importosfrompprintimportpprintimportnumpyfrompandasimportDataFrameimportmatplotlib.pyplotaspltfromonnximportloadfromonnx_array_api.ext_test_caseimportexample_pathfromonnx_array_api.plotting.text_plotimportonnx_simple_text_plotfromonnx_array_api.validation.diffimporttext_diff,html_difffromonnxruntimeimportGraphOptimizationLevel,InferenceSession,SessionOptionsfromonnx_array_api.ext_test_caseimportmeasure_timefromonnx_array_api.ort.ort_optimizersimportort_optimized_modelfilename=example_path("data/small.onnx")optimized=filename+".optimized.onnx"ifnotos.path.exists(optimized):ort_optimized_model(filename,output=optimized)print(optimized)

data/small.onnx.optimized.onnx

so=SessionOptions()so.graph_optimization_level=GraphOptimizationLevel.ORT_ENABLE_ALLimg=numpy.random.random((1,3,112,112)).astype(numpy.float32)sess=InferenceSession(filename,so,providers=["CPUExecutionProvider"])sess_opt=InferenceSession(optimized,so,providers=["CPUExecutionProvider"])input_name=sess.get_inputs()[0].nameout=sess.run(None,{input_name:img})[0]out_opt=sess_opt.run(None,{input_name:img})[0]ifout.shape!=out_opt.shape:print("ERROR shape are different{out.shape} !={out_opt.shape}")diff=numpy.abs(out-out_opt).max()print(f"Differences:{diff}")

Differences: 0.0

withopen(filename,"rb")asf:model=load(f)print("first model to text...")text1=onnx_simple_text_plot(model,indent=False)print(text1)

first model to text...opset: domain='' version=11input: name='input' type=dtype('float32') shape=['None', 3, 112, 112]init: name='i0' type=float32 shape=(64,)init: name='i1' type=float32 shape=(64,)init: name='i2' type=float32 shape=(64,)init: name='i3' type=float32 shape=(64,)init: name='i4' type=float32 shape=(1, 2, 7, 7)init: name='i5' type=float32 shape=(64, 3, 3, 3)init: name='i6' type=float32 shape=(64,)init: name='i7' type=float32 shape=(64, 64, 3, 3)init: name='i8' type=float32 shape=(64,)init: name='i9' type=float32 shape=(64, 64, 3, 3)init: name='i10' type=float32 shape=(64,)init: name='i11' type=float32 shape=(64, 64, 1, 1)init: name='i12' type=float32 shape=(64,)init: name='i13' type=float32 shape=(64, 1, 1)init: name='i14' type=float32 shape=(64, 1, 1)Conv(input, i5, i6, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[1,1]) -> r0PRelu(r0, i13) -> r1ReduceMean(r1, axes=[1], keepdims=1) -> r2ReduceMax(r1, axes=[1], keepdims=1) -> r3Concat(r2, r3, axis=1) -> r4Conv(r4, i4, dilations=[1,1], group=1, kernel_shape=[7,7], pads=[3,3,3,3], strides=[1,1]) -> r5Sigmoid(r5) -> r6Mul(r6, r1) -> r7BatchNormalization(r7, i0, i1, i2, i3, epsilon=0.00, momentum=0.90) -> r8Conv(r8, i7, i8, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[1,1]) -> r9PRelu(r9, i14) -> r10Conv(r10, i9, i10, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[2,2]) -> r11Conv(r7, i11, i12, dilations=[1,1], group=1, kernel_shape=[1,1], pads=[0,0,0,0], strides=[2,2]) -> r12Add(r11, r12) -> onnx::BatchNormalization_1830output: name='onnx::BatchNormalization_1830' type=dtype('float32') shape=['None', 64, 56, 56]

withopen(optimized,"rb")asf:model=load(f)print("second model to text...")text2=onnx_simple_text_plot(model,indent=False)print(text2)

second model to text...opset: domain='' version=11opset: domain='ai.onnx.ml' version=5opset: domain='ai.onnx.training' version=1opset: domain='ai.onnx.preview.training' version=1opset: domain='com.microsoft' version=1opset: domain='com.microsoft.experimental' version=1opset: domain='com.microsoft.nchwc' version=1opset: domain='org.pytorch.aten' version=1input: name='input' type=dtype('float32') shape=['None', 3, 112, 112]init: name='i0' type=float32 shape=(64,)init: name='i1' type=float32 shape=(64,)init: name='i2' type=float32 shape=(64,)init: name='i3' type=float32 shape=(64,)init: name='reorder_token_10' type=float32 shape=(64, 64, 3, 3)init: name='reorder_token_6' type=float32 shape=(64, 64, 3, 3)init: name='i6' type=float32 shape=(64,)init: name='reorder_token_1' type=float32 shape=(8, 2, 7, 7)init: name='i8' type=float32 shape=(64,)init: name='reorder' type=float32 shape=(64, 3, 3, 3)init: name='i10' type=float32 shape=(64,)init: name='reorder_token_3' type=float32 shape=(64, 64, 1, 1)init: name='i12' type=float32 shape=(64,)init: name='i13' type=float32 shape=(64, 1, 1)init: name='i14' type=float32 shape=(64, 1, 1)Conv[com.microsoft.nchwc](input, reorder, i6, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_0ReorderOutput[com.microsoft.nchwc](reorder_token_0, channels_last=0, channels=64) -> r0PRelu(r0, i13) -> r1ReduceMax(r1, keepdims=1, axes=[1]) -> r3ReduceMean(r1, keepdims=1, axes=[1]) -> r2Concat(r2, r3, axis=1) -> r4Conv[com.microsoft.nchwc](r4, reorder_token_1, activation=b'Sigmoid', auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[7,7], pads=[3,3,3,3]) -> reorder_token_2ReorderOutput[com.microsoft.nchwc](reorder_token_2, channels_last=0, channels=1) -> r6Mul(r6, r1) -> r7BatchNormalization(r7, i0, i1, i2, i3, momentum=0.90, epsilon=0.00) -> r8ReorderInput[com.microsoft.nchwc](r8, channels_last=0) -> reorder_token_7Conv[com.microsoft.nchwc](reorder_token_7, reorder_token_6, i8, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_9ReorderOutput[com.microsoft.nchwc](reorder_token_9, channels_last=0, channels=64) -> r9PRelu(r9, i14) -> r10ReorderInput[com.microsoft.nchwc](r10, channels_last=0) -> reorder_token_11ReorderInput[com.microsoft.nchwc](r7, channels_last=0) -> reorder_token_4Conv[com.microsoft.nchwc](reorder_token_4, reorder_token_3, i12, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[2,2], kernel_shape=[1,1], pads=[0,0,0,0]) -> reorder_token_5Conv[com.microsoft.nchwc](reorder_token_11, reorder_token_10, i10, reorder_token_5, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[2,2], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_13ReorderOutput[com.microsoft.nchwc](reorder_token_13, channels_last=0, channels=64) -> onnx::BatchNormalization_1830output: name='onnx::BatchNormalization_1830' type=dtype('float32') shape=['None', 64, 56, 56]

print("differences...")print(text_diff(text1,text2))

differences...  opset: domain='' version=11+ opset: domain='ai.onnx.ml' version=5+ opset: domain='ai.onnx.training' version=1+ opset: domain='ai.onnx.preview.training' version=1+ opset: domain='com.microsoft' version=1+ opset: domain='com.microsoft.experimental' version=1+ opset: domain='com.microsoft.nchwc' version=1+ opset: domain='org.pytorch.aten' version=1  input: name='input' type=dtype('float32') shape=['None', 3, 112, 112]  init: name='i0' type=float32 shape=(64,)  init: name='i1' type=float32 shape=(64,)  init: name='i2' type=float32 shape=(64,)  init: name='i3' type=float32 shape=(64,)- init: name='i4' type=float32 shape=(1, 2, 7, 7)?             ^^                      ^  ^  ^  ^+ init: name='reorder_token_10' type=float32 shape=(64, 64, 3, 3)?             ^^^^^^^^^^^^^^^^                      ^^  ^^  ^  ^- init: name='i5' type=float32 shape=(64, 3, 3, 3)?             ^^                          ^+ init: name='reorder_token_6' type=float32 shape=(64, 64, 3, 3)?             ^^^^^^^^^^^^^^^                          ^^  init: name='i6' type=float32 shape=(64,)- init: name='i7' type=float32 shape=(64, 64, 3, 3)?             ^^                      ^^  ^^  ^  ^+ init: name='reorder_token_1' type=float32 shape=(8, 2, 7, 7)?             ^^^^^^^^^^^^^^^                      ^  ^  ^  ^  init: name='i8' type=float32 shape=(64,)- init: name='i9' type=float32 shape=(64, 64, 3, 3)?             ^^                          ^^+ init: name='reorder' type=float32 shape=(64, 3, 3, 3)?             ^^^^^^^                          ^  init: name='i10' type=float32 shape=(64,)- init: name='i11' type=float32 shape=(64, 64, 1, 1)?             ^^^+ init: name='reorder_token_3' type=float32 shape=(64, 64, 1, 1)?             ^^^^^^^^^^^^^^^  init: name='i12' type=float32 shape=(64,)  init: name='i13' type=float32 shape=(64, 1, 1)  init: name='i14' type=float32 shape=(64, 1, 1)- Conv(input, i5, i6, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[1,1]) -> r0+ Conv[com.microsoft.nchwc](input, reorder, i6, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_0+ ReorderOutput[com.microsoft.nchwc](reorder_token_0, channels_last=0, channels=64) -> r0  PRelu(r0, i13) -> r1+ ReduceMax(r1, keepdims=1, axes=[1]) -> r3- ReduceMean(r1, axes=[1], keepdims=1) -> r2?                ----------+ ReduceMean(r1, keepdims=1, axes=[1]) -> r2?                          ++++++++++- ReduceMax(r1, axes=[1], keepdims=1) -> r3  Concat(r2, r3, axis=1) -> r4- Conv(r4, i4, dilations=[1,1], group=1, kernel_shape=[7,7], pads=[3,3,3,3], strides=[1,1]) -> r5- Sigmoid(r5) -> r6+ Conv[com.microsoft.nchwc](r4, reorder_token_1, activation=b'Sigmoid', auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[7,7], pads=[3,3,3,3]) -> reorder_token_2+ ReorderOutput[com.microsoft.nchwc](reorder_token_2, channels_last=0, channels=1) -> r6  Mul(r6, r1) -> r7- BatchNormalization(r7, i0, i1, i2, i3, epsilon=0.00, momentum=0.90) -> r8?                                        --------------+ BatchNormalization(r7, i0, i1, i2, i3, momentum=0.90, epsilon=0.00) -> r8?                                                     ++++++++++++++- Conv(r8, i7, i8, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[1,1]) -> r9+ ReorderInput[com.microsoft.nchwc](r8, channels_last=0) -> reorder_token_7+ Conv[com.microsoft.nchwc](reorder_token_7, reorder_token_6, i8, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[1,1], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_9+ ReorderOutput[com.microsoft.nchwc](reorder_token_9, channels_last=0, channels=64) -> r9  PRelu(r9, i14) -> r10- Conv(r10, i9, i10, dilations=[1,1], group=1, kernel_shape=[3,3], pads=[1,1,1,1], strides=[2,2]) -> r11- Conv(r7, i11, i12, dilations=[1,1], group=1, kernel_shape=[1,1], pads=[0,0,0,0], strides=[2,2]) -> r12- Add(r11, r12) -> onnx::BatchNormalization_1830+ ReorderInput[com.microsoft.nchwc](r10, channels_last=0) -> reorder_token_11+ ReorderInput[com.microsoft.nchwc](r7, channels_last=0) -> reorder_token_4+ Conv[com.microsoft.nchwc](reorder_token_4, reorder_token_3, i12, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[2,2], kernel_shape=[1,1], pads=[0,0,0,0]) -> reorder_token_5+ Conv[com.microsoft.nchwc](reorder_token_11, reorder_token_10, i10, reorder_token_5, auto_pad=b'NOTSET', dilations=[1,1], group=1, strides=[2,2], kernel_shape=[3,3], pads=[1,1,1,1]) -> reorder_token_13+ ReorderOutput[com.microsoft.nchwc](reorder_token_13, channels_last=0, channels=64) -> onnx::BatchNormalization_1830  output: name='onnx::BatchNormalization_1830' type=dtype('float32') shape=['None', 64, 56, 56]

print("html differences...")output=html_diff(text1,text2)withopen("diff_html.html","w",encoding="utf-8")asf:f.write(output)print("done.")

html differences...done.

img=numpy.random.random((1,3,112,112)).astype(numpy.float32)t1=measure_time(lambda:sess.run(None,{input_name:img}),repeat=25,number=25)t1["name"]="original"print("Original model")pprint(t1)t2=measure_time(lambda:sess_opt.run(None,{input_name:img}),repeat=25,number=25)t2["name"]="optimized"print("Optimized")pprint(t2)

Original model{'average': np.float64(0.0056790061488049106), 'context_size': 64, 'deviation': np.float64(0.0009349826756959479), 'max_exec': np.float64(0.008692911559919594), 'min_exec': np.float64(0.004763298119942192), 'name': 'original', 'number': 25, 'repeat': 25, 'ttime': np.float64(0.14197515372012276)}Optimized{'average': np.float64(0.005954185577592579), 'context_size': 64, 'deviation': np.float64(0.0014574969062307263), 'max_exec': np.float64(0.0113777117599966), 'min_exec': np.float64(0.0048157119199458975), 'name': 'optimized', 'number': 25, 'repeat': 25, 'ttime': np.float64(0.14885463943981447)}

fig,ax=plt.subplots(1,1,figsize=(12,4))df=DataFrame([t1,t2]).set_index("name")df

ax.bar(df.index,df["average"].values,yerr=df["deviation"].values,capsize=6)ax.set_title("Measure performance of optimized model\nlower is better")plt.grid()fig.savefig("plot_optimization.png")

DownloadJupyternotebook:plot_optimization.ipynb

DownloadPythonsourcecode:plot_optimization.py

Downloadzipped:plot_optimization.zip