Last active
August 20, 2024 03:56
-
-
Save fengyuentau/a119f94fd16721ec9974b8c7b0a45d4c to your computer and use it in GitHub Desktop.
ONNXRuntime performance tests
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import onnxscript as ost | |
| from onnxscript import opset19 as op | |
| import numpy as np | |
| import onnx | |
| import onnxruntime as ort | |
| import time | |
| arr0 = np.array([500], dtype=np.int64) | |
| arr1 = np.array([50], dtype=np.int64) | |
| arr2 = np.array([5], dtype=np.int64) | |
| @ost.script() | |
| def make_topk_2d_axis0(x: ost.FLOAT[1000, 100]) -> (ost.FLOAT[500, 100], ost.INT64[500, 100]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr0)), axis=0) | |
| return values, indices | |
| @ost.script() | |
| def make_topk_2d_axis0_k5(x: ost.FLOAT[1000, 100]) -> (ost.FLOAT[5, 100], ost.INT64[5, 100]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr2)), axis=0) | |
| return values, indices | |
| @ost.script() | |
| def make_topk_2d_axis1(x: ost.FLOAT[1000, 100]) -> (ost.FLOAT[1000, 50], ost.INT64[1000, 50]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr1)), axis=1) | |
| return values, indices | |
| @ost.script() | |
| def make_topk_3d_axis0(x: ost.FLOAT[100, 100, 100]) -> (ost.FLOAT[50, 100, 100], ost.INT64[50, 100, 100]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr1)), axis=0) | |
| return values, indices | |
| @ost.script() | |
| def make_topk_3d_axis1(x: ost.FLOAT[100, 100, 100]) -> (ost.FLOAT[100, 50, 100], ost.INT64[100, 50, 100]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr1)), axis=1) | |
| return values, indices | |
| @ost.script() | |
| def make_topk_3d_axis2(x: ost.FLOAT[100, 100, 100]) -> (ost.FLOAT[100, 100, 50], ost.INT64[100, 100, 50]): | |
| values, indices = op.TopK(x, op.Constant(value=onnx.numpy_helper.from_array(arr1)), axis=2) | |
| return values, indices | |
| models = dict( | |
| topk_2d_axis0=make_topk_2d_axis0, | |
| topk_2d_axis0_k5=make_topk_2d_axis0_k5, | |
| topk_2d_axis1=make_topk_2d_axis1, | |
| topk_3d_axis0=make_topk_3d_axis0, | |
| topk_3d_axis1=make_topk_3d_axis1, | |
| topk_3d_axis2=make_topk_3d_axis2, | |
| ) | |
| input_2d_tensor = np.random.randn(1000, 100).astype(np.float32) | |
| input_3d_tensor = np.random.randn(100, 100, 100).astype(np.float32) | |
| def perf_model(k): | |
| model_proto = models[k].to_model_proto() | |
| net = ort.InferenceSession(model_proto.SerializeToString(), providers=["CPUExecutionProvider"]) | |
| x = input_2d_tensor if "2d" in k else input_3d_tensor | |
| print(x.shape) | |
| for _ in range(10): | |
| net.run([], {"x": x}) | |
| elapsed_times = [] | |
| for _ in range(100): | |
| start = time.time() | |
| net.run([], {"x": x}) | |
| end = time.time() | |
| elapsed_times.append(1000 * (end - start)) | |
| elapsed_times = sorted(elapsed_times) | |
| mean = sum(elapsed_times) / len(elapsed_times) | |
| median = (elapsed_times[49] + elapsed_times[50]) / 2 | |
| minimum = elapsed_times[0] | |
| print(f"{k}: mean={mean:.2f}, median={median:.2f}, min={minimum:.2f}") | |
| for key in models.keys(): | |
| perf_model(key) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment