In watermarking, the watermark embedding strength is crucial, and the introduction of the strength factor can adjust the trade-off between the quality of the encoded image and the accuracy of the recovered message, thus enabling good imperceptibility and robustness of the encoded image. In traditional watermarking methods, the strength factor is selected in relation to the cover image, and based on different images, different strength factors are manually selected or algorithmically derived to adjust the visual effect of the watermarked image. However, due to the subjectivity and inflexibility of traditional algorithms, they can not achieve the effect of adaptive adjustment of watermarked images. Recently, watermarking methods combined with deep learning have gradually occupied the mainstream of this field. In the testing stage, to balance the overall robustness and imperceptibility, the strength factor is no longer selected based on the cover image as in traditional methods. Instead, it is set to a universal value. Therefore, the watermarking method based on deep learning is still in the primary stage of the trial-and-error method. To solve the subjectivity of the hand-designed embedding strength algorithm of the traditional watermarking methods and the low elasticity of the strength factor of the learning method so as to realize the adaptive embedding of watermarks, we propose an adaptive watermarking method with separate training. The proposed method adds a new component, the Adaptor, compared to other frameworks. The Adaptor can adaptively select strength factors to control the embedding strength of the watermark relying on the cover image and secret message. A two-stage training method is used to maintain the stability of the training and to achieve the best results for each component. With the results obtained from our experiments, our proposed method can find the appropriate strength factor and optimize it, resulting in improved robustness and imperceptibility of the watermark. The proposed method shows better results compared to the current state-of-the-art algorithms.