def loop_test(data):
ox = data.ox
shape = ox.shape(data)
dim_0 = ox.gather([shape, ox.constant(value=0)], axis=0)
dim_1 = ox.gather(
[shape,
ox.constant(value=np.array([1], dtype=np.int64))],
axis=0)
zeros = ox.constant_of_shape(dim_1, value=0.0)
is_true = ox.constant(value=True)
@onnx_function(outputs=['c_o', 'total_o', 'scan_o'],
output_types=[_Ty.b,
_Ty.F([None]),
_Ty.F([None])],
input_types=[_Ty.I([1]), _Ty.b,
_Ty.F([None])])
def range_body(iter_n, cond, total):
ox = iter_n.ox
iter_scalar = ox.squeeze(iter_n, axes=[0])
col = ox.gather([data, iter_scalar], axis=0)
total = ox.add([total, col])
return (is_true, total, total)
final_total, scan_res = ox.loop(
dim_0,
is_true,
range_body,
inputs=[zeros],
outputs=['final_total', 'scan_res'])
return final_total, scan_res
def test_float16_with_loop(self):
@onnx_function(outputs=['y1', 'y2'],
input_types=[_Ty.F([None, None])],
output_types=[_Ty.F([None]),
_Ty.F([None, None])])
def loop_test(data):
ox = data.ox
shape = ox.shape(data)
dim_0 = ox.gather([shape, ox.constant(value=0)], axis=0)
dim_1 = ox.gather(
[shape,
ox.constant(value=np.array([1], dtype=np.int64))],
axis=0)
zeros = ox.constant_of_shape(dim_1, value=0.0)
is_true = ox.constant(value=True)
@onnx_function(outputs=['c_o', 'total_o', 'scan_o'],
output_types=[_Ty.b,
_Ty.F([None]),
_Ty.F([None])],
input_types=[_Ty.I([1]), _Ty.b,
_Ty.F([None])])
def range_body(iter_n, cond, total):
ox = iter_n.ox
iter_scalar = ox.squeeze(iter_n, axes=[0])
col = ox.gather([data, iter_scalar], axis=0)
total = ox.add([total, col])
return (is_true, total, total)
final_total, scan_res = ox.loop(
dim_0,
is_true,
range_body,
inputs=[zeros],
outputs=['final_total', 'scan_res'])
return final_total, scan_res
m1 = np.array([[2, 3], [4, 5], [6, 7]], dtype=np.float32)
expected_res = loop_test(m1)
model = loop_test.to_model()
f16model = convert_float_to_float16(copy.deepcopy(model))
actual_res = _ort_inference(f16model, {'data': m1.astype(np.float16)})
for expected, actual in zip(expected_res, actual_res):
self.assertTrue(np.allclose(expected, actual))
self.assertTrue(actual.dtype == np.float16)
f16model2 = convert_float_to_float16(copy.deepcopy(model),
keep_io_types=True)
actual_res2 = _ort_inference(f16model2, {'data': m1})
for expected, actual2 in zip(expected_res, actual_res2):
self.assertTrue(np.allclose(expected, actual2))
self.assertTrue(actual2.dtype == np.float32)
def test_float16(self):
@onnx_function(outputs=['z'],
input_types=(_Ty.F([1, 1, 6, 1])),
output_types=[_Ty.f])
def transpose_n_matmul(x):
ox = x.ox # type: OnnxOperatorBuilderX
wm = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
12]).astype(np.float32).reshape([2, 6])
b = ox.constant(value=wm)
a = ox.transpose(x, perm=[0, 1, 3, 2])
c = ox.transpose(b, perm=[1, 0])
return ox.matmul([a, c])
m1 = np.array([[2, 3], [4, 5],
[6, 7]]).astype(np.float32).reshape([1, 1, 6, 1])
expected = transpose_n_matmul(m1)
model = transpose_n_matmul.to_model()
f16model = convert_float_to_float16(copy.deepcopy(model))
actual = _ort_inference(f16model, {'x': m1.astype(np.float16)})
self.assertTrue(np.allclose(expected, actual))
f16model2 = convert_float_to_float16(copy.deepcopy(model),
keep_io_types=True)
actual2 = _ort_inference(f16model2, {'x': m1})
self.assertTrue(np.allclose(expected, actual2))
def loop_test(len):
ox = len.ox
s_len = ox.squeeze(len, axes=[0])
is_true = ox.constant(value=True)
@onnx_function(outputs=['c_o', 'i_o', 'j_o', 'all_i', 'all_j'],
output_types=[_Ty.b, _Ty.f, _Ty.f, _Ty.f, _Ty.f],
input_types=[_Ty.I([1]), _Ty.b, _Ty.F([1]), _Ty.F([1])])
def range_body(iter_n, cond, i, j):
return (is_true,
i + i.ox.constant(value=1.0), j + 2.0, i, j)
one_c = ox.constant(value=-1.0)
y1, y2, y3, y4 = ox.loop(s_len, is_true, range_body, inputs=[one_c, one_c],
outputs=['y1_o', 'y2_o', 'y3_o', 'y4_o'])
return y1, y2, y3, y4
def test_loop(self):
@onnx_function(outputs=['y1', 'y2', 'y3', 'y4'],
input_types=[_Ty.I([1])],
output_types=[_Ty.F([None]), _Ty.F([None]), _Ty.F([None, 1]), _Ty.F([None, 1])])
def loop_test(len):
ox = len.ox
s_len = ox.squeeze(len, axes=[0])
is_true = ox.constant(value=True)
@onnx_function(outputs=['c_o', 'i_o', 'j_o', 'all_i', 'all_j'],
output_types=[_Ty.b, _Ty.f, _Ty.f, _Ty.f, _Ty.f],
input_types=[_Ty.I([1]), _Ty.b, _Ty.F([1]), _Ty.F([1])])
def range_body(iter_n, cond, i, j):
return (is_true,
i + i.ox.constant(value=1.0), j + 2.0, i, j)
one_c = ox.constant(value=-1.0)
y1, y2, y3, y4 = ox.loop(s_len, is_true, range_body, inputs=[one_c, one_c],
outputs=['y1_o', 'y2_o', 'y3_o', 'y4_o'])
return y1, y2, y3, y4
self.assertEqual(
loop_test(np.array([16], dtype=np.int64))[2][4], 3.0)
def test_matmul_opt(self):
@onnx_function(outputs=['z'],
input_types=(_Ty.F([1, 1, 6, 1])),
output_types=[_Ty.f])
def transpose_n_matmul(x):
ox = x.ox # type: OnnxOperatorBuilderX
wm = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]).astype(np.float32).reshape([2, 6])
b = ox.constant(value=wm)
a = ox.transpose(x, perm=[0, 1, 3, 2])
c = ox.transpose(b, perm=[1, 0])
return ox.matmul([a, c])
m1 = np.array([[2, 3], [4, 5], [6, 7]]).astype(np.float32).reshape([1, 1, 6, 1])
expected = transpose_n_matmul(m1)
opted = optimize_onnx_model(transpose_n_matmul.to_model())
actual = _ort_inference(opted, {'x': m1})
self.assertTrue(np.allclose(expected, actual), "The result mismatch")
cast_batch, op_version=operator.target_opset, axes=[0])
apply_cast(scope, cast_batch, operator.output_full_names[2], container, to=onnx_proto.TensorProto.INT32)
apply_identity(scope, box_batch, operator.output_full_names[0], container)
apply_identity(scope, score_batch, operator.output_full_names[1], container)
set_converter(YOLONMSLayer, convert_NMSLayer)
yolo_model_graph_tiny = None
evaluation_model_graph_tiny = None
nms_model_graph_tiny = None
num_classes = 20
@Graph.trace(
input_types=[_Ty.F(shape=['N', 3, 'M1', 'M2']), _Ty.F(shape=['N', 2])],
output_types=[_Ty.F(shape=[1, 'M1', 4]), _Ty.F(shape=[1, num_classes, 'M2']), _Ty.I32(shape=[1, 'M3', 3])],
outputs=["yolonms_layer_1", "yolonms_layer_1_1", "yolonms_layer_1_2"])
def combine_model_tiny(input_1, image_shape):
global yolo_model_graph_tiny
global evaluation_model_graph_tiny
global nms_model_graph_tiny
output_1 = yolo_model_graph_tiny(input_1)
input_2 = output_1 + (image_shape,)
yolo_evaluation_layer_1, yolo_evaluation_layer_2 = evaluation_model_graph_tiny(*input_2)
nms_layer_1_1, nms_layer_1_2, nms_layer_1_3 = nms_model_graph_tiny(yolo_evaluation_layer_1, yolo_evaluation_layer_2)
return nms_layer_1_1, nms_layer_1_2, nms_layer_1_3
yolo_model_graph = None
evaluation_model_graph = None
to=onnx_proto.TensorProto.INT32)
apply_identity(scope, box_batch, operator.output_full_names[0], container)
apply_identity(scope, score_batch, operator.output_full_names[1],
container)
set_converter(YOLONMSLayer, convert_NMSLayer)
yolo_model_graph_tiny = None
evaluation_model_graph_tiny = None
nms_model_graph_tiny = None
@Graph.trace(
input_types=[_Ty.F(shape=['N', 3, 'M1', 'M2']),
_Ty.F(shape=['N', 2])],
output_types=[
_Ty.F(shape=[1, 'M1', 4]),
_Ty.F(shape=[1, 80, 'M2']),
_Ty.I32(shape=[1, 'M3', 3])
],
outputs=["yolonms_layer_1", "yolonms_layer_1_1", "yolonms_layer_1_2"])
def combine_model_tiny(input_1, image_shape):
global yolo_model_graph_tiny
global evaluation_model_graph_tiny
global nms_model_graph_tiny
output_1 = yolo_model_graph_tiny(input_1)
input_2 = output_1 + (image_shape, )
yolo_evaluation_layer_1, yolo_evaluation_layer_2 = evaluation_model_graph_tiny(
*input_2)