diff --git a/examples/python/paddle_densenet.py b/examples/python/paddle_densenet.py
new file mode 100644
index 00000000..c193df8f
--- /dev/null
+++ b/examples/python/paddle_densenet.py
@@ -0,0 +1,80 @@
+
+import paddle
+import paddle.vision.transforms as T
+from paddle.vision.datasets import Cifar10
+from pyinfinitensor.onnx import OnnxStub, backend
+import onnx
+import itertools
+
+def run_cifar_train_and_infer():
+    
+    paddle.device.set_device("gpu")
+
+    transform = T.Compose(
+        [
+            T.Resize(224),
+            T.ToTensor(),
+            T.Normalize(
+                mean=[0.5, 0.5, 0.5],
+                std=[0.5, 0.5, 0.5],
+                to_rgb=True,
+            ),
+        ]
+    )
+    
+    # 下载数据集并初始化 DataSet
+    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
+    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
+
+    # 模型组网并初始化网络
+    densenet = paddle.vision.models.DenseNet(num_classes=10)
+    model = paddle.Model(densenet)
+
+    # 模型训练的配置准备，准备损失函数，优化器和评价指标
+    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
+                paddle.nn.CrossEntropyLoss(),
+                paddle.metric.Accuracy())
+
+    # 模型训练
+    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
+    # 模型评估
+    model.evaluate(test_dataset, batch_size=64, verbose=1)
+
+    # export to ONNX
+    save_path = 'onnx.save/densenet' # 需要保存的路径
+    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
+    paddle.onnx.export(densenet, save_path, input_spec=[x_spec], opset_version=11)
+
+    # 加载onnx模型并放到Infinitensor中
+    model_path = save_path + ".onnx"
+    onnx_model = onnx.load(model_path)
+    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
+    model = gofusion_model
+    model.init()
+
+    # 启动推理
+    cifar10_test = Cifar10(
+        mode="test",
+        transform=transform,  # apply transform to every image
+        backend="cv2",  # use OpenCV as image transform backend
+    )
+    batch_size = 1
+    total_size = 0
+    total_acc = 0.0
+    for data in itertools.islice(iter(cifar10_test), 10000):
+        images, labels = data
+        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
+        model.run()
+        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
+        outputs = paddle.to_tensor(outputs)
+        outputs = paddle.reshape(outputs, (1, 10))
+        labels = paddle.to_tensor(labels)
+        labels = paddle.reshape(labels, (1,1))
+        acc = paddle.metric.accuracy(outputs, labels)
+        total_acc += acc
+        total_size += batch_size
+    print("test acc: {}".format(total_acc.numpy() / total_size))
+
+
+if __name__ == "__main__":
+    run_cifar_train_and_infer()
diff --git a/examples/python/paddle_inception.py b/examples/python/paddle_inception.py
new file mode 100644
index 00000000..9f4d9ca5
--- /dev/null
+++ b/examples/python/paddle_inception.py
@@ -0,0 +1,80 @@
+import paddle
+import paddle.vision.transforms as T
+from paddle.vision.datasets import Cifar10
+from pyinfinitensor.onnx import OnnxStub, backend
+import onnx
+import itertools
+
+def run_cifar_train_and_infer():
+    
+    paddle.device.set_device("gpu")
+
+    transform = T.Compose(
+        [
+            T.Resize(224),
+            T.ToTensor(),
+            T.Normalize(
+                mean=[0.5, 0.5, 0.5],
+                std=[0.5, 0.5, 0.5],
+                to_rgb=True,
+            ),
+        ]
+    )
+    
+    # 下载数据集并初始化 DataSet
+    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
+    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
+
+    # 模型组网并初始化网络
+    inception = paddle.vision.models.InceptionV3(num_classes=10)
+    model = paddle.Model(inception)
+
+    # 模型训练的配置准备，准备损失函数，优化器和评价指标
+    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
+                paddle.nn.CrossEntropyLoss(),
+                paddle.metric.Accuracy())
+
+    # 模型训练
+    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
+    # 模型评估
+    model.evaluate(test_dataset, batch_size=64, verbose=1)
+
+    # export to ONNX
+    save_path = 'onnx.save/inception' # 需要保存的路径
+    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
+    paddle.onnx.export(inception, save_path, input_spec=[x_spec], opset_version=11)
+
+    # 加载onnx模型并放到Infinitensor中
+    model_path = save_path + ".onnx"
+    onnx_model = onnx.load(model_path)
+    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
+    model = gofusion_model
+    model.init()
+
+    # 启动推理
+    cifar10_test = Cifar10(
+        mode="test",
+        transform=transform,  # apply transform to every image
+        backend="cv2",  # use OpenCV as image transform backend
+    )
+    batch_size = 1
+    total_size = 0
+    total_acc = 0.0
+    for data in itertools.islice(iter(cifar10_test), 10000):
+        images, labels = data
+        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
+        model.run()
+        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
+        outputs = paddle.to_tensor(outputs)
+        outputs = paddle.reshape(outputs, (1, 10))
+        labels = paddle.to_tensor(labels)
+        labels = paddle.reshape(labels, (1,1))
+        acc = paddle.metric.accuracy(outputs, labels)
+        total_acc += acc
+        total_size += batch_size
+    print("test acc: {}".format(total_acc.numpy() / total_size))
+
+
+
+if __name__ == "__main__":
+    run_cifar_train_and_infer() 
diff --git a/examples/python/paddle_model_dev.md b/examples/python/paddle_model_dev.md
new file mode 100644
index 00000000..f22355ea
--- /dev/null
+++ b/examples/python/paddle_model_dev.md
@@ -0,0 +1,31 @@
+## Description
+
+This is a doc to tell you how to run paddle*.py in your machine. If your model run on other machines except Nvidia, you may need to make some change.
+
+## What do we do in paddle*.py files?
+
+1. Train model and evalute model with Cifar10 dataset
+
+2. Export paddle model to onnx model
+
+3. Load onnx model, infer with InfiniTensor and calculate the inference accuracy
+
+## Command
+
+1. Go to `/examples/python` folder 
+
+2. Run the following command
+   
+   1. ```
+      python paddle_resnet.py
+      python paddle_densenet.py
+      python paddle_inception.py
+      ```
+
+## What should I do if I use other device(MLU, XPU, NPU)?
+
+You need to change this code:
+
+```
+paddle.device.set_device("gpu") # Change gpu to mlu, xpu or npu
+```
diff --git a/examples/python/paddle_resnet.py b/examples/python/paddle_resnet.py
new file mode 100644
index 00000000..8005f341
--- /dev/null
+++ b/examples/python/paddle_resnet.py
@@ -0,0 +1,81 @@
+
+import paddle
+import paddle.vision.transforms as T
+from paddle.vision.datasets import Cifar10
+from pyinfinitensor.onnx import OnnxStub, backend
+import onnx
+import itertools
+from paddle.vision.models.resnet import BasicBlock
+
+def run_cifar_train_and_infer():
+    
+    paddle.device.set_device("gpu")
+
+    transform = T.Compose(
+        [
+            T.Resize(224),
+            T.ToTensor(),
+            T.Normalize(
+                mean=[0.5, 0.5, 0.5],
+                std=[0.5, 0.5, 0.5],
+                to_rgb=True,
+            ),
+        ]
+    )
+    
+    # 下载数据集并初始化 DataSet
+    train_dataset = paddle.vision.datasets.Cifar10(mode='train', transform=transform)
+    test_dataset = paddle.vision.datasets.Cifar10(mode='test', transform=transform)
+
+    # 模型组网并初始化网络
+    resnet = paddle.vision.models.ResNet(BasicBlock, depth=18, num_classes=10)
+    model = paddle.Model(resnet)
+
+    # 模型训练的配置准备，准备损失函数，优化器和评价指标
+    model.prepare(paddle.optimizer.Adam(parameters=model.parameters()), 
+                paddle.nn.CrossEntropyLoss(),
+                paddle.metric.Accuracy())
+
+    # 模型训练
+    model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)
+    # 模型评估
+    model.evaluate(test_dataset, batch_size=64, verbose=1)
+
+    # export to ONNX
+    save_path = 'onnx.save/resnet' # 需要保存的路径
+    x_spec = paddle.static.InputSpec([1, 3, 224, 224], 'float32', 'x') # 为模型指定输入的形状和数据类型，支持持 Tensor 或 InputSpec ，InputSpec 支持动态的 shape。
+    paddle.onnx.export(resnet, save_path, input_spec=[x_spec], opset_version=11)
+
+    # 加载onnx模型并放到Infinitensor中
+    model_path = save_path + ".onnx"
+    onnx_model = onnx.load(model_path)
+    gofusion_model = OnnxStub(onnx_model, backend.cuda_runtime())
+    model = gofusion_model
+    model.init()
+
+    # 启动推理
+    cifar10_test = Cifar10(
+        mode="test",
+        transform=transform,  # apply transform to every image
+        backend="cv2",  # use OpenCV as image transform backend
+    )
+    batch_size = 1
+    total_size = 0
+    total_acc = 0.0
+    for data in itertools.islice(iter(cifar10_test), 10000):
+        images, labels = data
+        next(model.inputs.items().__iter__())[1].copyin_float(images.reshape([3*224*224]).tolist())
+        model.run()
+        outputs = next(model.outputs.items().__iter__())[1].copyout_float()
+        outputs = paddle.to_tensor(outputs)
+        outputs = paddle.reshape(outputs, (1, 10))
+        labels = paddle.to_tensor(labels)
+        labels = paddle.reshape(labels, (1,1))
+        acc = paddle.metric.accuracy(outputs, labels)
+        total_acc += acc
+        total_size += batch_size
+    print("test acc: {}".format(total_acc.numpy() / total_size))
+
+
+if __name__ == "__main__":
+    run_cifar_train_and_infer()