make_node("ConstantFill", [], ["out"], shape=input_dim, value=value, **kwargs)
else:
fill_node = helper.make_node("ConstantFill", ["input_a"], ["out"], value=value, dtype=dtype, **kwargs)
graph = helper.make_graph([fill_node],
"fill_test",
inputs = [helper.make_tensor_value_info("input_a",
TensorProto.FLOAT, list(input_dim))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(out.shape))])
model = helper.make_model(graph, producer_name='fill_test')
for target, ctx in ctx_list():
if is_shape == True:
tvm_out = get_tvm_output(model, [], target, ctx, out.shape)
else:
tvm_out = get_tvm_output(model, [input_a], target, ctx, out.shape)
tvm.testing.assert_allclose(out, tvm_out, rtol=1e-5, atol=1e-5)
def _optimized(self, graph):
orig_model = helper.make_model(graph, producer_name='onnx-to-caffe2-test')
orig_model_str = orig_model.SerializeToString()
optimized_model_str = c2.Caffe2Backend.optimize_onnx(orig_model_str)
optimized_model = ModelProto()
optimized_model.ParseFromString(optimized_model_str)
return optimized_model
def verify_lrn(shape, nsize, dtype, alpha=None, beta=None, bias=None):
in_array = np.random.uniform(size=shape).astype(dtype)
if alpha == None and beta == None and bias==None:
alpha = 0.0001
beta = 0.75
bias = 1.0
node = onnx.helper.make_node('LRN', inputs=['in'], outputs=['out'], size=nsize)
else:
node = onnx.helper.make_node('LRN', inputs=['in'], outputs=['out'], alpha=alpha,
beta=beta, bias=bias, size=nsize)
graph = helper.make_graph([node],
"lrn_test",
inputs = [helper.make_tensor_value_info("in", TensorProto.FLOAT, list(shape))],
outputs = [helper.make_tensor_value_info("out", TensorProto.FLOAT, list(shape))])
model = helper.make_model(graph, producer_name='lrn_test')
def _get_python_lrn():
square_sum = np.zeros(shape).astype(dtype)
for n, c, h, w in np.ndindex(in_array.shape):
square_sum[n, c, h, w] = sum(in_array[n,
max(0, c - int(math.floor((nsize - 1) / 2))): \
min(5, c + int(math.ceil((nsize - 1) / 2)) + 1),
h,
w] ** 2)
py_out = in_array / ((bias + (alpha / nsize) * square_sum) ** beta)
return py_out
for target, ctx in ctx_list():
input_name = model.graph.input[0].name
py_out = _get_python_lrn()
tvm_out = get_tvm_output(model, in_array, target, ctx, py_out.shape, 'float32')
tvm.testing.assert_allclose(py_out, tvm_out, rtol=1e-5, atol=1e-5)
def test_shape():
in_shape = (4, 3, 3, 4)
ref_shape = (6, 2, 4, 3)
ref_array = np.array(ref_shape)
ref_node = onnx.helper.make_node('Constant',
inputs=[],
outputs=['ref_in'],
value=onnx.helper.make_tensor(name = 'const_tensor',
data_type = onnx.TensorProto.INT32,
dims = ref_array.shape,
vals = ref_array.flatten().astype(int)))
reshape_node = helper.make_node("Reshape", ["in", "ref_in"], ["out"])
shape_node = helper.make_node("Shape", ['out'], ['final_out'])
graph = helper.make_graph([ref_node, reshape_node, shape_node],
"shape_test",
inputs = [helper.make_tensor_value_info("in",
TensorProto.FLOAT, list(in_shape))],
outputs = [helper.make_tensor_value_info("final_out",
TensorProto.FLOAT, list(ref_shape))])
model = helper.make_model(graph, producer_name='shape_test')
for target, ctx in ctx_list():
x = np.random.uniform(size=in_shape).astype('int32')
tvm_out = get_tvm_output(model, x, target, ctx, ref_shape, 'int32')
tvm.testing.assert_allclose(ref_shape, tvm_out)
def get_onnx_graph(testname, input_names, inputs, output_name, output_shape, attr):
outputs = [helper.make_tensor_value_info("output", TensorProto.FLOAT, shape=output_shape)]
nodes = [helper.make_node(output_name, input_names, ["output"], **attr)]
graph = helper.make_graph(nodes, testname, inputs, outputs)
model = helper.make_model(graph)
return model
def caffe2_net_to_onnx_model(cls, *args, **kwargs):
opset_id = OperatorSetIdProto()
opset_id.domain = '' # ONNX default domain
opset_id.version = cls.target_opset_version
model = make_model(cls.caffe2_net_to_onnx_graph(*args, **kwargs),
opset_imports=[opset_id], # current supported opset version
producer_name='onnx-caffe2', # producer name
)
checker.check_model(model)
return model
def test_check_model(self): # type: () -> None
node = helper.make_node(
"Relu", ["X"], ["Y"], name="test")
graph = helper.make_graph(
[node],
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 2])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 2])])
model = helper.make_model(graph, producer_name='test')
checker.check_model(model)
def test_ops(op_name, inputs, input_tensors, numpy_op):
outputs = [helper.make_tensor_value_info("output", TensorProto.FLOAT, shape=np.shape(inputs[0]))]
nodes = [helper.make_node(op_name, ["input"+str(i+1) for i in range(len(inputs))], ["output"])]
graph = helper.make_graph(nodes,
op_name + "_test",
input_tensors,
outputs)
model = helper.make_model(graph)
bkd_rep = backend.prepare(model)
output = bkd_rep.run(inputs)
npt.assert_almost_equal(output[0], numpy_op)
def test_model(self): # type: () -> None
node_def = helper.make_node(
"Relu", ["X"], ["Y"])
graph_def = helper.make_graph(
[node_def],
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 2])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 2])])
self.assertRaises(AttributeError, helper.make_model, graph_def, xxx=1)
model_def = helper.make_model(graph_def, producer_name='test')
self.assertEqual(model_def.producer_name, 'test')
def test_model_metadata_props(self): # type: () -> None
graph = helper.make_graph([], "my graph", [], [])
model_def = helper.make_model(graph, doc_string='test')
helper.set_model_props(model_def, {'Title': 'my graph', 'Keywords': 'test;graph'})
checker.check_model(model_def)
helper.set_model_props(model_def, {'Title': 'my graph', 'Keywords': 'test;graph'})
checker.check_model(model_def) # helper replaces, so no dupe
dupe = model_def.metadata_props.add()
dupe.key = 'Title'
dupe.value = 'Other'
self.assertRaises(checker.ValidationError, checker.check_model, model_def)
def verify_single_ops(op, x, out_np, rtol=1e-7, atol=1e-7):
z = helper.make_node(op, ['in1'], ['out'])
graph = helper.make_graph([z],
'_test',
inputs = [helper.make_tensor_value_info("in1",
TensorProto.FLOAT, list(in_shape)),],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(out_shape))])
model = helper.make_model(graph, producer_name='_test')
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, [x], target, ctx)
tvm.testing.assert_allclose(out_np, tvm_out, rtol=rtol, atol=atol)
def test_check_old_model(self): # type: () -> None
node = helper.make_node(
"Pad", ["X"], ["Y"], paddings=(0, 0, 0, 0))
graph = helper.make_graph(
[node],
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 2])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 2])])
onnx_id = helper.make_opsetid("", 1)
model = helper.make_model(graph, producer_name='test', opset_imports=[onnx_id])
checker.check_model(model)
def test_polish_model(self): # type: () -> None
node_def = helper.make_node(
"Relu", ["X"], ["Y"], doc_string="ABC")
graph_def = helper.make_graph(
[node_def],
"test",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 2])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 2])])
model_def = helper.make_model(graph_def, producer_name='test')
polished_def = onnx.utils.polish_model(model_def)
self.assertEqual(polished_def.producer_name, 'test')
self.assertEqual(len(polished_def.graph.node), 1)
self.assertFalse(polished_def.graph.node[0].HasField('doc_string'))
def run_node(cls, node, input, device='CPU'):
input_tensors = []
if len(node.input) != len(input):
raise ValueError(
"Unexpected Input Size: Op_Type = {0}, Expected = {1}, Received = {2}"
.format(node.op_type, len(node.input), len(input))
)
for i in range(len(input)):
input_tensors.append(helper.make_tensor_value_info(node.input[i], TensorProto.FLOAT, input[i].shape))
onnx_graph = helper.make_graph([node], "test_{}".format(node.op_type), input_tensors, [])
onnx_model = helper.make_model(onnx_graph)
return CNTKBackend.run_model(onnx_model, input, device)
def verify_binary_ops(op, x, y, out_np, broadcast=None, rtol=1e-7, atol=1e-7):
if broadcast is None:
z = helper.make_node(op, ['in1', 'in2'], ['out'])
else:
z = helper.make_node(op, ['in1', 'in2'], ['out'], broadcast=1)
graph = helper.make_graph([z],
'_test',
inputs = [helper.make_tensor_value_info("in1",
TensorProto.FLOAT, list(in_shape)),
helper.make_tensor_value_info("in2",
TensorProto.FLOAT, list(in_shape))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.FLOAT, list(out_shape))])
model = helper.make_model(graph, producer_name='_test')
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, [x, y], target, ctx)
tvm.testing.assert_allclose(out_np, tvm_out, rtol=rtol, atol=atol)
def verify_argmax(input_dim, axis=None, keepdims=None):
def _argmax_numpy(data, axis=0, keepdims=True):
result = np.argmax(data, axis=axis)
if (keepdims == 1):
result = np.expand_dims(result, axis)
return result.astype(data.dtype)
a_np1 = np.random.uniform(-10, 10, input_dim).astype(np.int32)
if keepdims is None and axis is None:
b_np = _argmax_numpy(a_np1)
node = onnx.helper.make_node('ArgMax',
inputs=['a_np1'],
outputs=['out'])
elif axis is None:
b_np = _argmax_numpy(a_np1, keepdims=keepdims)
node = onnx.helper.make_node('ArgMax',
inputs=['a_np1'],
outputs=['out'],
keepdims=keepdims)
elif keepdims is None:
b_np = _argmax_numpy(a_np1, axis=axis)
node = onnx.helper.make_node('ArgMax',
inputs=['a_np1'],
outputs=['out'],
axis=axis)
else:
b_np = _argmax_numpy(a_np1, axis=axis, keepdims=keepdims)
node = onnx.helper.make_node('ArgMax',
inputs=['a_np1'],
outputs=['out'],
axis=axis,
keepdims=keepdims)
graph = helper.make_graph([node],
"argmax_test",
inputs = [helper.make_tensor_value_info("a_np1",
TensorProto.INT32, list(a_np1.shape))],
outputs = [helper.make_tensor_value_info("out",
TensorProto.INT32, list(b_np.shape))])
model = helper.make_model(graph, producer_name='argmax_test')
for target, ctx in ctx_list():
tvm_out = get_tvm_output(model, [a_np1], target, ctx, b_np.shape, b_np.dtype)
tvm.testing.assert_allclose(b_np, tvm_out, rtol=1e-5, atol=1e-5)
def gen_gru_onnx_test_model(model_path,
seq_length,
batch_size,
hidden_size,
input_size,
direction,
has_bias,
has_sequence_lens,
has_initial_h,
linear_before_reset=False):
# Validate parameters
assert direction in GRU_DIRS, 'ONNX GRU direction invalid!'
assert not has_sequence_lens, 'ONNX GRU Variable sequence length not supported'
# Get number of directions
num_directions = 2 if (direction == GRU_DIR_BIDIRECTIONAL) else 1
# Tensor sizes
X_shape = [seq_length, batch_size, input_size]
W_shape = [num_directions, 3 * hidden_size, input_size]
R_shape = [num_directions, 3 * hidden_size, hidden_size]
B_shape = [num_directions, 6 * hidden_size]
sequence_lens_shape = [batch_size]
initial_h_shape = [num_directions, batch_size, hidden_size]
Y_shape = [seq_length, num_directions, batch_size, hidden_size]
# Generate random inputs (weights are assumed concatenated in ONNX format: z,r,h)
np.random.seed(1)
X = np.random.randn(*X_shape)
W = np.random.randn(*W_shape)
R = np.random.randn(*R_shape)
B = np.random.randn(*B_shape) if has_bias else np.zeros(B_shape)
sequence_lens = np.random.randint(
1, seq_length, batch_size) if has_sequence_lens else np.tile(
seq_length, batch_size)
initial_h = np.random.randn(
*initial_h_shape) if has_initial_h else np.zeros(initial_h_shape)
# Function to get all the weight components for the given direction
def get_weights(dir_idx):
Wz = np.reshape(W[dir_idx, 0 * hidden_size:1 * hidden_size, :],
[hidden_size, input_size])
Wr = np.reshape(W[dir_idx, 1 * hidden_size:2 * hidden_size, :],
[hidden_size, input_size])
Wh = np.reshape(W[dir_idx, 2 * hidden_size:3 * hidden_size, :],
[hidden_size, input_size])
Rz = np.reshape(R[dir_idx, 0 * hidden_size:1 * hidden_size, :],
[hidden_size, hidden_size])
Rr = np.reshape(R[dir_idx, 1 * hidden_size:2 * hidden_size, :],
[hidden_size, hidden_size])
Rh = np.reshape(R[dir_idx, 2 * hidden_size:3 * hidden_size, :],
[hidden_size, hidden_size])
bWz = np.reshape(B[dir_idx, 0 * hidden_size:1 * hidden_size],
[hidden_size])
bWr = np.reshape(B[dir_idx, 1 * hidden_size:2 * hidden_size],
[hidden_size])
bWh = np.reshape(B[dir_idx, 2 * hidden_size:3 * hidden_size],
[hidden_size])
bRz = np.reshape(B[dir_idx, 3 * hidden_size:4 * hidden_size],
[hidden_size])
bRr = np.reshape(B[dir_idx, 4 * hidden_size:5 * hidden_size],
[hidden_size])
bRh = np.reshape(B[dir_idx, 5 * hidden_size:6 * hidden_size],
[hidden_size])
return Wz, Wr, Wh, Rz, Rr, Rh, bWz, bWr, bWh, bRz, bRr, bRh
# Function to get PyTorch weights (which are in the r,z,h order)
def get_torch_weights(dir_idx):
Wz, Wr, Wh, Rz, Rr, Rh, bWz, bWr, bWh, bRz, bRr, bRh = get_weights(
dir_idx)
W_torch = np.concatenate((Wr, Wz, Wh), 0)
R_torch = np.concatenate((Rr, Rz, Rh), 0)
bW_torch = np.concatenate((bWr, bWz, bWh), 0)
bR_torch = np.concatenate((bRr, bRz, bRh), 0)
return (W_torch, R_torch, bW_torch, bR_torch)
# ----------------------------------------- COMPUTE pyTORCH REFERENCE ----------------------------------------------
# Compute reference using Pytorch. Pytorch GRU has only forward/bidirectional so we will do the reverse GRU using
# a Pytorch forward GRU.
gru = torch.nn.GRU(input_size=input_size,
hidden_size=hidden_size,
num_layers=1,
bias=True,
batch_first=False,
dropout=0,
bidirectional=(direction == GRU_DIR_BIDIRECTIONAL))
# Get GRU state dictionary
gru_state_dict = gru.state_dict()
# Assign forward weights
forwardEnabled = direction in [GRU_DIR_FORWARD, GRU_DIR_BIDIRECTIONAL]
if forwardEnabled:
forward_dir_idx = 0
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(forward_dir_idx)
gru_state_dict['weight_ih_l0'] = torch.tensor(W_torch,
dtype=torch.float32)
gru_state_dict['weight_hh_l0'] = torch.tensor(R_torch,
dtype=torch.float32)
gru_state_dict['bias_ih_l0'] = torch.tensor(bW_torch,
dtype=torch.float32)
gru_state_dict['bias_hh_l0'] = torch.tensor(bR_torch,
dtype=torch.float32)
# Assign reverse weights
reverseEnabled = direction in [GRU_DIR_REVERSE, GRU_DIR_BIDIRECTIONAL]
if reverseEnabled:
if direction == GRU_DIR_REVERSE:
reverse_dir_idx = 0
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(reverse_dir_idx)
gru_state_dict['weight_ih_l0'] = torch.tensor(W_torch,
dtype=torch.float32)
gru_state_dict['weight_hh_l0'] = torch.tensor(R_torch,
dtype=torch.float32)
gru_state_dict['bias_ih_l0'] = torch.tensor(bW_torch,
dtype=torch.float32)
gru_state_dict['bias_hh_l0'] = torch.tensor(bR_torch,
dtype=torch.float32)
else:
reverse_dir_idx = 1
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(reverse_dir_idx)
gru_state_dict['weight_ih_l0_reverse'] = torch.tensor(
W_torch, dtype=torch.float32)
gru_state_dict['weight_hh_l0_reverse'] = torch.tensor(
R_torch, dtype=torch.float32)
gru_state_dict['bias_ih_l0_reverse'] = torch.tensor(
bW_torch, dtype=torch.float32)
gru_state_dict['bias_hh_l0_reverse'] = torch.tensor(
bR_torch, dtype=torch.float32)
# Set GRU state dictionary
gru.load_state_dict(gru_state_dict, strict=True)
# Perform inference
X_torch = torch.tensor(X, dtype=torch.float32)
initial_h_torch = torch.tensor(initial_h, dtype=torch.float32)
if direction == GRU_DIR_REVERSE:
Y, next_h = gru(X_torch.flip([0]), initial_h_torch)
Y = Y.flip([0])
else:
Y, next_h = gru(X_torch, initial_h_torch)
# Reshape output to ONNX format [seq_length, num_directions, batch_size, hidden_size]
Y_ref = Y.detach().numpy()
Y_ref = np.reshape(Y_ref,
[seq_length, batch_size, num_directions, hidden_size])
Y_ref = np.transpose(Y_ref, [0, 2, 1, 3])
# Reshape states to ONNX format
Y_h_ref = next_h.detach().numpy()
# --------------------------------------- COMPUTE PYTHON-NUMPY REFERENCE -------------------------------------------
# Create X slices
Xslices = list()
for t in range(seq_length):
Xslices.append(np.reshape(X[t, :, :], [batch_size, input_size]))
# Function to compute one GRU cell
def compute_gru(forward):
dir_idx = 0 if forward else (0 if direction == GRU_DIR_REVERSE else 1)
Wz, Wr, Wh, Rz, Rr, Rh, bWz, bWr, bWh, bRz, bRr, bRh = get_weights(
dir_idx)
def f(x):
return (1 / (1 + np.exp(-x)))
def g(x):
return np.tanh(x)
def mm(x, w):
return np.matmul(x, w.transpose())
Ht = np.reshape(initial_h[dir_idx, :, :], [batch_size, hidden_size])
Yslices = list()
for t in range(seq_length):
xt = Xslices[t] if forward else Xslices[seq_length - 1 - t]
zt = f(mm(xt, Wz) + bWz + mm(Ht, Rz) + bRz)
rt = f(mm(xt, Wr) + bWr + mm(Ht, Rr) + bRr)
if linear_before_reset:
htild = g(mm(xt, Wh) + bWh + rt * (mm(Ht, Rh) + bRh))
else:
htild = g(mm(xt, Wh) + bWh + mm(rt * Ht, Rh) + bRh)
Ht = (1 - zt) * htild + zt * Ht
Yslices.append(Ht)
return Yslices, Ht
Yslices = list()
Hslices = list()
# Compute forward GRU
forwardYslices = list()
if forwardEnabled:
Yt, Ht = compute_gru(True)
forwardYslices += Yt
Hslices.append(Ht)
# Compute reverse GRU
reverseYslices = list()
if reverseEnabled:
Yt, Ht = compute_gru(False)
reverseYslices += Yt
Hslices.append(Ht)
# Concatenate slices
for t in range(seq_length):
if forwardEnabled:
Yslices.append(forwardYslices[t])
if reverseEnabled:
Yslices.append(reverseYslices[seq_length - 1 - t])
Y_ref_np = np.concatenate(Yslices, 0).reshape(
[seq_length, num_directions, batch_size, hidden_size])
Y_h_ref_np = np.concatenate(Hslices, 0).reshape(
[num_directions, batch_size, hidden_size])
# Use numpy implementation when linear_before_reset = False, else assert errors
if linear_before_reset is False:
Y_ref = Y_ref_np
Y_h_ref = Y_h_ref_np
else:
assert np.max(
np.abs(Y_ref - Y_ref_np)
) < 1e-6, "Mismatch between Pytorch and Numpy GRU implementation"
assert np.max(
np.abs(Y_h_ref - Y_h_ref_np)
) < 1e-6, "Mismatch between Pytorch and Numpy GRU implementation"
# ---------------------------------------------- NODE DEFINITION --------------------------------------------------
# Node inputs
node_inputs = [
'X', 'W', 'R', 'B' if has_bias else '', '',
'initial_h' if has_initial_h else ''
]
# Node outputs
node_outputs = ['Y', 'Y_h']
# GRU node definition
gru_node_def = onnx.helper.make_node(
'GRU',
name='gru',
inputs=node_inputs,
outputs=node_outputs,
hidden_size=hidden_size,
direction=direction,
linear_before_reset=linear_before_reset)
# Error node definition
err_node_def = onnx.helper.make_node('Sub',
name='error',
inputs=['Y', 'Y_ref'],
outputs=['Y_err'])
# --------------------------------------------- GRAPH DEFINITION --------------------------------------------------
graph_input = list()
graph_init = list()
graph_output = list()
# GRU inputs
graph_input.append(
helper.make_tensor_value_info('X', TensorProto.FLOAT, X_shape))
graph_input.append(
helper.make_tensor_value_info('W', TensorProto.FLOAT, W_shape))
graph_input.append(
helper.make_tensor_value_info('R', TensorProto.FLOAT, R_shape))
if has_bias:
graph_input.append(
helper.make_tensor_value_info('B', TensorProto.FLOAT, B_shape))
if has_sequence_lens:
graph_input.append(
helper.make_tensor_value_info('sequence_lens', TensorProto.INT32,
sequence_lens_shape))
if has_initial_h:
graph_input.append(
helper.make_tensor_value_info('initial_h', TensorProto.FLOAT,
initial_h_shape))
# Reference input
graph_input.append(
helper.make_tensor_value_info('Y_ref', TensorProto.FLOAT, Y_shape))
# GRU initializers
graph_init.append(make_init('X', TensorProto.FLOAT, X))
graph_init.append(make_init('W', TensorProto.FLOAT, W))
graph_init.append(make_init('R', TensorProto.FLOAT, R))
if has_bias:
graph_init.append(make_init('B', TensorProto.FLOAT, B))
if has_sequence_lens:
graph_init.append(
make_init('sequence_lens', TensorProto.INT32, sequence_lens))
if has_initial_h:
graph_init.append(make_init('initial_h', TensorProto.FLOAT, initial_h))
# Reference initializer
graph_init.append(make_init('Y_ref', TensorProto.FLOAT, Y_ref))
# Graph outputs
graph_output.append(
helper.make_tensor_value_info('Y_err', TensorProto.FLOAT, Y_shape))
# Define graph (GraphProto)
graph_name = 'gru_test'
graph_def = helper.make_graph([gru_node_def, err_node_def],
graph_name,
inputs=graph_input,
outputs=graph_output)
# Set initializers
graph_def.initializer.extend(graph_init)
# --------------------------------------------- MODEL DEFINITION --------------------------------------------------
# Define model (ModelProto)
model_def = helper.make_model(graph_def, producer_name='onnx-gru')
# Check model
onnx.checker.check_model(model_def)
# Print model
with open(model_path, 'w') as f:
f.write(str(model_def))
from onnx import helper, TensorProto
INPUT_1 = helper.make_tensor_value_info('input1', TensorProto.FLOAT, [1])
INPUT_2 = helper.make_tensor_value_info('input2', TensorProto.FLOAT, [1])
OUTPUT_1 = helper.make_tensor_value_info('output1', TensorProto.FLOAT, [1])
OUTPUT_2 = helper.make_tensor_value_info('output2', TensorProto.FLOAT, [1])
nodes = [
helper.make_node(
'Mul',
['input1', 'input2'],
['output1'],
),
helper.make_node(
'Add',
['input1', 'input2'],
['output2'],
),
]
graph_def = helper.make_graph(
nodes,
'add_mul',
[INPUT_1, INPUT_2],
[OUTPUT_1, OUTPUT_2],
)
model_def = helper.make_model(
graph_def,
producer_name='add_mul.py',
opset_imports=[onnx.OperatorSetIdProto(version=12)])
onnx.save(model_def, 'add_mul.onnx')
def construct_model_conv_squeezes(self,
output_model_path,
conv_input_shape,
conv_weight_shape,
conv_output_shape,
opset=13):
# (input)
# / | \
# Conv1 conv2 conv3
# | | |
# Squeeze1 Squeeze2 |
# \ / |
# add1 |
# | |
# Unsqueeze |
# \ |
# add2
# |
# (output)
input_tensor = helper.make_tensor_value_info('input',
TensorProto.FLOAT,
conv_input_shape)
conv1_weight_arr = np.random.randint(-1, 2, conv_weight_shape).astype(
np.float32)
conv1_weight_initializer = onnx.numpy_helper.from_array(
conv1_weight_arr, name='conv1_weight')
conv1_node = onnx.helper.make_node('Conv', ['input', 'conv1_weight'],
['conv1_output'],
name='conv1_node')
conv2_weight_arr = np.random.randint(-1, 2, conv_weight_shape).astype(
np.float32)
conv2_weight_initializer = onnx.numpy_helper.from_array(
conv2_weight_arr, name='conv2_weight')
conv2_node = onnx.helper.make_node('Conv', ['input', 'conv2_weight'],
['conv2_output'],
name='conv2_node')
conv3_weight_arr = np.random.randint(-1, 2, conv_weight_shape).astype(
np.float32)
conv3_weight_initializer = onnx.numpy_helper.from_array(
conv3_weight_arr, name='conv3_weight')
conv3_node = onnx.helper.make_node('Conv', ['input', 'conv3_weight'],
['conv3_output'],
name='conv3_node')
if (opset >= 13):
squeeze_axes_initializer = onnx.numpy_helper.from_array(
np.array([0], dtype=np.int64), name='squeeze_axes')
squeeze1_node = helper.make_node('Squeeze',
['conv1_output', 'squeeze_axes'],
['squeeze1_output'],
name='suqeeze1_node')
squeeze2_node = helper.make_node('Squeeze',
['conv2_output', 'squeeze_axes'],
['squeeze2_output'],
name='suqeeze2_node')
else:
squeeze1_node = helper.make_node('Squeeze', ['conv1_output'],
['squeeze1_output'],
name='suqeeze1_node',
axes=[0])
squeeze2_node = helper.make_node('Squeeze', ['conv2_output'],
['squeeze2_output'],
name='suqeeze2_node',
axes=[0])
add1_node = helper.make_node('Add',
['squeeze1_output', 'squeeze2_output'],
['add1_output'],
name='add1_node')
if (opset >= 13):
unsqueeze_node = helper.make_node('Unsqueeze',
['add1_output', 'squeeze_axes'],
['unsqueeze_output'],
name='unsqueeze_node')
else:
unsqueeze_node = helper.make_node('Unsqueeze', ['add1_output'],
['unsqueeze_output'],
name='unsqueeze_node',
axes=[0])
output_tensor = helper.make_tensor_value_info('output',
TensorProto.FLOAT,
conv_output_shape)
add2_node = helper.make_node('Add',
['unsqueeze_output', 'conv3_output'],
['output'],
name='add2_node')
initializers = [
conv1_weight_initializer, conv2_weight_initializer,
conv3_weight_initializer
]
if (opset >= 13):
initializers.append(squeeze_axes_initializer)
graph = helper.make_graph([
conv1_node, conv2_node, conv3_node, squeeze1_node, squeeze2_node,
add1_node, unsqueeze_node, add2_node
],
'TestOpSuqeezes_test_model', [input_tensor],
[output_tensor],
initializer=initializers)
model = helper.make_model(
graph, opset_imports=[helper.make_opsetid("", opset)])
model.ir_version = 7 # use stable onnx ir version
onnx.save(model, output_model_path)
pdf_temp_2,
max_score,
normalized_scores, # top_action_1, # top_action_2,
# one_hot_top_action_int, one_hot_top_action_float,
one_hot_top_action,
exploit_prob,
exploit_top_action,
pdf, # node_rand_sub,
node,
node_rand,
add_node,
clip_node,
topk_node
],
'compute_graph',
input_tensors,
output_tensors,
initializer_tensors)
model = oh.make_model(graph, producer_name='explore')
f = open("model1.onnx", "wb")
f.write(model.SerializeToString())
f.close()
tf_model = onnx.load('model1.onnx')
tf_rep = prepare(tf_model)
sample = tf_rep.run(np.asarray([[.3, .3, .4]]))
print(sample)
def main(model_path,
model_save_path=None,
add_transpose_for_channel_last_first_issue=True,
bottom_nodes_name=None):
onnx_weight_node_list = []
output_tensor_value_info = []
onnx_node_list = []
inner_node_shape_value_info = []
# parse node information through tflite interpreter (tflite interpreter can't parse operator information in our target tensorflow version 1.15)
interpreter = tf.lite.Interpreter(model_path)
interpreter.allocate_tensors()
# get model input info(assume there is only one input)
input_details = interpreter.get_input_details()
model_input_name = input_details[0]['name']
input_tensor_value_info = None
# generate tree
tree_graph = Tree(model_path=model_path,
bottom_nodes_name=bottom_nodes_name,
defused=True)
# get sequential node name
sequential_keys = tree_graph.get_sequential_nodes_key()
# get tree node in the form of {node_name: op_node_obj}
tree_dict = tree_graph.get_nodes()
#############################
# build head transpose node #
#############################
for h_node in tree_graph.get_head_nodes():
# transpose for channel last to channel first
if add_transpose_for_channel_last_first_issue is True:
logging.getLogger('tflite2onnx').debug(
"generating transpose node for channel last first issue: " +
h_node.node_name)
input_tensor_value_info = helper.make_tensor_value_info(
model_input_name, TensorProto.FLOAT,
h_node.node_input_shape.tolist())
h_transpose_node = build_head_transpose_node_for_channel_last_2_channel_first(
input_tensor_value_info.name, h_node.node_name)
onnx_node_list.append(h_transpose_node)
h_node.input_nodes_name = [h_transpose_node.name]
else:
input_tensor_value_info = helper.make_tensor_value_info(
model_input_name, TensorProto.FLOAT,
tflite_utils.tflite2onnx_shape_map(
h_node.node_input_shape.tolist()))
h_node.input_nodes_name = [input_tensor_value_info.name]
############################
# build model node by node #
############################
for key in sequential_keys:
logging.getLogger('tflite2onnx').debug("generating: " + key)
nodes, val, weight = tree_dict[key].generate()
if (len(val) != 0) and (tree_dict[key].is_bottom_node is False):
inner_node_shape_value_info.extend(val)
if len(weight) != 0:
onnx_weight_node_list.extend(weight)
if len(nodes) != 0:
onnx_node_list.extend(nodes)
# sometimes, there are sub-node in one tree node, we need to find the last one
b_nodes = [node for node in tree_graph.get_bottom_nodes()]
###############################
# build bottom transpose node #
###############################
for b_node in b_nodes:
out_value_info = None
if add_transpose_for_channel_last_first_issue is True:
logging.getLogger('tflite2onnx').debug(
"generating transpose node for channel last first issue: " +
b_node.node_name)
out_value_info, transpose_node = build_button_transpose_node_for_channel_first_2_channel_last(
b_node.node_list[-1], b_node.node_output_shape.tolist())
if transpose_node != None:
onnx_node_list.append(transpose_node)
else:
out_value_info = set_end_node(b_node.node_list[-1],
b_node.node_output_shape.tolist())
output_tensor_value_info.append(out_value_info)
input_init = [input_tensor_value_info]
input_init.extend(onnx_weight_node_list)
onnx_inputs = tflite_utils.make_kneron_valid_onnx_input(input_init)
graph_cnn = helper.make_graph(onnx_node_list,
'cnn_test',
onnx_inputs,
output_tensor_value_info,
onnx_weight_node_list,
value_info=inner_node_shape_value_info)
cnn_model = helper.make_model(graph_cnn, producer_name='Kneron')
cnn_model.opset_import[0].version = 11
# add generated time to model meta data
helper.set_model_props(
cnn_model, {
'Generated Time':
datetime.utcnow().strftime("%m/%d/%Y, %H:%M:%S") + " (UTC+0)"
})
cnn_model = onnx.utils.polish_model(cnn_model)
# save
if model_save_path is not None:
onnx.save(cnn_model, model_save_path)
return cnn_model
Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1, 5, 32, 32])
conv1 = helper.make_node(
'Conv', # name
['X', 'W'], # inputs
['Y'], # outputs
## Node Attributes
kernel_shape=[3, 3],
strides=[1, 1],
pads=[1, 1, 1, 1],
)
graph_def = helper.make_graph(
nodes=[conv1], # graph nodes
name= 'conv_model', # graph name
inputs = [X, W_info], # graph inputs
outputs = [Y], # graph outputs
initializer = [W],
)
model_def = helper.make_model(graph_def, producer_name='benchmarks')
onnx.checker.check_model(model_def)
model_def = shape_inference.infer_shapes(model_def)
onnx.checker.check_model(model_def)
model_def = optimizer.optimize(model_def)
onnx.checker.check_model(model_def)
onnx.save_model(model_def, 'e2_conv.onnx')
def _make_module(in_shape, kernel_output_channel, bias_shape, auto_pad_mode,
dilation, group, kernel_shape, pad, stride):
inputs = []
initializers = []
# input
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, in_shape)
inputs.append('input')
# weight
w_shape = []
w_shape.append(kernel_output_channel)
w_shape.append(in_shape[1])
w_shape.extend(kernel_shape)
weight = helper.make_tensor('weight',
TensorProto.FLOAT,
dims=w_shape,
vals=np.random.rand(*w_shape).astype(
np.float32).flatten().tolist())
inputs.append('weight')
initializers.append(weight)
# bias
if bias_shape is not None:
bias = helper.make_tensor('bias',
TensorProto.FLOAT,
dims=bias_shape,
vals=np.random.rand(*bias_shape).astype(
np.float32).flatten().tolist())
inputs.append('bias')
initializers.append(bias)
# dilation
d = [1, 1] if dilation is None else dilation
# stride
s = [1, 1] if stride is None else stride
# pad
padding = [0, 0, 0, 0]
if (auto_pad_mode is None
or auto_pad_mode == 'NOTSET') and pad is not None:
padding = pad
elif auto_pad_mode == 'SAME_UPPER':
output_h = math.floor((in_shape[2] + s[0] - 1) / s[0])
pad_nums = max(0, (output_h - 1) * s[0] + (w_shape[2] - 1) * d[0] + 1 -
in_shape[2])
padding[0] = math.floor(pad_nums / 2)
padding[2] = pad_nums - padding[0]
output_w = math.floor((in_shape[3] + s[1] - 1) / s[1])
pad_nums = max(0, (output_w - 1) * s[1] + (w_shape[3] - 1) * d[1] + 1 -
in_shape[3])
padding[1] = math.floor(pad_nums / 2)
padding[3] = pad_nums - padding[1]
elif auto_pad_mode == 'SAME_LOWER':
output_h = math.floor((in_shape[2] + s[0] - 1) / s[0])
pad_nums = max(0, (output_h - 1) * s[0] + (w_shape[2] - 1) * d[0] + 1 -
in_shape[2])
padding[0] = math.ceil(pad_nums / 2)
padding[2] = pad_nums - padding[0]
output_w = math.floor((in_shape[3] + s[1] - 1) / s[1])
pad_nums = max(0, (output_w - 1) * s[1] + (w_shape[3] - 1) * d[1] + 1 -
in_shape[3])
padding[1] = math.ceil(pad_nums / 2)
padding[3] = pad_nums - padding[1]
# output
out_shape = []
out_shape.append(in_shape[0])
out_shape.append(w_shape[0])
out_shape.append(
math.floor((in_shape[2] + padding[0] + padding[2] -
((w_shape[2] - 1) * d[0] + 1) + s[0]) / s[0]))
out_shape.append(
math.floor((in_shape[3] + padding[1] + padding[3] -
((w_shape[3] - 1) * d[1] + 1) + s[1]) / s[1]))
output = helper.make_tensor_value_info('output', TensorProto.FLOAT,
out_shape)
attributes_dict = {}
if auto_pad_mode is not None:
attributes_dict['auto_pad'] = auto_pad_mode
if dilation is not None:
attributes_dict['dilations'] = dilation
if group is not None:
attributes_dict['group'] = group
if kernel_shape is not None:
attributes_dict['kernel_shape'] = kernel_shape
if pad is not None:
attributes_dict['pads'] = padding
if stride is not None:
attributes_dict['strides'] = stride
node = onnx.helper.make_node('Conv',
inputs=inputs,
outputs=['output'],
**attributes_dict)
nodes = []
nodes.append(node)
graph_def = helper.make_graph(nodes,
'test-model', [input], [output],
initializer=initializers)
model_def = helper.make_model(graph_def, producer_name='kendryte')
return model_def
def create_const_net(self, const_shapes, ir_version, opset=None):
"""
ONNX net IR net
Inputs->Concat with Sum of consts->Output => Input->Concat with consts
"""
#
# Create ONNX model
#
from onnx import helper
from onnx import TensorProto
shape_len = 0
for shape in const_shapes:
if len(shape) > shape_len:
shape_len = len(shape)
input_shape = shape
concat_axis = 0
output_shape = input_shape.copy()
output_shape[concat_axis] *= 2
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
nodes = list()
input_names = list()
consts = list()
for i, shape in enumerate(const_shapes):
const = np.random.randint(-127, 127, shape).astype(np.float)
const_name = 'const{}'.format(i + 1)
nodes.append(helper.make_node(
'Constant',
inputs=[],
outputs=[const_name],
value=helper.make_tensor(
name='const_tensor',
data_type=TensorProto.FLOAT,
dims=const.shape,
vals=const.flatten(),
),
))
input_names.append(const_name)
consts.append(const)
nodes.append(helper.make_node(
'Sum',
inputs=input_names,
outputs=['sum']
))
nodes.append(helper.make_node(
'Concat',
inputs=['input', 'sum'],
outputs=['output'],
axis=concat_axis
))
# Create the graph (GraphProto)
graph_def = helper.make_graph(
nodes,
'test_model',
[input],
[output],
)
# Create the model (ModelProto)
args = dict(producer_name='test_model')
if opset:
args['opset_imports'] = [helper.make_opsetid("", opset)]
onnx_net = helper.make_model(graph_def, **args)
# Create reference IR net
ref_net = None
return onnx_net, ref_net
def create_net(self, dyn_shapes, const_shapes, precision, ir_version, opset=None):
"""
ONNX net IR net
Inputs->Sum with consts->Output => Input->Eltwise
"""
#
# Create ONNX model
#
from onnx import helper
from onnx import TensorProto
inputs = list()
input_names = list()
out_shape_len = 0
for i, shape in enumerate(dyn_shapes):
input_name = 'input{}'.format(i + 1)
inputs.append(helper.make_tensor_value_info(input_name, TensorProto.FLOAT, shape))
input_names.append(input_name)
if len(shape) > out_shape_len:
out_shape_len = len(shape)
output_shape = shape
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
nodes = list()
consts = list()
for i, shape in enumerate(const_shapes):
const = np.random.randint(-127, 127, shape).astype(np.float)
const_name = 'const{}'.format(i + 1)
nodes.append(helper.make_node(
'Constant',
inputs=[],
outputs=[const_name],
value=helper.make_tensor(
name='const_tensor',
data_type=TensorProto.FLOAT,
dims=const.shape,
vals=const.flatten(),
),
))
input_names.append(const_name)
consts.append(const)
nodes.append(helper.make_node(
'Sum',
inputs=input_names,
outputs=['output']
))
# Create the graph (GraphProto)
graph_def = helper.make_graph(
nodes,
'test_model',
inputs,
[output],
)
# Create the model (ModelProto)
args = dict(producer_name='test_model')
if opset:
args['opset_imports'] = [helper.make_opsetid("", opset)]
onnx_net = helper.make_model(graph_def, **args)
# Create reference IR net
ref_net = None
# Too complicated IR to generate by hand
return onnx_net, ref_net
def convert(args):
# Read the model files.
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(args.fluid_model, exe)
# Using blocks in programs, create nodes using:
onnx_nodes = []
# Load parameters
global_block = inference_program.global_block()
for var_name in global_block.vars:
var = global_block.var(var_name)
if var_name not in ['feed', 'fetch'] and var.persistable:
param = fluid.executor.fetch_var(var_name, inference_scope)
param_node = helper.make_node(
'Constant',
inputs=[],
outputs=[var_name],
value=helper.make_tensor(
name=var_name,
dims=var.shape,
data_type=PADDLE_TO_ONNX_DTYPE[var.dtype],
vals=param.flatten().tolist()))
onnx_nodes.append(param_node)
# Create inputs
inputs = [
paddle_variable_to_onnx_tensor(v, global_block)
for v in feed_target_names
]
# Create outputs
fetch_target_names = [
fetch_target.name for fetch_target in fetch_targets
]
outputs = [
paddle_variable_to_onnx_tensor(v, global_block)
for v in fetch_target_names
]
# Create nodes
for block in inference_program.blocks:
for op in block.ops:
if op.type in ops.node_maker:
# TODO(kuke): deal with the corner case that vars in
# different blocks have the same name
node_proto = ops.node_maker[op.type](
inputs=op.input_arg_names,
attrs=op.attr_names,
outputs=op.output_arg_names)
onnx_nodes.append(node_proto)
else:
if op.type not in ['feed', 'fetch']:
raise NotImplementedError("OP[%s] is not supported in "
"the converter!" % op.type)
# Make graph
model_name = os.path.basename(
args.fluid_model.strip('/')).split('.')[0]
onnx_graph = helper.make_graph(onnx_nodes, model_name, inputs, outputs)
# Make model
onnx_model = helper.make_model(onnx_graph,
producer_name='PaddlePaddle')
# Model check
checker.check_model(onnx_model)
# Output readable model
print("The converted model is:\n{}".format(onnx_model))
# Save converted model
if args.onnx_model is not None:
try:
with open(args.onnx_model, 'wb') as f:
f.write(onnx_model.SerializeToString())
print("Saved converted model to path: %s" % args.onnx_model)
except (IOError), e:
print("Invalid ONNX model saving path: %s" % args.onnx_model)
def test_cast_errors():
np.random.seed(133391)
input_data = np.ceil(np.random.rand(2, 3, 4) * 16)
# missing 'to' attribute
node = onnx.helper.make_node('Cast', inputs=['A'], outputs=['B'])
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT16, value.shape)
for name, value in zip(node.output, ())
] # type: ignore
graph = make_graph([node], 'compute_graph', input_tensors, output_tensors)
model = make_model(graph, producer_name='NgraphBackend')
with pytest.raises(RuntimeError):
import_onnx_model(model)
# unsupported data type representation
node = onnx.helper.make_node('Cast',
inputs=['A'],
outputs=['B'],
to=1.2345)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(name, onnx.TensorProto.INT32, value.shape)
for name, value in zip(node.output, ())
] # type: ignore
graph = make_graph([node], 'compute_graph', input_tensors, output_tensors)
model = make_model(graph, producer_name='NgraphBackend')
with pytest.raises(RuntimeError):
import_onnx_model(model)
# unsupported input tensor data type:
node = onnx.helper.make_node('Cast',
inputs=['A'],
outputs=['B'],
to=onnx.TensorProto.INT32)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.COMPLEX64, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(name, onnx.TensorProto.INT32, value.shape)
for name, value in zip(node.output, ())
] # type: ignore
graph = make_graph([node], 'compute_graph', input_tensors, output_tensors)
model = make_model(graph, producer_name='NgraphBackend')
with pytest.raises((RuntimeError, NgraphTypeError)):
import_onnx_model(model)
# unsupported output tensor data type:
node = onnx.helper.make_node('Cast',
inputs=['A'],
outputs=['B'],
to=onnx.TensorProto.COMPLEX128)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(name, onnx.TensorProto.COMPLEX128, value.shape)
for name, value in zip(node.output, ())
] # type: ignore
graph = make_graph([node], 'compute_graph', input_tensors, output_tensors)
model = make_model(graph, producer_name='NgraphBackend')
with pytest.raises(RuntimeError):
import_onnx_model(model)
def construct_model_gemm(self, output_model_path):
# (input)
# |
# GEMM
# |
# Clip
# |
# GEMM
# |
# (output)
input_name = "input"
output_name = "output"
initializers = []
def make_gemm(input_name, weight_shape, weight_name, bias_shape,
bias_name, output_name):
weight_data = np.random.normal(0, 0.1,
weight_shape).astype(np.float32)
initializers.append(
onnx.numpy_helper.from_array(weight_data, name=weight_name))
bias_data = np.random.normal(0, 0.1, bias_shape).astype(np.float32)
initializers.append(
onnx.numpy_helper.from_array(bias_data, name=bias_name))
return onnx.helper.make_node(
"Gemm",
[input_name, weight_name, bias_name],
[output_name],
alpha=1.0,
beta=1.0,
transB=1,
)
# make gemm1 node
gemm1_output_name = "gemm1_output"
gemm1_node = make_gemm(
input_name,
[100, 10],
"linear1.weight",
[100],
"linear1.bias",
gemm1_output_name,
)
# make Clip
clip_min_name = "clip_min"
clip_max_name = "clip_max"
clip_output_name = "clip_output"
clip_inputs = [gemm1_output_name, clip_min_name, clip_max_name]
clip_outputs = [clip_output_name]
initializers.append(
onnx.numpy_helper.from_array(np.array(-1.0, dtype=np.float32),
name=clip_min_name))
initializers.append(
onnx.numpy_helper.from_array(np.array(1.0, dtype=np.float32),
name=clip_max_name))
clip_node = onnx.helper.make_node("Clip", clip_inputs, clip_outputs)
# make gemm2 node
gemm2_node = make_gemm(
clip_output_name,
[10, 100],
"linear2.weight",
[10],
"linear2.bias",
output_name,
)
# make graph
input_tensor = helper.make_tensor_value_info(input_name,
TensorProto.FLOAT,
[-1, 10])
output_tensor = helper.make_tensor_value_info(output_name,
TensorProto.FLOAT,
[-1, 10])
graph_name = "gemm_test"
graph = helper.make_graph(
[gemm1_node, clip_node, gemm2_node],
graph_name,
[input_tensor],
[output_tensor],
initializer=initializers,
)
model = helper.make_model(graph,
opset_imports=[helper.make_opsetid("", 13)])
model.ir_version = 7 # use stable onnx ir version
onnx.save(model, output_model_path)
def GenerateModel3(model_name):
nodes = [ # LayerNorm subgraph
helper.make_node("Shape", ["input_ids"], ["shape1_out"], "shape1"),
helper.make_node("Gather", ["shape1_out", "indices_0"],
["gather0_out"], "gather0"),
helper.make_node("Unsqueeze", ["gather0_out"], ["unsqueeze0_out"],
"unsqueeze0",
axes=[0]),
helper.make_node("Shape", ["input_ids"], ["shape2_out"], "shape2"),
helper.make_node("Gather", ["shape2_out", "indices_1"],
["gather1_out"], "gather1"),
helper.make_node("Unsqueeze", ["gather1_out"], ["unsqueeze1_out"],
"unsqueeze1",
axes=[0]),
helper.make_node("Concat", ["unsqueeze0_out", "unsqueeze1_out"],
["concat_out"],
"concat",
axis=0),
helper.make_node("Cast", ["gather1_out"], ["cast_out"], "cast", to=7),
helper.make_node("Range", ["start_0", "cast_out", "delta_1"],
["range_out"], "range"),
helper.make_node("Unsqueeze", ["range_out"], ["unsqueeze2_out"],
"unsqueeze2",
axes=[0]),
helper.make_node("Expand", ["unsqueeze2_out", "concat_out"],
["expand_out"], "expand"),
helper.make_node("Gather", ["pos_embed", "expand_out"],
["pos_gather_out"], "pos_gather"),
helper.make_node("Gather", ["word_embed", "input_ids"],
["word_gather_out"], "word_gather"),
helper.make_node("Add", ["word_gather_out", "pos_gather_out"],
["word_add_pos_out"], "word_add_pos"),
helper.make_node("Gather", ["seg_embed", "segment_ids"],
["seg_gather_out"], "seg_gather"),
helper.make_node("Add", ["word_add_pos_out", "seg_gather_out"],
["add3_out"], "add3"),
helper.make_node("LayerNormalization",
["add3_out", "layer_norm_weight", "layer_norm_bias"],
["layernorm_out"],
"layernorm",
axis=-1,
epsion=0.000009999999747378752),
helper.make_node("Cast", ["input_mask"], ["mask_cast_out"],
"mask_cast",
to=6),
helper.make_node("ReduceSum", ["mask_cast_out"], ["mask_index_out"],
"mask_index",
axes=[1],
keepdims=0),
helper.make_node(
"Attention",
["layernorm_out", "qkv_weights", "qkv_bias", "mask_index_out"],
["att_out"],
"att",
domain="com.microsoft",
num_heads=2),
helper.make_node("MatMul", ["att_out", "matmul_weight"],
["matmul_out"], "matmul"),
helper.make_node("Add", ["matmul_out", "add_bias"], ["add_out"],
"add"),
helper.make_node("Add", ["add_out", "layernorm_out"], ["add2_out"],
"add2")
]
# hidden_size=4, num_heads=2, max_seq_length=3
initializers = [ # initializers
helper.make_tensor('indices_0', TensorProto.INT64, [], [0]),
helper.make_tensor('indices_1', TensorProto.INT64, [], [1]),
helper.make_tensor('start_0', TensorProto.INT64, [], [0]),
helper.make_tensor('delta_1', TensorProto.INT64, [], [1]),
helper.make_tensor('word_embed', TensorProto.FLOAT, [2, 4],
[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]),
helper.make_tensor('pos_embed', TensorProto.FLOAT, [4, 4], [
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0,
2.0, 3.0, 4.0
]),
helper.make_tensor('seg_embed', TensorProto.FLOAT, [2, 4],
[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]),
helper.make_tensor('layer_norm_weight', TensorProto.FLOAT, [4],
[1.0, 2.0, 3.0, 4.0]),
helper.make_tensor('layer_norm_bias', TensorProto.FLOAT, [4],
[0.1, 0.2, 0.3, 0.4]),
helper.make_tensor('qkv_weights', TensorProto.FLOAT, [4, 4], [
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0,
2.0, 3.0, 4.0
]),
helper.make_tensor('qkv_bias', TensorProto.FLOAT, [4],
[0.1, 0.2, 0.3, 0.4]),
helper.make_tensor('matmul_weight', TensorProto.FLOAT, [4, 4], [
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0,
2.0, 3.0, 4.0
]),
helper.make_tensor('add_bias', TensorProto.FLOAT, [4],
[0.1, 0.2, 0.3, 0.4]),
]
graph = helper.make_graph(
nodes,
"EmbedLayerNorm_format3", #name
[ # inputs
helper.make_tensor_value_info('input_ids', TensorProto.INT64,
['batch', 3]),
helper.make_tensor_value_info('segment_ids', TensorProto.INT64,
['batch', 3]),
helper.make_tensor_value_info('input_mask', TensorProto.INT64,
['batch', 3]),
],
[ # outputs
helper.make_tensor_value_info('add2_out', TensorProto.FLOAT,
['batch', 3, 4]),
],
initializers)
model = helper.make_model(graph)
onnx.save(model, model_name)
def create_net_const(self, shape, scale, precision, ir_version):
"""
ONNX net IR net
Input->Concat(+scaled const)->Output => Input->Concat(+const)
"""
#
# Create ONNX model
#
import onnx
from onnx import helper
from onnx import TensorProto
import numpy as np
concat_axis = 0
output_shape = shape.copy()
output_shape[concat_axis] *= 2
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
constant = np.random.randint(-127, 127, shape).astype(np.float)
node_const_def = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=['const1'],
value=helper.make_tensor(
name='const_tensor',
data_type=TensorProto.FLOAT,
dims=constant.shape,
vals=constant.flatten(),
),
)
node_def = onnx.helper.make_node(
'Scale',
inputs=['const1'],
outputs=['scale'],
scale=scale
)
node_concat_def = onnx.helper.make_node(
'Concat',
inputs=['input', 'scale'],
outputs=['output'],
axis=concat_axis
)
# Create the graph (GraphProto)
graph_def = helper.make_graph(
[node_const_def, node_def, node_concat_def],
'test_model',
[input],
[output],
)
# Create the model (ModelProto)
onnx_net = helper.make_model(graph_def, producer_name='test_model')
#
# Create reference IR net
#
ir_const = constant.flatten() * scale
if precision == 'FP16':
ir_const = ir_const.astype(np.float16)
ref_net = None
return onnx_net, ref_net
def GenerateModel5(model_name):
batch_size = 2
hidden_size = 4
attention_heads = 2
sequence_length = 3
nodes = [
helper.make_node("Gather", ["word_embed", "input_ids"],
["word_gather_out"],
"word_gather",
axis=0),
helper.make_node("Add", ["word_gather_out", "pos_gather_out"],
["word_add_pos_out"], "word_add_pos"),
helper.make_node("Gather", ["seg_embed", "segment_ids"],
["seg_gather_out"],
"seg_gather",
axis=0),
helper.make_node("Add", ["word_add_pos_out", "seg_gather_out"],
["add3_out"], "add3"),
helper.make_node("LayerNormalization",
["add3_out", "layer_norm_weight", "layer_norm_bias"],
["layernorm_out"],
"layernorm",
axis=-1,
epsion=0.000009999999747378752),
helper.make_node("Cast", ["input_mask"], ["mask_cast_out"],
"mask_cast",
to=6),
helper.make_node("ReduceSum", ["mask_cast_out"], ["mask_index_out"],
"mask_index",
axes=[1],
keepdims=0),
helper.make_node(
"Attention",
["layernorm_out", "qkv_weights", "qkv_bias", "mask_index_out"],
["att_out"],
"att",
domain="com.microsoft",
num_heads=attention_heads),
helper.make_node("MatMul", ["att_out", "matmul_weight"],
["matmul_out"], "matmul"),
helper.make_node("Add", ["matmul_out", "add_bias"], ["add_out"],
"add"),
helper.make_node("Add", ["add_out", "layernorm_out"], ["add2_out"],
"add2")
]
qkv_weights = [
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0,
3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0,
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0,
3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0,
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0
]
initializers = [ # initializers
helper.make_tensor('word_embed', TensorProto.FLOAT, [2, hidden_size],
[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]),
helper.make_tensor(
'pos_gather_out', TensorProto.FLOAT,
[batch_size, sequence_length, hidden_size], [
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 8.0, 7.0, 6.0,
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 8.0, 7.0, 6.0
]),
helper.make_tensor('seg_embed', TensorProto.FLOAT, [2, hidden_size],
[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]),
helper.make_tensor('layer_norm_weight', TensorProto.FLOAT,
[hidden_size], [1.0, 2.0, 3.0, 4.0]),
helper.make_tensor('layer_norm_bias', TensorProto.FLOAT, [hidden_size],
[0.1, 0.2, 0.3, 0.4]),
helper.make_tensor('qkv_weights', TensorProto.FLOAT,
[hidden_size, 3 * hidden_size], qkv_weights),
helper.make_tensor(
'qkv_bias', TensorProto.FLOAT, [3 * hidden_size],
[0.1, 0.2, 0.3, 0.4, 0.1, 0.2, 0.3, 0.4, 0.1, 0.2, 0.3, 0.4]),
helper.make_tensor('matmul_weight', TensorProto.FLOAT,
[hidden_size, hidden_size], [
1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 1.0,
2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0
]),
helper.make_tensor('add_bias', TensorProto.FLOAT, [hidden_size],
[0.1, 0.2, 0.3, 0.4]),
]
graph = helper.make_graph(
nodes,
"EmbedLayerNorm_format5", #name
[ # inputs
helper.make_tensor_value_info('input_ids', TensorProto.INT64,
[batch_size, sequence_length]),
helper.make_tensor_value_info('segment_ids', TensorProto.INT64,
[batch_size, sequence_length]),
helper.make_tensor_value_info('input_mask', TensorProto.INT64,
[batch_size, sequence_length]),
],
[ # outputs
helper.make_tensor_value_info(
'add2_out', TensorProto.FLOAT,
[batch_size, sequence_length, hidden_size]),
],
initializers)
model = helper.make_model(graph)
onnx.save(model, model_name)
['Y'], # outputs
mode='constant', # attributes
value=1.5,
pads=[0, 1, 0, 1],
)
# Create the graph (GraphProto)
graph_def = helper.make_graph(
[node_def],
'test-model',
[X],
[Y],
)
# Create the model (ModelProto)
model_def = helper.make_model(graph_def, producer_name='onnx-example')
model_def.opset_import[0].version = 10
print('The model is:\n{}'.format(model_def))
onnx.checker.check_model(model_def)
print('The model is checked!')
#####################################
# Same example with sklearn-onnx
# ++++++++++++++++++++++++++++++
#
# Every operator has its own class in *sklearn-onnx*.
# The list is dynamically created based on the installed
# onnx package.
from skl2onnx.algebra.onnx_ops import OnnxPad # noqa
def gen_lstm_onnx_test_model(
model_path,
seq_length,
batch_size,
hidden_size,
input_size,
direction,
has_bias,
has_sequence_lens,
has_initial_h,
has_initial_c,
has_peephole,
input_forget=False,
):
# Validate parameters
assert direction in LSTM_DIRS, "ONNX LSTM direction invalid!"
assert not has_sequence_lens, "ONNX LSTM Variable sequence length not supported"
# Get number of directions
num_directions = 2 if (direction == LSTM_DIR_BIDIRECTIONAL) else 1
# Tensor sizes
X_shape = [seq_length, batch_size, input_size]
W_shape = [num_directions, 4 * hidden_size, input_size]
R_shape = [num_directions, 4 * hidden_size, hidden_size]
B_shape = [num_directions, 8 * hidden_size]
sequence_lens_shape = [batch_size]
initial_h_shape = [num_directions, batch_size, hidden_size]
initial_c_shape = [num_directions, batch_size, hidden_size]
P_shape = [num_directions, 3 * hidden_size]
Y_shape = [seq_length, num_directions, batch_size, hidden_size]
# Generate random inputs (weights are assumed concatenated in ONNX format: i,o,f,c)
np.random.seed(1)
X = np.random.randn(*X_shape)
W = np.random.randn(*W_shape)
R = np.random.randn(*R_shape)
B = np.random.randn(*B_shape) if has_bias else np.zeros(B_shape)
sequence_lens = (np.random.randint(1, seq_length, batch_size)
if has_sequence_lens else np.tile(seq_length, batch_size))
initial_h = (np.random.randn(
*initial_h_shape) if has_initial_h else np.zeros(initial_h_shape))
initial_c = (np.random.randn(
*initial_c_shape) if has_initial_c else np.zeros(initial_c_shape))
P = np.random.randn(*P_shape) if has_peephole else np.zeros(P_shape)
# Function to get all the weight components for the given direction
def get_weights(dir_idx):
Wi = np.reshape(W[dir_idx, 0 * hidden_size:1 * hidden_size, :],
[hidden_size, input_size])
Wo = np.reshape(W[dir_idx, 1 * hidden_size:2 * hidden_size, :],
[hidden_size, input_size])
Wf = np.reshape(W[dir_idx, 2 * hidden_size:3 * hidden_size, :],
[hidden_size, input_size])
Wc = np.reshape(W[dir_idx, 3 * hidden_size:4 * hidden_size, :],
[hidden_size, input_size])
Ri = np.reshape(R[dir_idx, 0 * hidden_size:1 * hidden_size, :],
[hidden_size, hidden_size])
Ro = np.reshape(R[dir_idx, 1 * hidden_size:2 * hidden_size, :],
[hidden_size, hidden_size])
Rf = np.reshape(R[dir_idx, 2 * hidden_size:3 * hidden_size, :],
[hidden_size, hidden_size])
Rc = np.reshape(R[dir_idx, 3 * hidden_size:4 * hidden_size, :],
[hidden_size, hidden_size])
bWi = np.reshape(B[dir_idx, 0 * hidden_size:1 * hidden_size],
[hidden_size])
bWo = np.reshape(B[dir_idx, 1 * hidden_size:2 * hidden_size],
[hidden_size])
bWf = np.reshape(B[dir_idx, 2 * hidden_size:3 * hidden_size],
[hidden_size])
bWc = np.reshape(B[dir_idx, 3 * hidden_size:4 * hidden_size],
[hidden_size])
bRi = np.reshape(B[dir_idx, 4 * hidden_size:5 * hidden_size],
[hidden_size])
bRo = np.reshape(B[dir_idx, 5 * hidden_size:6 * hidden_size],
[hidden_size])
bRf = np.reshape(B[dir_idx, 6 * hidden_size:7 * hidden_size],
[hidden_size])
bRc = np.reshape(B[dir_idx, 7 * hidden_size:8 * hidden_size],
[hidden_size])
Pi = np.tile(P[dir_idx, 0 * hidden_size:1 * hidden_size],
(batch_size, 1))
Po = np.tile(P[dir_idx, 1 * hidden_size:2 * hidden_size],
(batch_size, 1))
Pf = np.tile(P[dir_idx, 2 * hidden_size:3 * hidden_size],
(batch_size, 1))
return (
Wi,
Wo,
Wf,
Wc,
Ri,
Ro,
Rf,
Rc,
bWi,
bWo,
bWf,
bWc,
bRi,
bRo,
bRf,
bRc,
Pi,
Po,
Pf,
)
# Function to get PyTorch weights (which are in the i, f, c, o order)
def get_torch_weights(dir_idx):
(
Wi,
Wo,
Wf,
Wc,
Ri,
Ro,
Rf,
Rc,
bWi,
bWo,
bWf,
bWc,
bRi,
bRo,
bRf,
bRc,
Pi,
Po,
Pf,
) = get_weights(dir_idx)
W_torch = np.concatenate((Wi, Wf, Wc, Wo), 0)
R_torch = np.concatenate((Ri, Rf, Rc, Ro), 0)
bW_torch = np.concatenate((bWi, bWf, bWc, bWo), 0)
bR_torch = np.concatenate((bRi, bRf, bRc, bRo), 0)
return (W_torch, R_torch, bW_torch, bR_torch)
# ----------------------------------------- COMPUTE pyTORCH REFERENCE ----------------------------------------------
# Compute reference using Pytorch. Pytorch LSTM has only forward/bidirectional so we will do the reverse LSTM using
# a Pytorch forward LSTM.
lstm = torch.nn.LSTM(
input_size=input_size,
hidden_size=hidden_size,
num_layers=1,
bias=True,
batch_first=False,
dropout=0,
bidirectional=(direction == LSTM_DIR_BIDIRECTIONAL),
)
# Get LSTM state dictionary
lstm_state_dict = lstm.state_dict()
# Assign forward weights
forwardEnabled = direction in [LSTM_DIR_FORWARD, LSTM_DIR_BIDIRECTIONAL]
if forwardEnabled:
forward_dir_idx = 0
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(forward_dir_idx)
lstm_state_dict["weight_ih_l0"] = torch.tensor(W_torch,
dtype=torch.float32)
lstm_state_dict["weight_hh_l0"] = torch.tensor(R_torch,
dtype=torch.float32)
lstm_state_dict["bias_ih_l0"] = torch.tensor(bW_torch,
dtype=torch.float32)
lstm_state_dict["bias_hh_l0"] = torch.tensor(bR_torch,
dtype=torch.float32)
# Assign reverse weights
reverseEnabled = direction in [LSTM_DIR_REVERSE, LSTM_DIR_BIDIRECTIONAL]
if reverseEnabled:
if direction == LSTM_DIR_REVERSE:
reverse_dir_idx = 0
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(reverse_dir_idx)
lstm_state_dict["weight_ih_l0"] = torch.tensor(W_torch,
dtype=torch.float32)
lstm_state_dict["weight_hh_l0"] = torch.tensor(R_torch,
dtype=torch.float32)
lstm_state_dict["bias_ih_l0"] = torch.tensor(bW_torch,
dtype=torch.float32)
lstm_state_dict["bias_hh_l0"] = torch.tensor(bR_torch,
dtype=torch.float32)
else:
reverse_dir_idx = 1
(W_torch, R_torch, bW_torch,
bR_torch) = get_torch_weights(reverse_dir_idx)
lstm_state_dict["weight_ih_l0_reverse"] = torch.tensor(
W_torch, dtype=torch.float32)
lstm_state_dict["weight_hh_l0_reverse"] = torch.tensor(
R_torch, dtype=torch.float32)
lstm_state_dict["bias_ih_l0_reverse"] = torch.tensor(
bW_torch, dtype=torch.float32)
lstm_state_dict["bias_hh_l0_reverse"] = torch.tensor(
bR_torch, dtype=torch.float32)
# Set LSTM state dictionary
lstm.load_state_dict(lstm_state_dict, strict=True)
# Perform inference
X_torch = torch.tensor(X, dtype=torch.float32)
initial_h_torch = torch.tensor(initial_h, dtype=torch.float32)
initial_c_torch = torch.tensor(initial_c, dtype=torch.float32)
if direction == LSTM_DIR_REVERSE:
Y, (next_h, next_c) = lstm(X_torch.flip([0]),
(initial_h_torch, initial_c_torch))
Y = Y.flip([0])
else:
Y, (next_h, next_c) = lstm(X_torch, (initial_h_torch, initial_c_torch))
# Reshape output to ONNX format [seq_length, num_directions, batch_size, hidden_size]
Y_ref = Y.detach().numpy()
Y_ref = np.reshape(Y_ref,
[seq_length, batch_size, num_directions, hidden_size])
Y_ref = np.transpose(Y_ref, [0, 2, 1, 3])
# Reshape states to ONNX format
Y_h_ref = next_h.detach().numpy()
Y_c_ref = next_c.detach().numpy()
# --------------------------------------- COMPUTE PYTHON-NUMPY REFERENCE -------------------------------------------
# Create X slices
Xslices = list()
for t in range(seq_length):
Xslices.append(np.reshape(X[t, :, :], [batch_size, input_size]))
# Function to compute one LSTM cell
def compute_lstm(forward):
dir_idx = 0 if forward else (0 if direction == LSTM_DIR_REVERSE else 1)
(
Wi,
Wo,
Wf,
Wc,
Ri,
Ro,
Rf,
Rc,
bWi,
bWo,
bWf,
bWc,
bRi,
bRo,
bRf,
bRc,
Pi,
Po,
Pf,
) = get_weights(dir_idx)
def f(x):
return 1 / (1 + np.exp(-x))
def g(x):
return np.tanh(x)
def h(x):
return np.tanh(x)
def mm(x, w):
return np.matmul(x, w.transpose())
Ht = np.reshape(initial_h[dir_idx, :, :], [batch_size, hidden_size])
Ct = np.reshape(initial_c[dir_idx, :, :], [batch_size, hidden_size])
Yslices = list()
for t in range(seq_length):
xt = Xslices[t] if forward else Xslices[seq_length - 1 - t]
ft = f(mm(xt, Wf) + bWf + mm(Ht, Rf) + bRf + Pf * Ct)
if input_forget:
it = 1 - ft
else:
it = f(mm(xt, Wi) + bWi + mm(Ht, Ri) + bRi + Pi * Ct)
ctild = g(mm(xt, Wc) + bWc + mm(Ht, Rc) + bRc)
Ct = ft * Ct + it * ctild
ot = f(mm(xt, Wo) + bWo + mm(Ht, Ro) + bRo + Po * Ct)
Ht = ot * h(Ct)
Yslices.append(Ht)
return Yslices, Ht, Ct
Yslices = list()
Hslices = list()
Cslices = list()
# Compute forward LSTM
forwardYslices = list()
if forwardEnabled:
Yt, Ht, Ct = compute_lstm(True)
forwardYslices += Yt
Hslices.append(Ht)
Cslices.append(Ct)
# Compute reverse LSTM
reverseYslices = list()
if reverseEnabled:
Yt, Ht, Ct = compute_lstm(False)
reverseYslices += Yt
Hslices.append(Ht)
Cslices.append(Ct)
# Concatenate slices
for t in range(seq_length):
if forwardEnabled:
Yslices.append(forwardYslices[t])
if reverseEnabled:
Yslices.append(reverseYslices[seq_length - 1 - t])
Y_ref_np = np.concatenate(Yslices, 0).reshape(
[seq_length, num_directions, batch_size, hidden_size])
Y_h_ref_np = np.concatenate(Hslices, 0).reshape(
[num_directions, batch_size, hidden_size])
Y_c_ref_np = np.concatenate(Cslices, 0).reshape(
[num_directions, batch_size, hidden_size])
# Use numpy implementation when using peepholes or input_forget, else assert errors
if has_peephole or input_forget:
Y_ref = Y_ref_np
Y_h_ref = Y_h_ref_np
Y_c_ref = Y_c_ref_np
else:
assert (np.max(np.abs(Y_ref - Y_ref_np)) <
1e-6), "Mismatch between Pytorch and Numpy LSTM implementation"
assert (np.max(np.abs(Y_h_ref - Y_h_ref_np)) <
1e-6), "Mismatch between Pytorch and Numpy LSTM implementation"
assert (np.max(np.abs(Y_c_ref - Y_c_ref_np)) <
1e-6), "Mismatch between Pytorch and Numpy LSTM implementation"
# ---------------------------------------------- NODE DEFINITION --------------------------------------------------
# Node inputs
node_inputs = [
"X",
"W",
"R",
"B" if has_bias else "",
"",
"initial_h" if has_initial_h else "",
"initial_c" if has_initial_c else "",
"P" if has_peephole else "",
]
# Node outputs
node_outputs = ["Y", "Y_h", "Y_c"]
# LSTM node definition
lstm_node_def = onnx.helper.make_node(
"LSTM",
name="lstm",
inputs=node_inputs,
outputs=node_outputs,
hidden_size=hidden_size,
direction=direction,
input_forget=input_forget,
)
# Error node definition
err_node_def = onnx.helper.make_node("Sub",
name="error",
inputs=["Y", "Y_ref"],
outputs=["Y_err"])
# --------------------------------------------- GRAPH DEFINITION --------------------------------------------------
graph_input = list()
graph_init = list()
graph_output = list()
# LSTM inputs
graph_input.append(
helper.make_tensor_value_info("X", TensorProto.FLOAT, X_shape))
graph_input.append(
helper.make_tensor_value_info("W", TensorProto.FLOAT, W_shape))
graph_input.append(
helper.make_tensor_value_info("R", TensorProto.FLOAT, R_shape))
if has_bias:
graph_input.append(
helper.make_tensor_value_info("B", TensorProto.FLOAT, B_shape))
if has_sequence_lens:
graph_input.append(
helper.make_tensor_value_info("sequence_lens", TensorProto.INT32,
sequence_lens_shape))
if has_initial_h:
graph_input.append(
helper.make_tensor_value_info("initial_h", TensorProto.FLOAT,
initial_h_shape))
if has_initial_c:
graph_input.append(
helper.make_tensor_value_info("initial_c", TensorProto.FLOAT,
initial_c_shape))
if has_peephole:
graph_input.append(
helper.make_tensor_value_info("P", TensorProto.FLOAT, P_shape))
# Reference input
graph_input.append(
helper.make_tensor_value_info("Y_ref", TensorProto.FLOAT, Y_shape))
# LSTM initializers
graph_init.append(make_init("X", TensorProto.FLOAT, X))
graph_init.append(make_init("W", TensorProto.FLOAT, W))
graph_init.append(make_init("R", TensorProto.FLOAT, R))
if has_bias:
graph_init.append(make_init("B", TensorProto.FLOAT, B))
if has_sequence_lens:
graph_init.append(
make_init("sequence_lens", TensorProto.INT32, sequence_lens))
if has_initial_h:
graph_init.append(make_init("initial_h", TensorProto.FLOAT, initial_h))
if has_initial_c:
graph_init.append(make_init("initial_c", TensorProto.FLOAT, initial_c))
if has_peephole:
graph_init.append(make_init("P", TensorProto.FLOAT, P))
# Reference initializer
graph_init.append(make_init("Y_ref", TensorProto.FLOAT, Y_ref))
# Graph outputs
graph_output.append(
helper.make_tensor_value_info("Y_err", TensorProto.FLOAT, Y_shape))
# Define graph (GraphProto)
graph_name = "lstm_test"
graph_def = helper.make_graph(
[lstm_node_def, err_node_def],
graph_name,
inputs=graph_input,
outputs=graph_output,
)
# Set initializers
graph_def.initializer.extend(graph_init)
# --------------------------------------------- MODEL DEFINITION --------------------------------------------------
# Define model (ModelProto)
model_def = helper.make_model(graph_def, producer_name="onnx-lstm")
# Check model
onnx.checker.check_model(model_def)
# Print model
with open(model_path, "w") as f:
f.write(str(model_def))
opsets = []
onnxdomain = OperatorSetIdProto()
onnxdomain.version = 10
onnxdomain.domain = "" # The empty string ("") or absence of this field implies the operator set that is defined as part of the ONNX specification.
opsets.append(onnxdomain)
msdomain = OperatorSetIdProto()
msdomain.version = 1
msdomain.domain = "com.microsoft"
opsets.append(msdomain)
kwargs = {}
kwargs["opset_imports"] = opsets
model_def = helper.make_model(graph_def,
producer_name="onnx-example",
**kwargs)
file_name = "gelu_format2_0" if switch_order else "gelu_format2_1"
if has_bias:
file_name += "_with_bias"
if gelu_use_graph_input:
file_name += "_use_graph_input"
if node_has_graph_output:
file_name += "_use_graph_output"
file_name += ".onnx"
onnx.save(model_def, file_name)
print(file_name)
data_type=TensorProto.FLOAT,
dims=(1, ),
vals=B.reshape(1).tolist())
# Create the graph (GraphProto)
graph_def = helper.make_graph(
[node_def],
"test-model",
inputs=[
helper.make_tensor_value_info("data", TensorProto.FLOAT, [1, 1, 3, 3]),
helper.make_tensor_value_info("W", TensorProto.FLOAT, [1, 1, 2, 2]),
helper.make_tensor_value_info("B", TensorProto.FLOAT, [
1,
])
],
outputs=[
helper.make_tensor_value_info("y", TensorProto.FLOAT, [1, 1, 4, 4])
])
graph_def.initializer.extend([weight_tensor])
graph_def.initializer.extend([bias_tensor])
graph_def.initializer[0].name = 'W'
graph_def.initializer[1].name = 'B'
# Create the model (ModelProto)
model_def = helper.make_model(graph_def, producer_name='onnx-conv')
with open('simpleConv.onnxtxt', 'w') as f:
f.write(str(model_def))
onnx.checker.check_model(model_def)
print('The model is checked!')
def save_model(graph, file_name):
model = helper.make_model(graph)
onnx.checker.check_model(model)
onnx.save(model, file_name)
def test_merge_drop_unnecessary_initializers_and_value_info(
self): # type: () -> None
'''
Tests automatic removal of initializers when merging graphs
'''
ops = [helper.make_opsetid("", 10)]
g = GraphProto()
g.input.extend(
[helper.make_tensor_value_info('x', TensorProto.FLOAT, [])])
g.output.extend(
[helper.make_tensor_value_info('y', TensorProto.FLOAT, [])])
g.node.extend(
[helper.make_node('Identity', inputs=['x'], outputs=['y'])])
g1 = GraphProto()
g1.CopyFrom(g)
g1.name = 'g1'
m1 = helper.make_model(g1, producer_name='test', opset_imports=ops)
checker.check_model(m1)
g2 = GraphProto()
g2.CopyFrom(g)
g2.name = 'g2'
g2.initializer.extend([
helper.make_tensor(name='x',
data_type=TensorProto.FLOAT,
dims=(),
vals=[0])
])
m2 = helper.make_model(g2, producer_name='test', opset_imports=ops)
checker.check_model(m2)
g3 = GraphProto()
g3.CopyFrom(g)
g3.name = 'g3'
g3.sparse_initializer.extend([_make_sparse_tensor('x')])
m3 = helper.make_model(g3, producer_name='test', opset_imports=ops)
checker.check_model(m3)
g4 = GraphProto()
g4.CopyFrom(g)
g4.name = 'g3'
g4.value_info.extend(
[helper.make_tensor_value_info('x', TensorProto.FLOAT, [])])
m4 = helper.make_model(g4, producer_name='test', opset_imports=ops)
checker.check_model(m4)
# Initializer 'x' from m1 is removed, because there is no longer an input with that name
out_m1 = compose.merge_models(m1,
m2,
prefix1='m1/',
io_map=[('y', 'x')])
self.assertEqual(0, len(out_m1.graph.initializer))
# Sparse initializer 'x' from m1 is removed, because there is no longer an input with that name
out_m2 = compose.merge_models(m1,
m3,
prefix1='m1/',
io_map=[('y', 'x')])
self.assertEqual(0, len(out_m2.graph.initializer))
# Value info 'x' from m1 is removed, because there is no longer an input with that name
out_m3 = compose.merge_models(m1,
m4,
prefix1='m1/',
io_map=[('y', 'x')])
self.assertEqual(0, len(out_m3.graph.value_info))
def test_training_info_proto(self): # type: () -> None
# Inference graph.
A_shape = [2, 2]
A_name = 'A'
A = np.random.rand(*A_shape).astype(np.float32)
A_initializer = numpy_helper.from_array(A, name=A_name)
A_value_info = helper.make_tensor_value_info(A_name, TensorProto.FLOAT,
A_shape)
B_shape = [2, 2]
B_name = 'B'
B = np.random.rand(*B_shape).astype(np.float32)
B_initializer = numpy_helper.from_array(B, name=B_name)
B_value_info = helper.make_tensor_value_info(B_name, TensorProto.FLOAT,
B_shape)
C_shape = [2, 2]
C_name = 'C'
C_value_info = helper.make_tensor_value_info(C_name, TensorProto.FLOAT,
C_shape)
inference_node = helper.make_node('MatMul',
inputs=[A_name, B_name],
outputs=[C_name])
inference_graph = helper.make_graph([inference_node],
'simple_inference',
[A_value_info, B_value_info],
[C_value_info],
[A_initializer, B_initializer])
# Training graph
X_shape = [2, 2]
X_name = 'X'
X = np.random.rand(*X_shape).astype(np.float32)
X_initializer = numpy_helper.from_array(X, name=X_name)
X_value_info = helper.make_tensor_value_info(X_name, TensorProto.FLOAT,
X_shape)
Y_shape = [2, 2]
Y_name = 'Y'
Y_value_info = helper.make_tensor_value_info(Y_name, TensorProto.FLOAT,
Y_shape)
node = helper.make_node(
'MatMul',
inputs=[X_name, C_name], # tensor "C" is from inference graph.
outputs=[Y_name])
training_graph = helper.make_graph([node], 'simple_training',
[X_value_info], [Y_value_info],
[X_initializer])
# Capture assignment of B <--- Y.
training_info = helper.make_training_info(training_graph,
[(B_name, Y_name)], None,
None)
# Create a model with both inference and training information.
model = helper.make_model(inference_graph)
# Check if the inference-only part is correct.
onnx.checker.check_model(model)
# Insert training information.
new_training_info = model.training_info.add()
new_training_info.CopyFrom(training_info)
# Generate the actual training graph from training information so that
# we can run onnx checker to check if the full training graph is a valid
# graph. As defined in spec, full training graph forms by concatenating
# corresponding fields.
full_training_graph = helper.make_graph(
list(model.graph.node) +
list(model.training_info[0].algorithm.node), 'full_training_graph',
list(model.graph.input) +
list(model.training_info[0].algorithm.input),
list(model.graph.output) +
list(model.training_info[0].algorithm.output),
list(model.graph.initializer) +
list(model.training_info[0].algorithm.initializer))
# Wrap full training graph as a ModelProto so that we can run checker.
full_training_model = helper.make_model(full_training_graph)
full_training_model_with_shapes = shape_inference.infer_shapes(
full_training_model)
onnx.checker.check_model(full_training_model_with_shapes)
def test_overlapping_function_names(self): # type: () -> None
'''
Tests error checking when the name of local function entries overlaps
'''
ops = [helper.make_opsetid("", 10), helper.make_opsetid("local", 10)]
def _make_function(
domain, # type: Text
fname, # type: Text
inputs, # type: List[Text]
outputs, # type: List[Text]
nodes, # type: List[NodeProto]
): # type: (...) -> FunctionProto
f = FunctionProto()
f.domain = domain
f.name = fname
f.input.extend(inputs)
f.output.extend(outputs)
f.node.extend(nodes)
f.opset_import.extend(ops)
return f
ops = [helper.make_opsetid("", 10), helper.make_opsetid("local", 10)]
g = GraphProto()
g.input.extend([
helper.make_tensor_value_info('x0', TensorProto.FLOAT, []),
helper.make_tensor_value_info('x1', TensorProto.FLOAT, [])
])
g.output.extend([
helper.make_tensor_value_info('y', TensorProto.FLOAT, []),
])
g.node.extend([
helper.make_node('f1',
domain='local',
inputs=['x0', 'x1'],
outputs=['y'])
])
g1 = GraphProto()
g1.CopyFrom(g)
g1.name = 'g1'
m1 = helper.make_model(g1, producer_name='test', opset_imports=ops)
m1.functions.extend([
_make_function(
'local', 'f1', ['x0', 'x1'], ['y'],
[helper.make_node('Add', inputs=['x0', 'x1'], outputs=['y'])])
])
checker.check_model(m1)
g2 = GraphProto()
g2.CopyFrom(g)
g2.name = 'g2'
m2 = helper.make_model(g2, producer_name='test', opset_imports=ops)
m2.functions.extend([
_make_function(
'local', 'f1', ['x0', 'x1'], ['y'],
[helper.make_node('Mul', inputs=['x0', 'x1'], outputs=['y'])])
])
checker.check_model(m2)
m = compose.merge_models(m1,
m2,
io_map=[('y', 'x0'), ('y', 'x1')],
prefix1='m1/',
prefix2='m2/')
checker.check_model(m)
nodes = [n.op_type for n in m.graph.node]
self.assertEqual(['m1/f1', 'm2/f1'], nodes)
functions = [f.name for f in m.functions]
self.assertEqual(['m1/f1', 'm2/f1'], functions)
g3 = GraphProto()
g3.CopyFrom(g)
g3.name = 'g3'
g3.node[0].op_type = 'f2'
m3 = helper.make_model(g3, producer_name='test', opset_imports=ops)
m3.functions.extend([
_make_function('local', 'f1', ['x0', 'x1'], ['y'], [
helper.make_node('Add', inputs=['x0', 'x1'], outputs=['y0']),
helper.make_node('Mul', inputs=['x0', 'x1'], outputs=['y1']),
helper.make_node('Add', inputs=['y0', 'y1'], outputs=['y'])
]),
_make_function('local', 'f2', ['x0', 'x1'], ['y'], [
helper.make_node(
'f1', domain='local', inputs=['x0', 'x1'], outputs=['y0']),
helper.make_node('Mul', inputs=['x0', 'x1'], outputs=['y1']),
helper.make_node('Add', inputs=['y0', 'y1'], outputs=['y'])
])
])
checker.check_model(m3)
m = compose.merge_models(m1,
m3,
io_map=[('y', 'x0'), ('y', 'x1')],
prefix1='m1/',
prefix2='m3/')
checker.check_model(m)
nodes = [n.op_type for n in m.graph.node]
self.assertEqual(['m1/f1', 'm3/f2'], nodes)
functions = [f.name for f in m.functions]
self.assertEqual(['m1/f1', 'm3/f1', 'm3/f2'], functions)
self.assertEqual(['Add'], [n.op_type for n in m.functions[0].node])
self.assertEqual(['Add', 'Mul', 'Add'],
[n.op_type for n in m.functions[1].node])
self.assertEqual(['m3/f1', 'Mul', 'Add'],
[n.op_type for n in m.functions[2].node])
def _test_overlapping_names(
self,
inputs0=['i0', 'i1'], # type: List[Text]
inputs1=['i2', 'i3'], # type: List[Text]
outputs0=['o0', 'o1'], # type: List[Text]
outputs1=['o2', 'o3'], # type: List[Text]
value_info0=['v0', 'v1'], # type: List[Text]
value_info1=['v2', 'v3'], # type: List[Text]
initializer0=['init0', 'init1'], # type: List[Text]
initializer1=['init2', 'init3'], # type: List[Text]
sparse_initializer0=['sparse_init0',
'sparse_init1'], # type: List[Text]
sparse_initializer1=['sparse_init2',
'sparse_init3'], # type: List[Text]
): # type: (...) -> None
n0 = [
helper.make_node('Identity',
inputs=[inputs0[i]],
outputs=[outputs0[i]])
for i in range(len(inputs0))
]
i0 = [
helper.make_tensor_value_info(inputs0[i], TensorProto.FLOAT, [])
for i in range(len(inputs0))
]
o0 = [
helper.make_tensor_value_info(outputs0[i], TensorProto.FLOAT, [])
for i in range(len(outputs0))
]
vi0 = [
helper.make_tensor_value_info(value_info0[i], TensorProto.FLOAT,
[]) for i in range(len(value_info0))
]
init0 = [
helper.make_tensor(name=initializer0[i],
data_type=TensorProto.INT64,
dims=(),
vals=[1]) for i in range(len(initializer0))
]
sparse_init0 = [
_make_sparse_tensor(sparse_initializer0[i])
for i in range(len(sparse_initializer0))
]
n1 = [
helper.make_node('Identity',
inputs=[inputs1[i]],
outputs=[outputs1[i]])
for i in range(len(inputs1))
]
i1 = [
helper.make_tensor_value_info(inputs1[i], TensorProto.FLOAT, [])
for i in range(len(inputs1))
]
o1 = [
helper.make_tensor_value_info(outputs1[i], TensorProto.FLOAT, [])
for i in range(len(outputs1))
]
vi1 = [
helper.make_tensor_value_info(value_info1[i], TensorProto.FLOAT,
[]) for i in range(len(value_info1))
]
init1 = [
helper.make_tensor(name=initializer1[i],
data_type=TensorProto.INT64,
dims=(),
vals=[1]) for i in range(len(initializer1))
]
sparse_init1 = [
_make_sparse_tensor(sparse_initializer1[i])
for i in range(len(sparse_initializer1))
]
ops = [helper.make_opsetid("", 10)]
m0 = helper.make_model(helper.make_graph(
nodes=n0,
name='g0',
inputs=i0,
outputs=o0,
value_info=vi0,
initializer=init0,
sparse_initializer=sparse_init0),
producer_name='test',
opset_imports=ops)
m1 = helper.make_model(helper.make_graph(
nodes=n1,
name='g1',
inputs=i1,
outputs=o1,
value_info=vi1,
initializer=init1,
sparse_initializer=sparse_init1),
producer_name='test',
opset_imports=ops)
overlap = compose.check_overlapping_names(m0.graph, m1.graph)
i = 0
overlapping_inputs = list(set(inputs0) & set(inputs1))
overlapping_outputs = list(set(outputs0) & set(outputs1))
overlapping_edges = list(set(overlapping_inputs + overlapping_outputs))
if len(overlapping_edges) > 0:
self.assertEqual(overlap[i], ('edge', overlapping_edges))
i += 1
overlapping_vis = list(set(value_info0) & set(value_info1))
if len(overlapping_vis) > 0:
self.assertEqual(overlap[i], ('value_info', overlapping_vis))
i += 1
overlapping_init = list(set(initializer0) & set(initializer1))
if len(overlapping_init) > 0:
self.assertEqual(overlap[i], ('initializer', overlapping_init))
i += 1
overlapping_sparse_init = list(
set(sparse_initializer0) & set(sparse_initializer1))
if len(overlapping_sparse_init) > 0:
expected_overlap = []
for overlapping_name in overlapping_sparse_init:
expected_overlap.append(overlapping_name + '_values')
expected_overlap.append(overlapping_name + '_idx')
self.assertEqual(overlap[i],
('sparse_initializer', expected_overlap))
i += 1
m0_new = compose.add_prefix(m0, prefix='g0/')
overlap = compose.check_overlapping_names(m0_new.graph, m1.graph)
self.assertEqual(0, len(overlap))
def test_cast_errors():
from onnx.onnx_cpp2py_export.checker import ValidationError
np.random.seed(133391)
input_data = np.ceil(np.random.rand(2, 3, 4) * 16)
# missing 'to' attribute
node = onnx.helper.make_node("Cast", inputs=["A"], outputs=["B"])
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(node.output[0], onnx.TensorProto.FLOAT16,
input_data.shape)
] # type: ignore
graph = make_graph([node], "compute_graph", input_tensors, output_tensors)
model = make_model(graph, producer_name="NgraphBackend")
with pytest.raises(ValidationError):
import_onnx_model(model)
# unsupported data type representation
node = onnx.helper.make_node("Cast",
inputs=["A"],
outputs=["B"],
to=1.2345)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(node.output[0], onnx.TensorProto.INT32,
input_data.shape)
] # type: ignore
graph = make_graph([node], "compute_graph", input_tensors, output_tensors)
model = make_model(graph, producer_name="NgraphBackend")
with pytest.raises(ValidationError):
import_onnx_model(model)
# unsupported input tensor data type:
node = onnx.helper.make_node("Cast",
inputs=["A"],
outputs=["B"],
to=onnx.TensorProto.INT32)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.COMPLEX64, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(node.output[0], onnx.TensorProto.INT32,
input_data.shape)
] # type: ignore
graph = make_graph([node], "compute_graph", input_tensors, output_tensors)
model = make_model(graph, producer_name="NgraphBackend")
with pytest.raises((RuntimeError, NgraphTypeError)):
import_onnx_model(model)
# unsupported output tensor data type:
node = onnx.helper.make_node("Cast",
inputs=["A"],
outputs=["B"],
to=onnx.TensorProto.COMPLEX128)
input_tensors = [
make_tensor_value_info(name, onnx.TensorProto.FLOAT, value.shape)
for name, value in zip(node.input, [input_data])
]
output_tensors = [
make_tensor_value_info(node.output[0], onnx.TensorProto.COMPLEX128,
input_data.shape)
] # type: ignore
graph = make_graph([node], "compute_graph", input_tensors, output_tensors)
model = make_model(graph, producer_name="NgraphBackend")
with pytest.raises(RuntimeError):
import_onnx_model(model)
def convert_topology(topology,
model_name,
doc_string,
target_opset,
targeted_onnx,
channel_first_inputs=None):
'''
This function is used to convert our Topology object defined in _parser.py into a ONNX model (type: ModelProto).
:param topology: The Topology object we are going to convert
:param model_name: GraphProto's name. Let "model" denote the returned model. The string "model_name" would be
assigned to "model.graph.name."
:param doc_string: A string attached to the produced model
:param target_opset: number, for example, 7 for ONNX 1.2, and 8 for ONNX 1.3.
:param targeted_onnx[deprecated]: A string, which specifies the targeted ONNX version of the produced model. Possible values
include '1.1.2', '1.2', and so on.
:return: a ONNX ModelProto
'''
if targeted_onnx is not None and StrictVersion(
targeted_onnx) != StrictVersion(onnx.__version__):
warnings.warn(
'targeted_onnx is deprecated, please specify target_opset for the target model.\n'
+
'*** ONNX version conflict found. The installed version is %s while the targeted version is %s'
% (onnx.__version__, targeted_onnx))
opset_from_onnx_version = onnx.defs.onnx_opset_version()
if target_opset is None:
target_opset = opset_from_onnx_version
elif target_opset > opset_from_onnx_version:
raise RuntimeError(
"target_opset %d is higher than the number of the installed onnx package."
)
topology._initialize_graph_status_for_traversing()
container = ModelComponentContainer(target_opset)
# Put roots and leaves as ONNX's model into buffers. They will be added into ModelComponentContainer later.
tensor_inputs = {}
other_inputs = {}
tensor_outputs = {}
other_outputs = {}
for scope in topology.scopes:
for variable in scope.variables.values():
if variable.is_root:
if isinstance(variable.type,
(TensorType, Int64Type, FloatType, StringType)):
tensor_inputs[variable.raw_name] = variable
else:
other_inputs[variable.raw_name] = variable
if variable.is_leaf:
if isinstance(variable.type,
(TensorType, Int64Type, FloatType, StringType)):
tensor_outputs[variable.raw_name] = variable
else:
other_outputs[variable.raw_name] = variable
# Add roots the graph according to their order in the original model
invalid_name = []
nhwc_inputs = []
if channel_first_inputs is None:
channel_first_inputs = []
for name in topology.raw_model.input_names:
# Check input naming convention
input_name = name.replace('_', '').replace(":", "").replace("/", "")
if input_name and (input_name[0].isdigit() or
(not input_name.isalnum())):
invalid_name.append(name)
if name in tensor_inputs:
onnx_input = tensor_inputs[name] # type: Variable
if name in channel_first_inputs or \
(name.endswith(':0') and name[:-2] in channel_first_inputs):
nhwc_inputs.append(onnx_input.full_name)
s = onnx_input.type.shape
onnx_input.type.shape = [s[0], s[3], s[1], s[2]]
container.add_input(onnx_input)
if invalid_name:
warnings.warn(
'Some input names are not compliant with ONNX naming convention: %s'
% invalid_name)
for name in topology.raw_model.input_names:
if name in other_inputs:
container.add_input(other_inputs[name])
# Add leaves the graph according to their order in the original model
invalid_name = []
for name in topology.raw_model.output_names:
# Check output naming convention
output_name = name.replace('_', '').replace(":", "").replace("/", "")
if output_name and (output_name[0].isdigit() or
(not output_name.isalnum())):
invalid_name.append(name)
if name in tensor_outputs:
container.add_output(tensor_outputs[name])
if invalid_name:
warnings.warn(
'Some output names are not compliant with ONNX naming convention: %s'
% invalid_name)
for name in topology.raw_model.output_names:
if name in other_outputs:
container.add_output(other_outputs[name])
# Traverse the graph from roots to leaves
for operator in topology.topological_operator_iterator():
scope = next(scope for scope in topology.scopes
if scope.name == operator.scope)
if operator.type in topology.custom_conversion_functions:
topology.custom_conversion_functions[operator.type](scope,
operator,
container)
else:
# Convert the selected operator into some ONNX objects and save them into the container
registration.get_converter(operator.type)(scope, operator,
container)
# When calling ModelComponentContainer's add_initializer(...), nothing is added into the input list.
# However, for ONNX target opset < 9, initializers should also be model's (GraphProto) inputs.
# Thus, we create ValueInfoProto objects from initializers (type: TensorProto) directly and
# then add them into model's input list.
extra_inputs = [] # ValueInfoProto list of the initializers
for tensor in container.initializers:
# Sometimes (especially when creating optional input values such as RNN's initial hidden state), an initializer
# is also one of the original model's input, so it has been added into the container's input list. If this is
# the case, we need to skip one iteration to avoid duplicated inputs.
if tensor.name in [value_info.name for value_info in container.inputs]:
continue
# Initializers are always tensors so we can just call make_tensor_value_info(...)
value_info = helper.make_tensor_value_info(tensor.name,
tensor.data_type,
tensor.dims)
extra_inputs.append(value_info)
# enable the ONNX optimizations
nodes = optimize_onnx(container.nodes, nhwc_inputs,
container.inputs + extra_inputs, container.outputs)
# Create a graph from its main components
if container.target_opset < 9:
# Before ONNX opset 9, initializers need to be passed in with inputs
graph = helper.make_graph(nodes, model_name,
container.inputs + extra_inputs,
container.outputs, container.initializers)
else:
# In ONNX opset 9 and above, initializers are included as operator
# inputs, and therefore do not need to be passed as extra_inputs
graph = helper.make_graph(nodes, model_name, container.inputs,
container.outputs, container.initializers)
# Add extra information related to the graph
graph.value_info.extend(container.value_info)
# Create model
onnx_model = helper.make_model(graph)
# Merge operator sets for the same domain, the largest version number would be kept
purified_operator_set = dict()
for op_domain, op_version in container.node_domain_version_pair_sets:
if op_domain not in purified_operator_set:
purified_operator_set[op_domain] = op_version
else:
purified_operator_set[op_domain] = max(
purified_operator_set[op_domain], op_version)
# Fill operator sets
i = 0
for op_domain, op_version in purified_operator_set.items():
if i == 0 and len(onnx_model.opset_import) == 1:
# Overwrite the default operator set created by helper.make_model(...)
op_set = onnx_model.opset_import[0]
else:
# Just create one ONNX element in opset_import
op_set = onnx_model.opset_import.add()
op_set.domain = op_domain
op_set.version = op_version
i += 1
if container.target_opset < op_version:
raise RuntimeError(
('The specified opset %d is too low to convert this model, ' +
'which requires at least opset %d.') %
(container.target_opset, op_version))
elif container.target_opset > op_version:
getLogger('onnxmltools').warning(
'The maximum opset needed by this model is only %d.' %
op_version)
# Add extra information
add_metadata_props(onnx_model, topology.metadata_props, target_opset)
onnx_model.ir_version = onnx_proto.IR_VERSION
onnx_model.producer_name = utils.get_producer()
onnx_model.producer_version = utils.get_producer_version()
onnx_model.domain = utils.get_domain()
onnx_model.model_version = utils.get_model_version()
onnx_model.doc_string = doc_string
return onnx_model
def test_imagescaler1(self):
# test import model with imagescaler
try:
import onnx
from onnx import helper, TensorProto
except:
unittest.TestCase.skipTest(self, 'onnx not found')
if self.data_dir_local is None:
unittest.TestCase.skipTest(self, 'DLPY_DATA_DIR_LOCAL is not set in '
'the environment variables')
import numpy as np
n1 = helper.make_node('ImageScaler',
['X'],
['X1'],
bias=[0., 0., 0.],
scale=1.)
n2 = helper.make_node('Conv',
inputs=['X1', 'W1'],
outputs=['X2'],
kernel_shape=[3, 3],
pads=[0, 0, 0, 0])
n3 = helper.make_node('MatMul',
inputs=['X2', 'W2'],
outputs=['X3'])
W1 = np.ones((3, 3, 3)).astype(np.float32)
W2 = np.ones((9, 2)).astype(np.float32)
graph_def = helper.make_graph(
[n1, n2, n3],
name='test',
inputs=[
helper.make_tensor_value_info('X',
TensorProto.FLOAT,
[1, 3, 10, 10]),
helper.make_tensor_value_info('W1',
TensorProto.FLOAT,
[3, 3, 3]),
helper.make_tensor_value_info('W2',
TensorProto.FLOAT,
[9, 2])],
outputs=[
helper.make_tensor_value_info('X3',
TensorProto.FLOAT,
[1, 2])],
initializer=[
helper.make_tensor('W1',
TensorProto.FLOAT,
[3, 3, 3],
W1.flatten().astype(np.float32)),
helper.make_tensor('W2',
TensorProto.FLOAT,
[9, 2],
W2.flatten().astype(np.float32))])
onnx_model = helper.make_model(graph_def)
model1 = Model.from_onnx_model(self.s, onnx_model)
l1 = model1.layers[0]
self.assertTrue(l1.type == 'input')
self.assertTrue(l1.config['offsets'] == [0., 0., 0.])
self.assertTrue(l1.config['scale'] == 1.)
def test_sequence_ops(self):
# test SequenceConstruct and SequenceAt
a = np.random.randn(2, 1, 2).astype(np.float32)
b = np.random.randn(1, 1, 2).astype(np.float32)
c = np.random.randn(3, 1, 2).astype(np.float32)
seq_construct_node = helper.make_node('SequenceConstruct',
['a', 'b', 'c'], ['S'])
seq_at_node = helper.make_node('SequenceAt', ['S', 'at'], ['Y'])
out_value_info = helper.make_tensor_value_info('Y',
onnx.TensorProto.FLOAT,
[None])
a_value_info = helper.make_tensor_value_info('a',
onnx.TensorProto.FLOAT,
[2, 1, 2])
b_value_info = helper.make_tensor_value_info('b',
onnx.TensorProto.FLOAT,
[1, 1, 2])
c_value_info = helper.make_tensor_value_info('c',
onnx.TensorProto.FLOAT,
[3, 1, 2])
at_value_info = helper.make_tensor_value_info('at',
onnx.TensorProto.INT32,
[])
graph = helper.make_graph(
[seq_construct_node, seq_at_node],
name='seq_construct_at_test',
inputs=[a_value_info, b_value_info, c_value_info, at_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'at': 0})
np.testing.assert_almost_equal(output["Y"], a)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'at': -2})
np.testing.assert_almost_equal(output["Y"], b)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'at': 2})
np.testing.assert_almost_equal(output["Y"], c)
# test SequenceEmpty, SequenceInsert, and SequenceAt
p = np.int32(0)
seq_empty_node = helper.make_node('SequenceEmpty', [], ['S'])
seq_insert_node1 = helper.make_node('SequenceInsert', ['S', 'a'],
['S1'])
seq_insert_node2 = helper.make_node('SequenceInsert', ['S1', 'b'],
['S2'])
seq_insert_node3 = helper.make_node('SequenceInsert', ['S2', 'c', 'p'],
['S3'])
seq_at_node = helper.make_node('SequenceAt', ['S3', 'at'], ['Y'])
p_value_info = helper.make_tensor_value_info('p',
onnx.TensorProto.INT32,
[])
graph = helper.make_graph([
seq_empty_node, seq_insert_node1, seq_insert_node2,
seq_insert_node3, seq_at_node
],
name='seq_empty_insert_at_test',
inputs=[
a_value_info, b_value_info, c_value_info,
p_value_info, at_value_info
],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'p': p, 'at': 0})
np.testing.assert_almost_equal(output["Y"], c)
# test SequenceConstruct, SequenceErase, and SequenceLength
seq_construct_node = helper.make_node('SequenceConstruct',
['a', 'b', 'c'], ['S'])
seq_erase_node = helper.make_node('SequenceErase', ['S', 'p'], ['S1'])
seq_length_node = helper.make_node('SequenceLength', ['S1'], ['Y'])
graph = helper.make_graph(
[seq_construct_node, seq_erase_node, seq_length_node],
name='seq_construct_erase_length_test',
inputs=[a_value_info, b_value_info, c_value_info, p_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'p': p})
np.testing.assert_almost_equal(output["Y"], 2)
# test SequenceConstruct and SequenceErase
seq_construct_node = helper.make_node('SequenceConstruct',
['a', 'b', 'c'], ['S'])
seq_erase_node = helper.make_node('SequenceErase', ['S', 'p'], ['S1'])
seq_at_node = helper.make_node('SequenceAt', ['S1', 'at'], ['Y'])
graph = helper.make_graph(
[seq_construct_node, seq_erase_node, seq_at_node],
name='seq_construct_erase_test',
inputs=[
a_value_info, b_value_info, c_value_info, p_value_info,
at_value_info
],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'p': p, 'at': 0})
np.testing.assert_almost_equal(output["Y"], b)
output = tf_rep.run({'a': a, 'b': b, 'c': c, 'p': p, 'at': 1})
np.testing.assert_almost_equal(output["Y"], c)
# test SequenceConstruct and ConcatFromSequence
seq_construct_node = helper.make_node('SequenceConstruct',
['a', 'b', 'c'], ['S'])
concat_from_seq_node = helper.make_node('ConcatFromSequence', ['S'],
['Y'],
axis=1)
a = np.array([[1, 2], [3, 4]]).astype(np.float32)
b = np.array([[5, 6], [7, 8]]).astype(np.float32)
c = np.array([[9, 10], [11, 12]]).astype(np.float32)
a_value_info = helper.make_tensor_value_info('a',
onnx.TensorProto.FLOAT,
[2, 2])
b_value_info = helper.make_tensor_value_info('b',
onnx.TensorProto.FLOAT,
[2, 2])
c_value_info = helper.make_tensor_value_info('c',
onnx.TensorProto.FLOAT,
[2, 2])
graph = helper.make_graph(
[seq_construct_node, concat_from_seq_node],
name='seq_construct_concat_test',
inputs=[a_value_info, b_value_info, c_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'c': c})
d = np.concatenate((a, b, c), axis=1).astype(np.float32)
np.testing.assert_almost_equal(output["Y"], d)
# test SplitToSequence and SequenceAt
a = np.array([[1, 2, 3, 4, 5, 6, 7], [11, 12, 13, 14, 15, 16, 17],
[21, 22, 23, 24, 25, 26, 27]]).astype(np.float32)
b = np.int32([2, 1])
seq_split_node = helper.make_node('SplitToSequence', ['a', 'b'], ['S'])
seq_at_node = helper.make_node('SequenceAt', ['S', 'at'], ['Y'])
a_value_info = helper.make_tensor_value_info('a',
onnx.TensorProto.FLOAT,
[3, 7])
b_value_info = helper.make_tensor_value_info('b',
onnx.TensorProto.INT32,
[2])
at_value_info = helper.make_tensor_value_info('at',
onnx.TensorProto.INT32,
[])
graph = helper.make_graph(
[seq_split_node, seq_at_node],
name='split_to_seq_test',
inputs=[a_value_info, b_value_info, at_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'b': b, 'at': 1})
np.testing.assert_almost_equal(output["Y"], np.split(a, [2, 3])[1])
axis = 1
seq_split_node = helper.make_node('SplitToSequence', ['a'], ['S'],
axis=axis)
seq_at_node = helper.make_node('SequenceAt', ['S', 'at'], ['Y'])
at_value_info = helper.make_tensor_value_info('at',
onnx.TensorProto.INT32,
[])
graph = helper.make_graph([seq_split_node, seq_at_node],
name='split_to_seq_test',
inputs=[a_value_info, at_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'at': 0})
np.testing.assert_almost_equal(output["Y"], np.split(a, 7, axis=1)[0])
seq_split_node = helper.make_node('SplitToSequence', ['a'], ['S'],
keepdims=0)
seq_at_node = helper.make_node('SequenceAt', ['S', 'at'], ['Y'])
at_value_info = helper.make_tensor_value_info('at',
onnx.TensorProto.INT32,
[])
graph = helper.make_graph([seq_split_node, seq_at_node],
name='split_to_seq_test',
inputs=[a_value_info, at_value_info],
outputs=[out_value_info])
model = helper.make_model(graph, producer_name='backend-test')
tf_rep = prepare(model)
output = tf_rep.run({'a': a, 'at': 0})
expected = [np.squeeze(res) for res in np.split(a, 3)]
np.testing.assert_almost_equal(output["Y"], expected[0])
def construct_model_attention_and_matmul(self, output_model_path):
# (input)
# |
# Attention
# |
# MatMul
# |
# (output)
input_name = "input"
output_name = "output"
initializers = []
def make_attention_node(input_name, weight_shape, weight_name,
bias_shape, bias_name, output_name):
weight_data = np.random.normal(0, 0.1,
weight_shape).astype(np.float32)
initializers.append(
onnx.numpy_helper.from_array(weight_data, name=weight_name))
bias_data = np.random.normal(0, 0.1, bias_shape).astype(np.float32)
initializers.append(
onnx.numpy_helper.from_array(bias_data, name=bias_name))
return onnx.helper.make_node("Attention",
[input_name, weight_name, bias_name],
[output_name])
def make_matmul_node(input_name, weight_shape, weight_name,
output_name):
weight_data = np.random.normal(0, 0.1,
weight_shape).astype(np.float32)
initializers.append(
onnx.numpy_helper.from_array(weight_data, name=weight_name))
return onnx.helper.make_node("MatMul", [input_name, weight_name],
[output_name])
# make attention node
attention_output_name = "attention_output"
attention_node = make_attention_node(input_name, [10, 30],
"qkv.weight", [30], "qkv.bias",
attention_output_name)
attention_node.domain = "com.microsoft"
attention_node.attribute.extend(
[helper.make_attribute("num_heads", 5)])
# make matmul node
matmul_node = make_matmul_node(attention_output_name, [10, 10],
"matmul.weight", output_name)
# make graph
input_tensor = helper.make_tensor_value_info(input_name,
TensorProto.FLOAT,
[1, -1, 10])
output_tensor = helper.make_tensor_value_info(output_name,
TensorProto.FLOAT,
[1, -1, 10])
graph_name = "attention_test"
graph = helper.make_graph(
[attention_node, matmul_node],
graph_name,
[input_tensor],
[output_tensor],
initializer=initializers,
)
model = helper.make_model(graph,
opset_imports=[helper.make_opsetid("", 13)])
model.ir_version = onnx.IR_VERSION
onnx.save(model, output_model_path)
def create_neg(self, shape, ir_version):
"""
ONNX net IR net
Input->Neg->Output => Input->Power(scale=-1, shift=0, power=1)
"""
#
# Create ONNX model
#
import onnx
from onnx import helper
from onnx import TensorProto
input = helper.make_tensor_value_info('input', TensorProto.FLOAT,
shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT,
shape)
node_reduce_mean_def = onnx.helper.make_node(
'Neg',
inputs=['input'],
outputs=['output'],
)
# Create the graph (GraphProto)
graph_def = helper.make_graph(
[node_reduce_mean_def],
'test_neg_model',
[input],
[output],
)
# Create the model (ModelProto)
onnx_net = helper.make_model(graph_def, producer_name='test_neg_model')
#
# Create reference IR net
# Please, specify 'type': 'Input' for input node
# Moreover, do not forget to validate ALL layer attributes!!!
#
ref_net = None
if check_ir_version(10, None, ir_version):
nodes_attributes = {
'input': {
'kind': 'op',
'type': 'Parameter'
},
'input_data': {
'shape': shape,
'kind': 'data'
},
'neg': {
'kind': 'op',
'type': 'Negative'
},
'neg_data': {
'shape': shape,
'kind': 'data'
},
'result': {
'kind': 'op',
'type': 'Result'
}
}
ref_net = build_graph(nodes_attributes, [('input', 'input_data'),
('input_data', 'neg'),
('neg', 'neg_data'),
('neg_data', 'result')])
return onnx_net, ref_net
