Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Add contrib Q/DQ ops to symbolic shape inference tool #19340

Merged
merged 4 commits into from
Jan 31, 2024
Merged

Add contrib Q/DQ ops to symbolic shape inference tool #19340

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f7e14db
Add Q/DQ operators to ORT's symbolic shape inference script
adrianlizarraga Jan 30, 2024
c9a1dad
Properly propagate the data types for Q/DQ ops
adrianlizarraga Jan 31, 2024
b5a91d2
Add tests
adrianlizarraga Jan 31, 2024
a14b8c0
Merge branch 'main' into adrianl/add-qdq-to-sym-shape-infer
adrianlizarraga Jan 31, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Add tests
  • Loading branch information
@adrianlizarraga
adrianlizarraga committed Jan 31, 2024
commit b5a91d2bbc9b12097f83cfc64a17c18c07520fe1
202 changes: 202 additions & 0 deletions onnxruntime/test/python/onnxruntime_test_python_symbolic_shape_infer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -392,6 +392,208 @@ def test_div_precision(self):
self.assertEqual(len(output_dims), 1)
self.assertEqual(output_dims[0].dim_value, 512)

def test_quantize_linear(self):
"""
Test ONNX QuantizeLinear op.
Check that the output shape is propagated from the first input and that the output data
type comes from the zero-point input.
"""
initializers = [
helper.make_tensor(
"scale",
TensorProto.FLOAT,
[],
[1.0],
),
helper.make_tensor(
"zero_point",
TensorProto.INT8,
[],
[16],
),
]

nodes = [
helper.make_node(
"QuantizeLinear",
inputs=[
"input_f32",
"scale",
"zero_point",
],
outputs=["output_s8"],
),
]

inputs = [
helper.make_tensor_value_info("input_f32", TensorProto.FLOAT, ["b", 2, 3, 4]),
]

outputs = [
helper.make_tensor_value_info("output_s8", TensorProto.UNDEFINED, None),
]

graph = helper.make_graph(nodes, "QuantizeLinear_Test", inputs, outputs, initializers)
model = helper.make_model(graph)

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)

expected_shapes = [
helper.make_tensor_value_info("output_s8", TensorProto.INT8, ["b", 2, 3, 4]),
]
self._check_shapes(graph, inferred.graph, expected_shapes)

def test_quantize_linear_ms_domain(self):
"""
Test QuantizeLinear op ('com.microsoft' domain).
Check that the output shape is propagated from the first input and that the output data
type comes from the zero-point input.
"""
initializers = [
helper.make_tensor(
"scale",
TensorProto.FLOAT,
[],
[1.0],
),
helper.make_tensor(
"zero_point",
TensorProto.UINT16,
[],
[16],
),
]

nodes = [
helper.make_node(
"QuantizeLinear",
inputs=[
"input_f32",
"scale",
"zero_point",
],
outputs=["output_u16"],
domain="com.microsoft",
),
]

inputs = [
helper.make_tensor_value_info("input_f32", TensorProto.FLOAT, ["b", 2, 3, 4]),
]

outputs = [
helper.make_tensor_value_info("output_u16", TensorProto.UNDEFINED, None),
]

graph = helper.make_graph(nodes, "QuantizeLinear_MSDomain_Test", inputs, outputs, initializers)
model = helper.make_model(graph)

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)

expected_shapes = [
helper.make_tensor_value_info("output_u16", TensorProto.UINT16, ["b", 2, 3, 4]),
]
self._check_shapes(graph, inferred.graph, expected_shapes)

def test_quantize_linear_no_zp_input(self):
"""
Test QuantizeLinear op ('com.microsoft' domain).
Check that the output shape is propagated from the first input.
The zero-point input is missing, so the output data type should default to uint8.
"""
initializers = [
helper.make_tensor(
"scale",
TensorProto.FLOAT,
[],
[1.0],
),
]

nodes = [
helper.make_node(
"QuantizeLinear",
inputs=[
"input_f32",
"scale",
],
outputs=["output_u8"],
domain="com.microsoft",
),
]

inputs = [
helper.make_tensor_value_info("input_f32", TensorProto.FLOAT, ["b", 2, 3, 4]),
]

outputs = [
helper.make_tensor_value_info("output_u8", TensorProto.UNDEFINED, None),
]

graph = helper.make_graph(nodes, "QuantizeLinear_NoZP_Test", inputs, outputs, initializers)
model = helper.make_model(graph)

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)

# Check that the output shape is propagated from the first input and that the
# output data type comes from the zero-point input.
expected_shapes = [
helper.make_tensor_value_info("output_u8", TensorProto.UINT8, ["b", 2, 3, 4]),
]
self._check_shapes(graph, inferred.graph, expected_shapes)

def test_dequantize_linear_ms_domain(self):
"""
Test DequantizeLinear operator ('com.microsoft' domain).
Check that the output shape is propagated from the first input and that the output data
type comes from the scale input.
"""
initializers = [
helper.make_tensor(
"scale",
TensorProto.FLOAT,
[],
[1.0],
),
helper.make_tensor(
"zero_point",
TensorProto.UINT16,
[],
[16],
),
]

nodes = [
helper.make_node(
"DequantizeLinear",
inputs=[
"input_u16",
"scale",
"zero_point",
],
outputs=["output_f32"],
domain="com.microsoft",
),
]

inputs = [
helper.make_tensor_value_info("input_u16", TensorProto.UINT16, ["b", 2, 3, 4]),
]

outputs = [
helper.make_tensor_value_info("output_f32", TensorProto.UNDEFINED, None),
]

graph = helper.make_graph(nodes, "DequantizeLinear_MSDomain_Test", inputs, outputs, initializers)
model = helper.make_model(graph)

inferred = SymbolicShapeInference.infer_shapes(model, auto_merge=True)

expected_shapes = [
helper.make_tensor_value_info("output_f32", TensorProto.FLOAT, ["b", 2, 3, 4]),
]
self._check_shapes(graph, inferred.graph, expected_shapes)


class TestSymbolicShapeInferenceForSlice(unittest.TestCase):
def check_slice_of_concat(self, input_dims, start, end, step, expected_output_dim):
Expand Down
Loading