Skip to content

Commit 88f2652

Browse files
adrianlizarragaadrianlizarraga
authored
Regenerate test model with ONNX IR < 12 (#26149)
### Description - Regenerates the `input_propagate_to_output.onnx` model used in [this unit test](https://github.com/microsoft/onnxruntime/blob/35dcab5088118117acc6086c9b6dd6dd92c7060f/onnxruntime/test/shared_lib/test_inference.cc#L497-L506) so that it uses an ONNX IR version compatible with ONNX 1.18.0 (i.e., IR version < 12). - Adds script `input_propagate_to_output.py` that can be used to regenerate the `input_propagate_to_output.onnx` model. - Embed missing weight values that are needed to run the existing `test_dangling_input_segment_ids.py` script. ### Motivation and Context The main branch is using ONNX 1.19. However, this unit test also needs to pass in the `rel-1.23.1` branch, which is still using ONNX 1.18.0. So, by downgrading the model's IR version, the unit test can run in both branches. See original PR that added the test models: #26021
1 parent 35dcab5 commit 88f2652

File tree

3 files changed

+186
-23
lines changed

3 files changed

+186
-23
lines changed

‎onnxruntime/test/testdata/input_propagated_to_output.onnx‎

-169 Bytes
Binary file not shown.

‎onnxruntime/test/testdata/input_propagated_to_output.py‎

Lines changed: 113 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,113 @@
1+
"""
2+
Run this script to recreate the original onnx model.
3+
Example usage:
4+
python input_propagated_to_output.py input_propagated_to_output.onnx
5+
"""
6+
7+
import sys
8+
9+
import numpy as np
10+
import onnx
11+
12+
13+
def order_repeated_field(repeated_proto, key_name, order):
14+
order = list(order)
15+
repeated_proto.sort(key=lambda x: order.index(getattr(x, key_name)))
16+
17+
18+
def make_node(op_type, inputs, outputs, name=None, doc_string=None, domain=None, **kwargs):
19+
node = onnx.helper.make_node(op_type, inputs, outputs, name, doc_string, domain, **kwargs)
20+
if doc_string == "":
21+
node.doc_string = ""
22+
order_repeated_field(node.attribute, "name", kwargs.keys())
23+
return node
24+
25+
26+
def make_graph(*args, doc_string=None, **kwargs):
27+
graph = onnx.helper.make_graph(*args, doc_string=doc_string, **kwargs)
28+
if doc_string == "":
29+
graph.doc_string = ""
30+
return graph
31+
32+
33+
W1 = np.array(
34+
[
35+
[[[0.3258337378501892]], [[0.1461111307144165]], [[-0.4239698648452759]]],
36+
[[[0.14769716560840607]], [[0.20565544068813324]], [[-0.5241780877113342]]],
37+
[[[0.07987150549888611]], [[-0.17475983500480652]], [[0.005230882670730352]]],
38+
],
39+
dtype=np.float32,
40+
)
41+
42+
B1 = np.array(
43+
[-0.3170531392097473, -0.2701416313648224, -0.14249320328235626],
44+
dtype=np.float32,
45+
)
46+
47+
W3 = np.array(
48+
[
49+
[[[0.14025720953941345]], [[0.1433156430721283]], [[-0.1403128057718277]]],
50+
[[[-0.07530076801776886]], [[0.11853527277708054]], [[-0.19437682628631592]]],
51+
[[[0.5786639451980591]], [[-0.28565627336502075]], [[0.9048876166343689]]],
52+
],
53+
dtype=np.float32,
54+
)
55+
56+
B3 = np.array(
57+
[-0.13307525217533112, 0.5522456169128418, 0.6449958086013794],
58+
dtype=np.float32,
59+
)
60+
61+
W5 = np.array(
62+
[
63+
[[[-0.08959630876779556]], [[0.07607565075159073]], [[0.24446037411689758]]],
64+
[[[-0.06293385475873947]], [[-0.41520264744758606]], [[-0.83400559425354]]],
65+
[[[-0.031176576390862465]], [[-0.04187283664941788]], [[-0.439873069524765]]],
66+
],
67+
dtype=np.float32,
68+
)
69+
70+
B5 = np.array(
71+
[0.5949633717536926, -0.40198755264282227, -0.20182392001152039],
72+
dtype=np.float32,
73+
)
74+
75+
model = onnx.helper.make_model(
76+
opset_imports=[onnx.helper.make_operatorsetid("", 14)],
77+
ir_version=7,
78+
graph=make_graph(
79+
name="input_propagated_to_output",
80+
inputs=[
81+
onnx.helper.make_tensor_value_info("input", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
82+
],
83+
outputs=[
84+
onnx.helper.make_tensor_value_info("X6", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
85+
onnx.helper.make_tensor_value_info("X1", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
86+
onnx.helper.make_tensor_value_info("input", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
87+
onnx.helper.make_tensor_value_info("X2", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
88+
onnx.helper.make_tensor_value_info("X4", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
89+
onnx.helper.make_tensor_value_info("X3", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
90+
onnx.helper.make_tensor_value_info("X5", onnx.TensorProto.FLOAT, shape=[1, 3, 1, 3]),
91+
],
92+
initializer=[
93+
onnx.numpy_helper.from_array(W1, name="W1"),
94+
onnx.numpy_helper.from_array(W3, name="W3"),
95+
onnx.numpy_helper.from_array(W5, name="W5"),
96+
onnx.numpy_helper.from_array(B1, name="B1"),
97+
onnx.numpy_helper.from_array(B3, name="B3"),
98+
onnx.numpy_helper.from_array(B5, name="B5"),
99+
],
100+
nodes=[
101+
make_node("Relu", inputs=["input"], outputs=["X1"], name="Relu1"),
102+
make_node("Conv", inputs=["X1", "W1", "B1"], outputs=["X2"], name="Conv1"),
103+
make_node("Relu", inputs=["X2"], outputs=["X3"], name="Relu2"),
104+
make_node("Conv", inputs=["X3", "W3", "B3"], outputs=["X4"], name="Conv2"),
105+
make_node("Conv", inputs=["X1", "W5", "B5"], outputs=["X5"], name="Conv3"),
106+
make_node("Add", inputs=["X4", "X5"], outputs=["X6"], name="Add"),
107+
],
108+
),
109+
)
110+
111+
if __name__ == "__main__" and len(sys.argv) == 2:
112+
_, out_path = sys.argv
113+
onnx.save(model, out_path)

‎onnxruntime/test/testdata/test_dangling_input_segment_ids.py‎

Lines changed: 73 additions & 23 deletions
Original file line numberDiff line numberDiff line change
@@ -1,17 +1,13 @@
11
"""
22
Run this script to recreate the original onnx model.
33
Example usage:
4-
python test_dangling_input_segment_ids.py out_model_path.onnx
4+
python test_dangling_input_segment_ids.py test_dangling_input_segment_ids.onnx
55
"""
66

7-
import os
87
import sys
98

109
import numpy as np
1110
import onnx
12-
from onnx import TensorProto, helper, numpy_helper
13-
14-
DATA_DIR = os.path.join(os.path.dirname(os.path.realpath(__file__)), "test_dangling_input_segment_ids")
1511

1612

1713
def order_repeated_field(repeated_proto, key_name, order):
@@ -20,50 +16,104 @@ def order_repeated_field(repeated_proto, key_name, order):
2016

2117

2218
def make_node(op_type, inputs, outputs, name=None, doc_string=None, domain=None, **kwargs):
23-
node = helper.make_node(op_type, inputs, outputs, name, doc_string, domain, **kwargs)
19+
node = onnx.helper.make_node(op_type, inputs, outputs, name, doc_string, domain, **kwargs)
2420
if doc_string == "":
2521
node.doc_string = ""
2622
order_repeated_field(node.attribute, "name", kwargs.keys())
2723
return node
2824

2925

3026
def make_graph(*args, doc_string=None, **kwargs):
31-
graph = helper.make_graph(*args, doc_string=doc_string, **kwargs)
27+
graph = onnx.helper.make_graph(*args, doc_string=doc_string, **kwargs)
3228
if doc_string == "":
3329
graph.doc_string = ""
3430
return graph
3531

3632

37-
model = helper.make_model(
38-
opset_imports=[helper.make_operatorsetid("", 14), helper.make_operatorsetid("com.microsoft", 1)],
33+
WORD_EMBED = np.array(
34+
[
35+
[0.31524479389190674, 0.8928887248039246, 0.5778571963310242, 0.18401020765304565],
36+
[0.7879292368888855, 0.6120311617851257, 0.05390927195549011, 0.4201936721801758],
37+
[0.6790688633918762, 0.9186017513275146, 0.0004020248888991773, 0.976759135723114],
38+
[0.3765803277492523, 0.973783552646637, 0.6047161221504211, 0.8288457989692688],
39+
[0.5747115015983582, 0.6280761957168579, 0.2855762839317322, 0.5868333578109741],
40+
[0.750021755695343, 0.8583138585090637, 0.7550821900367737, 0.698057234287262],
41+
[0.8644794225692749, 0.3226810097694397, 0.6707887649536133, 0.4508739411830902],
42+
[0.38210275769233704, 0.4108113646507263, 0.401479572057724, 0.31738394498825073],
43+
[0.6219193935394287, 0.4302472770214081, 0.9738020896911621, 0.6778008937835693],
44+
[0.1985698938369751, 0.42670100927352905, 0.3433462381362915, 0.7976388335227966],
45+
[0.8799982666969299, 0.9038419723510742, 0.6627197861671448, 0.2702082693576813],
46+
[0.25236669182777405, 0.8548979163169861, 0.5277146697044373, 0.8021610975265503],
47+
[0.57248854637146, 0.7331425547599792, 0.5190116167068481, 0.7708839178085327],
48+
[0.5688579678535461, 0.4657098650932312, 0.3426889181137085, 0.06820935010910034],
49+
[0.3779241740703583, 0.07962607592344284, 0.9828171133995056, 0.18161284923553467],
50+
[0.8118587136268616, 0.8749616742134094, 0.6884132623672485, 0.5694944262504578],
51+
[0.16097143292427063, 0.46688002347946167, 0.34517204761505127, 0.22503995895385742],
52+
[0.5925118923187256, 0.31226983666419983, 0.9163055419921875, 0.9096355438232422],
53+
[0.257118284702301, 0.11089129745960236, 0.19296273589134216, 0.4995841681957245],
54+
[0.7285856604576111, 0.20819443464279175, 0.2480335533618927, 0.8516718745231628],
55+
[0.4158487319946289, 0.6166850924491882, 0.23366613686084747, 0.10196726024150848],
56+
[0.5158570408821106, 0.47714099287986755, 0.15267165005207062, 0.6218062043190002],
57+
[0.5440101027488708, 0.654137372970581, 0.1445455402135849, 0.7515278458595276],
58+
[0.22204914689064026, 0.5193518400192261, 0.7852960228919983, 0.022330427542328835],
59+
[0.32436245679855347, 0.8729223608970642, 0.8447096347808838, 0.5384405851364136],
60+
[0.8666082620620728, 0.9498059749603271, 0.8264070153236389, 0.8541154265403748],
61+
[0.09874340146780014, 0.651304304599762, 0.703516960144043, 0.6102408170700073],
62+
[0.7996152639389038, 0.034571219235658646, 0.7702387571334839, 0.7317286133766174],
63+
[0.25969839096069336, 0.25706928968429565, 0.6323032975196838, 0.3452974557876587],
64+
[0.796588659286499, 0.4461462199687958, 0.7827494144439697, 0.9904717803001404],
65+
[0.30024832487106323, 0.143005833029747, 0.9013084173202515, 0.5415593981742859],
66+
[0.9747403860092163, 0.6366044282913208, 0.9939129948616028, 0.5460708141326904],
67+
],
68+
dtype=np.float32,
69+
)
70+
71+
POS_EMBED = np.array(
72+
[
73+
[0.5264259576797485, 0.13542790710926056, 0.3557051718235016, 0.026218567043542862],
74+
[0.16039517521858215, 0.7456371784210205, 0.030399689450860023, 0.36654308438301086],
75+
[0.8623462319374084, 0.6926777362823486, 0.6909421682357788, 0.18863679468631744],
76+
[0.4419042766094208, 0.5815774202346802, 0.9897516965866089, 0.20390622317790985],
77+
[0.24773290753364563, 0.2621730864048004, 0.7501724362373352, 0.4569753408432007],
78+
[0.056929439306259155, 0.508516252040863, 0.21196016669273376, 0.7986042499542236],
79+
[0.29733139276504517, 0.027606012299656868, 0.5934324264526367, 0.8438404202461243],
80+
[0.3810161352157593, 0.7498583197593689, 0.5111414790153503, 0.5409517884254456],
81+
[0.9594343304634094, 0.803960919380188, 0.032323066145181656, 0.7093872427940369],
82+
[0.46500149369239807, 0.9475489258766174, 0.22143273055553436, 0.26707202196121216],
83+
[0.08147396147251129, 0.42861881852149963, 0.10901876538991928, 0.6337867379188538],
84+
[0.8029632568359375, 0.6968004703521729, 0.7662113904953003, 0.34245410561561584],
85+
[0.845851480960846, 0.4287687838077545, 0.824009895324707, 0.6264961361885071],
86+
[0.14342305064201355, 0.07838690280914307, 0.018332643434405327, 0.0667250007390976],
87+
[0.458583801984787, 0.11334192007780075, 0.0277833491563797, 0.7548614740371704],
88+
[0.394850492477417, 0.7469384670257568, 0.45240482687950134, 0.4500867426395416],
89+
],
90+
dtype=np.float32,
91+
)
92+
93+
model = onnx.helper.make_model(
94+
opset_imports=[onnx.helper.make_operatorsetid("", 14), onnx.helper.make_operatorsetid("com.microsoft", 1)],
3995
ir_version=7,
4096
graph=make_graph(
4197
name="embed_layernorm_graph",
4298
inputs=[
43-
helper.make_tensor_value_info("input_ids", TensorProto.INT32, shape=[1, 4]),
44-
helper.make_tensor_value_info("segment_ids", TensorProto.INT32, shape=[1, 4]),
99+
onnx.helper.make_tensor_value_info("input_ids", onnx.TensorProto.INT32, shape=[1, 4]),
100+
onnx.helper.make_tensor_value_info("segment_ids", onnx.TensorProto.INT32, shape=[1, 4]),
45101
],
46102
outputs=[
47-
helper.make_tensor_value_info("layernorm_out", TensorProto.FLOAT, shape=[1, 4, 4]),
48-
helper.make_tensor_value_info("mask_index_out", TensorProto.INT32, shape=[1]),
103+
onnx.helper.make_tensor_value_info("layernorm_out", onnx.TensorProto.FLOAT, shape=[1, 4, 4]),
104+
onnx.helper.make_tensor_value_info("mask_index_out", onnx.TensorProto.INT32, shape=[1]),
49105
],
50106
initializer=[
51-
numpy_helper.from_array(
52-
np.load(os.path.join(DATA_DIR, "const0_word_embed.npy")).astype("float32").reshape([32, 4]),
53-
name="word_embed",
54-
),
55-
numpy_helper.from_array(
56-
np.load(os.path.join(DATA_DIR, "const1_pos_embed.npy")).astype("float32").reshape([16, 4]),
57-
name="pos_embed",
58-
),
59-
numpy_helper.from_array(
107+
onnx.numpy_helper.from_array(WORD_EMBED, name="word_embed"),
108+
onnx.numpy_helper.from_array(POS_EMBED, name="pos_embed"),
109+
onnx.numpy_helper.from_array(
60110
np.array(
61111
[0.6185135841369629, 0.010364261455833912, 0.5386272668838501, 0.0030179566238075495],
62112
dtype="float32",
63113
),
64114
name="gamma",
65115
),
66-
numpy_helper.from_array(
116+
onnx.numpy_helper.from_array(
67117
np.array(
68118
[0.9511938095092773, 0.9054020047187805, 0.7959669232368469, 0.9152743220329285], dtype="float32"
69119
),

0 commit comments

Comments
 (0)