This paper proposes a deep-learning-based MAC paradigm, DL-MAC, to address the performance degradation of traditional MAC protocols in dense and heterogeneous unlicensed wireless networks. Traditional protocols like CSMA/CA in Wi-Fi networks suffer from increased collisions and extended backoff times, leading to reduced spectrum efficiency. DL-MAC leverages spectrum sensing data from energy detection modules to achieve channel access, rate adaptation, and channel switch functionalities. It integrates a deep neural network (DNN) for channel selection and a recurrent neural network (RNN) for joint design of channel access and rate adaptation. Real-world data collected in the 2.4 GHz frequency band validates DL-MAC's superior performance over traditional algorithms in both single and multi-channel environments, as well as outperforming single-function approaches. DL-MAC remains robust against channel switch overhead, demonstrating effectiveness in multi-channel settings. The paper also introduces a real-world dataset for open-source use, enabling further research. DL-MAC's contributions include a holistic approach integrating channel access, rate adaptation, and channel switch, validated performance on real-world data, and improved throughput and lower latency compared to traditional algorithms. The protocol's joint design of channel access and rate adaptation outperforms single-feature designs, and its multi-channel capabilities are confirmed through experiments showing no significant performance impact from channel switch overhead. The paper also discusses related works, including traditional MAC protocols, rate adaptation algorithms, and channel switch mechanisms, highlighting the limitations of existing approaches and the advantages of DL-MAC. The DL-MAC framework is detailed, including data preprocessing, neural network structures, and implementation steps for channel access, rate adaptation, and channel switch. Experimental results demonstrate DL-MAC's superior performance in throughput, mean delay, and generalization across different channels, compared to traditional and single-function approaches. The study concludes that DL-MAC offers a promising solution for improving spectrum efficiency and adaptability in complex and heterogeneous wireless networks.This paper proposes a deep-learning-based MAC paradigm, DL-MAC, to address the performance degradation of traditional MAC protocols in dense and heterogeneous unlicensed wireless networks. Traditional protocols like CSMA/CA in Wi-Fi networks suffer from increased collisions and extended backoff times, leading to reduced spectrum efficiency. DL-MAC leverages spectrum sensing data from energy detection modules to achieve channel access, rate adaptation, and channel switch functionalities. It integrates a deep neural network (DNN) for channel selection and a recurrent neural network (RNN) for joint design of channel access and rate adaptation. Real-world data collected in the 2.4 GHz frequency band validates DL-MAC's superior performance over traditional algorithms in both single and multi-channel environments, as well as outperforming single-function approaches. DL-MAC remains robust against channel switch overhead, demonstrating effectiveness in multi-channel settings. The paper also introduces a real-world dataset for open-source use, enabling further research. DL-MAC's contributions include a holistic approach integrating channel access, rate adaptation, and channel switch, validated performance on real-world data, and improved throughput and lower latency compared to traditional algorithms. The protocol's joint design of channel access and rate adaptation outperforms single-feature designs, and its multi-channel capabilities are confirmed through experiments showing no significant performance impact from channel switch overhead. The paper also discusses related works, including traditional MAC protocols, rate adaptation algorithms, and channel switch mechanisms, highlighting the limitations of existing approaches and the advantages of DL-MAC. The DL-MAC framework is detailed, including data preprocessing, neural network structures, and implementation steps for channel access, rate adaptation, and channel switch. Experimental results demonstrate DL-MAC's superior performance in throughput, mean delay, and generalization across different channels, compared to traditional and single-function approaches. The study concludes that DL-MAC offers a promising solution for improving spectrum efficiency and adaptability in complex and heterogeneous wireless networks.