The radio-over-fiber (RoF) system is promising to support broadband transmission and increased flexibility. To boost channel capacity in multi-carrier RoF systems with variable-rate forward error correction, probabilistic shaping and water-filling-based entropy loading outperforms bit-power loading in terms of achievable information rate. However, its reliance on specific channel conditions limits practical use in channel-dynamic RoF systems, highlighting the need for adaptive entropy loading that requires minimal channel state information. This paper presents a deep neural network-based transfer learning model for adaptive entropy prediction in discrete multi-tone signals, addressing frequency-selective responses in RoF systems. Numerical and experimental results confirm capacity-approaching generalized mutual information (GMI) and smoother normalized GMI (NGMI) performances, consistently achieving the 0.83 NGMI threshold across subcarriers. Unlike traditional methods requiring pre-measured signal-to-noise ratios (SNR), this approach simplifies implementation by using only demodulated data and the received SNR, providing a more channel-independent entropy loading option in dynamic RoF systems.