From f2344aa12bee6b51ed680a09863682b2fab5aa5a Mon Sep 17 00:00:00 2001 From: admin Date: Fri, 20 Sep 2024 11:13:21 +0800 Subject: [PATCH] upd md future work --- README.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index cdd9aca..fd4bf2e 100644 --- a/README.md +++ b/README.md @@ -52,10 +52,14 @@ In different control modes, generate music segments using specified emotional pr | None | 0.910 | 0.909 | ![](./figs/mat-none.jpg) | ## Future work -Referring to the RLBH of InstructGPT, we will introduce a PPO reinforcement learning fine-tuning optimization for the tunesformer model as well. +However, our current work still faces several limitations. For instance, conclusions derived from statistical correlations only provide a rough guide for designing emotional templates and do not fully reflect the true distribution of features within the emotional space. Additionally, due to the relatively small amount of data and the concentration on pop and game music styles, our analysis results are susceptible to [Simpson's Paradox](https://en.wikipedia.org/wiki/Simpson%27s_paradox). Furthermore, melody generation based on emotional control templates often results in music that is concentrated on a few specified emotions, rather than representing the complete emotional quadrant. For instance, when aiming to generate music for the Q2 quadrant, specifying templates may lead to a concentration on tense music, whereas anger and some other emotions also fall within Q2. Although this approach allows for high precision in 4Q representation, it may lead to a lack of emotional diversity in the generated music. + +To address these issues, we have released an application demonstration on [HuggingFace](https://huggingface.co/spaces/monetjoe/EMusicGen) based on the inference code of our generation system. This demonstration enables users to design and specify emotional templates, utilizing large-scale data to progressively refine feature distributions for greater accuracy. Additionally, the error-free rate is merely a necessary condition for quality but does not fully reflect the true quality of the generated melodies. Future work could incorporate reinforcement learning feedback in the demonstration to adjust the system's generation quality based on user-generated evaluations. Furthermore, while this study focuses on melody, chords are a crucial factor influencing musical emotion. Therefore, our demonstration also includes an option to add chords, and their impact will be considered in future research. ![](./figs/ppo.png) +Referring to the RLBH of InstructGPT, we will introduce a PPO reinforcement learning fine-tuning optimization for the tunesformer model as well. + ## Cite ```bibtex @article{Zhou2024EMusicGen,