This paper presents Xception, a novel deep convolutional neural network architecture inspired by Inception modules, where Inception modules have been replaced with depthwise separable convolutions. The authors interpret Inception modules as an intermediate step between regular convolutions and depthwise separable convolutions, which consist of a depthwise convolution followed by a pointwise convolution. This observation leads to the proposal of Xception, which slightly outperforms Inception V3 on the ImageNet dataset and significantly outperforms it on a larger dataset with 350 million images and 17,000 classes. Xception has the same number of parameters as Inception V3, indicating that the performance gains are due to more efficient use of model parameters rather than increased capacity. The paper discusses the Inception hypothesis, which posits that cross-channel and spatial correlations can be decoupled. It also explores the continuum between regular convolutions and depthwise separable convolutions, noting that depthwise separable convolutions are an extreme form of Inception modules. The authors argue that replacing Inception modules with depthwise separable convolutions can improve performance, and present Xception as a model based on this idea. Xception is a convolutional neural network architecture entirely based on depthwise separable convolution layers. It is a linear stack of depthwise separable convolution layers with residual connections, making it easy to define and modify. The architecture is evaluated on two image classification tasks: ImageNet and the JFT dataset. Xception shows marginally better results than Inception V3 on ImageNet and a 4.3% relative improvement on the JFT dataset. The paper also discusses the effects of residual connections and intermediate activations on performance, and notes that Xception is slightly slower than Inception V3 but has the same number of parameters. The authors conclude that depthwise separable convolutions may become a cornerstone of convolutional neural network architecture design in the future, as they offer similar properties to Inception modules but are as easy to use as regular convolution layers. The paper also highlights the importance of residual connections and the potential for further improvements through hyperparameter tuning.This paper presents Xception, a novel deep convolutional neural network architecture inspired by Inception modules, where Inception modules have been replaced with depthwise separable convolutions. The authors interpret Inception modules as an intermediate step between regular convolutions and depthwise separable convolutions, which consist of a depthwise convolution followed by a pointwise convolution. This observation leads to the proposal of Xception, which slightly outperforms Inception V3 on the ImageNet dataset and significantly outperforms it on a larger dataset with 350 million images and 17,000 classes. Xception has the same number of parameters as Inception V3, indicating that the performance gains are due to more efficient use of model parameters rather than increased capacity. The paper discusses the Inception hypothesis, which posits that cross-channel and spatial correlations can be decoupled. It also explores the continuum between regular convolutions and depthwise separable convolutions, noting that depthwise separable convolutions are an extreme form of Inception modules. The authors argue that replacing Inception modules with depthwise separable convolutions can improve performance, and present Xception as a model based on this idea. Xception is a convolutional neural network architecture entirely based on depthwise separable convolution layers. It is a linear stack of depthwise separable convolution layers with residual connections, making it easy to define and modify. The architecture is evaluated on two image classification tasks: ImageNet and the JFT dataset. Xception shows marginally better results than Inception V3 on ImageNet and a 4.3% relative improvement on the JFT dataset. The paper also discusses the effects of residual connections and intermediate activations on performance, and notes that Xception is slightly slower than Inception V3 but has the same number of parameters. The authors conclude that depthwise separable convolutions may become a cornerstone of convolutional neural network architecture design in the future, as they offer similar properties to Inception modules but are as easy to use as regular convolution layers. The paper also highlights the importance of residual connections and the potential for further improvements through hyperparameter tuning.