Post-quantum key-encapsulation mechanisms (KEMs) usually use the well-known Fujisaki-Okamoto (FO) transformation during key decapsulation to achieve chosen ciphertext attack (CCA) security. In the FO transformation, the re-encryption procedure depends on the message. The side-channel leakage of re-encryption can be exploited to recover the coefficients of the secret key under chosen ciphertexts, even if the KEM scheme is CCA secure. However, it still requires a large amount of trace during the profiling and attacking phase. In this work, we introduce the first non-profiled deep learning-based attack on lattice-based KEMs. We propose a chosen ciphertexts method that is suitable for non-profiled deep learning-based attacks. The horizontal chunking strategy is employed to partition the coefficients into chunks, enabling the independent recovery of multiple secret key coefficients within each chunk. Then, we adopt an incremental learning strategy to allow the deep learning model to gradually learn the knowledge of each chunk. Moreover, we push the limits of traditional non-profiled deep learning-based attacks by combining the unsupervised domain adaptation with the correlation distinguisher, eliminating the necessity of neural network training for each possible key guess. The feasibility of the attack is verified by practical experiments for the unprotected and masked implementations of Kyber on the ARM Cortex-M4.