def test_less():
g = graph.create_graph()
a = graph.create_initializer(g, "a", onnx.TensorProto.FLOAT, [4],
[1.1, 2.3, 3.5, 9.6])
b = graph.create_input(g, "b", onnx.TensorProto.FLOAT, [None, 1])
l = ops.less()(graph.merge(a, b))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.BOOL,
[None, 4])
assert_model_result(l,
input={'b': [
[0.1],
[1.2],
[11],
[4.2],
[np.NaN],
]},
expected_result=[[
[False, False, False, False],
[True, False, False, False],
[True, True, True, True],
[True, True, True, False],
[False, False, False, False],
]])
def _get_bin_index_on_continuous_value(g):
bin_count = len(bin_edges)
index_range = list(range(bin_count))
init_bin_index_range = graph.create_initializer(
g, "bin_index_range", onnx.TensorProto.FLOAT, [bin_count],
index_range)
init_bin_edges = graph.create_initializer(g, "bin_edges",
onnx.TensorProto.DOUBLE,
[bin_count], bin_edges)
g = ops.cast(onnx.TensorProto.DOUBLE)(g)
g = ops.less_or_equal()(graph.merge(init_bin_edges, g))
g = ops.cast(onnx.TensorProto.FLOAT)(g)
g = ops.mul()(graph.merge(g, init_bin_index_range))
g = ops.argmax(axis=1)(g)
return g
def _get_bin_score_1d(g):
init_bin_scores = graph.create_initializer(g, "bin_scores",
onnx.TensorProto.FLOAT,
bin_scores.shape,
bin_scores.flatten())
init_reshape = graph.create_initializer(
g,
"score_reshape",
onnx.TensorProto.INT64,
[3],
[-1, 1, bin_scores.shape[1]],
)
g = ops.gather_nd()(graph.merge(init_bin_scores,
g)) # gather score for each class
g = ops.reshape()(graph.merge(g, init_reshape))
return g
def _get_bin_score_2d(g):
init_bin_scores = graph.create_initializer(
g,
"bin_scores",
onnx.TensorProto.FLOAT,
[bin_scores.shape[0], bin_scores.shape[1]],
bin_scores.flatten(),
)
init_reshape = graph.create_initializer(
g,
"score_reshape",
onnx.TensorProto.INT64,
[3],
[-1, 1, 1],
)
g = ops.concat(axis=1)(g)
g = ops.gather_nd()(graph.merge(init_bin_scores, g))
g = ops.reshape()(graph.merge(g, init_reshape))
return g
def _predict_value(g):
init_reshape = graph.create_initializer(
g,
"reshape",
onnx.TensorProto.INT64,
[1],
[0],
)
g = ops.reshape()(graph.merge(g, init_reshape))
g = ops.identity("predict")(g)
return g
def _compute_class_score(g):
init_intercept = graph.create_initializer(
g,
"intercept",
onnx.TensorProto.FLOAT,
[intercept.shape[0]],
intercept,
)
init_sum_axis = graph.create_initializer(
g,
"sum_axis",
onnx.TensorProto.INT64,
[1],
[1],
)
g = ops.concat(axis=1)(g)
scores_output_name = g.transients[0].name
g = ops.reduce_sum(keepdims=0)(graph.merge(g, init_sum_axis))
g = ops.add()(graph.merge(g, init_intercept))
return g, scores_output_name
def test_create_initializer():
g = graph.create_graph()
init = graph.create_initializer(g, "foo", onnx.TensorProto.FLOAT, [4], [0.1, 0.2, 0.3, 0.4])
assert len(init.initializers) == 1
assert init.initializers == [onnx.helper.make_tensor(
'foo_0' ,
onnx.TensorProto.FLOAT,
[4],
[0.1, 0.2, 0.3, 0.4]
)]
assert init.initializers == init.transients
def _predict_proba(g):
if binary is True:
init_zeros = graph.create_initializer(
g,
"zeros",
onnx.TensorProto.FLOAT,
[2],
[0.0, 1.0],
)
init_reshape = graph.create_initializer(
g,
"reshape",
onnx.TensorProto.INT64,
[1],
[0],
)
g = ops.mul()(graph.merge(g, init_zeros))
g = ops.softmax(axis=1)(g)
g = ops.identity("predict")(g)
return g
def test_add():
g = graph.create_graph()
a = graph.create_initializer(g, "a", onnx.TensorProto.FLOAT, [1], [0.3])
i = graph.create_input(g, "i", onnx.TensorProto.FLOAT, [None])
l = ops.add()(graph.merge(i, a))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None])
assert_model_result(l,
input={
'i': [0.1, 1.2, 11, 4.2],
},
expected_result=[[0.4, 1.5, 11.3, 4.5]])
def test_reshape():
g = graph.create_graph()
shape = graph.create_initializer(g, "shape", onnx.TensorProto.INT64, [1],
[0])
i = graph.create_input(g, "i", onnx.TensorProto.FLOAT, [None, 1])
l = ops.reshape()(graph.merge(i, shape))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None])
assert_model_result(l,
input={'i': [
[0.1],
[1.2],
[11],
[4.2],
]},
expected_result=[[0.1, 1.2, 11, 4.2]])
def test_mul():
g = graph.create_graph()
a = graph.create_initializer(g, "a", onnx.TensorProto.FLOAT, [3],
[1.0, 2.0, 3.0])
b = graph.create_input(g, "b", onnx.TensorProto.FLOAT, [None, 3])
l = ops.mul()(graph.merge(a, b))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None, 3])
assert_model_result(l,
input={'b': [
[0.1, 0.1, 0.1],
[0.1, 0.2, 0.3],
]},
expected_result=[[
[0.1, 0.2, 0.3],
[0.1, 0.4, 0.9],
]])
def test_reduce_sum():
g = graph.create_graph()
axis = graph.create_initializer(g, "axis", onnx.TensorProto.INT64, [1],
[1])
i = graph.create_input(g, "i", onnx.TensorProto.FLOAT, [None, 3])
l = ops.reduce_sum(keepdims=0)(graph.merge(i, axis))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None])
assert_model_result(l,
input={
'i': [
[0.1, 1.0, 1.2],
[1.2, 0.4, 0.9],
[11, 0.8, -0.2],
[4.2, 3.2, -6.4],
]
},
expected_result=[[2.3, 2.5, 11.6, 1.0]])
def test_gather_elements():
g = graph.create_graph()
a = graph.create_initializer(g, "a", onnx.TensorProto.FLOAT, [3, 1],
[0.1, 0.2, 0.3])
b = graph.create_input(g, "b", onnx.TensorProto.INT64, [None, 1])
l = ops.gather_elements()(graph.merge(a, b))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None, 1])
assert_model_result(l,
input={'b': [
[2],
[1],
[0],
]},
expected_result=[[
[0.3],
[0.2],
[0.1],
]])
def test_gather_nd():
g = graph.create_graph()
a = graph.create_initializer(
g, "a", onnx.TensorProto.FLOAT, [3, 3],
np.array([
[0.1, 0.2, 0.3],
[1.1, 2.2, 3.3],
[0.1, 20.2, 30.3],
]).flatten())
b = graph.create_input(g, "b", onnx.TensorProto.INT64, [None, 2])
l = ops.gather_nd()(graph.merge(a, b))
l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT,
[None])
assert_model_result(l,
input={'b': [
[2, 0],
[1, 1],
[0, 1],
]},
expected_result=np.array([[0.1, 2.2, 0.2]]))
def test_merge():
g = graph.create_graph()
init1 = graph.create_initializer(g, "foo", onnx.TensorProto.FLOAT, [4], [0.1, 0.2, 0.3, 0.4])
init2 = graph.create_initializer(g, "foo", onnx.TensorProto.FLOAT, [4], [1.1, 1.2, 3.3, 4.4])
input1 = graph.create_input(g, "bar1", onnx.TensorProto.FLOAT, [None, 3])
input2 = graph.create_input(g, "bar2", onnx.TensorProto.FLOAT, [None, 4])
m = graph.merge(init1, input1, init2, input2)
assert len(m.initializers) == 2
assert len(m.inputs) == 2
assert len(m.transients) == 4
assert m.initializers == [
onnx.helper.make_tensor(
'foo_0' ,
onnx.TensorProto.FLOAT,
[4],
[0.1, 0.2, 0.3, 0.4]
),
onnx.helper.make_tensor(
'foo_1' ,
onnx.TensorProto.FLOAT,
[4],
[1.1, 1.2, 3.3, 4.4]
),
]
assert m.inputs == [
onnx.helper.make_tensor_value_info(
'bar1' ,
onnx.TensorProto.FLOAT,
[None, 3],
),
onnx.helper.make_tensor_value_info(
'bar2' ,
onnx.TensorProto.FLOAT,
[None, 4],
),
]
assert m.transients == [
onnx.helper.make_tensor(
'foo_0' ,
onnx.TensorProto.FLOAT,
[4],
[0.1, 0.2, 0.3, 0.4]
),
onnx.helper.make_tensor_value_info(
'bar1' ,
onnx.TensorProto.FLOAT,
[None, 3],
),
onnx.helper.make_tensor(
'foo_1' ,
onnx.TensorProto.FLOAT,
[4],
[1.1, 1.2, 3.3, 4.4]
),
onnx.helper.make_tensor_value_info(
'bar2' ,
onnx.TensorProto.FLOAT,
[None, 4],
),
]
