# Codename Fork

Last update: Dec 27, 2024

---

# ‎

• The codename fork is a fork of Applio made by Codename.

• This fork has more features compared to others and changes to increase quality.

• This guide will be only talking about the new features since everything else has been covered in the Applio guide.

---

# Are RVC Models Safe?

RVC Models are PyTorch Models, a Python library used for AI. PyTorch uses serialization via Pythons' Pickle Module, converting the model to a file. Since pickle can execute arbitrary code when loading a model, it could be theoretically used for malware, but this fork has a built-in feature to prevent code execution along the model. Also, HuggingFace has a Security Scanner which scans for any unsafe pickle exploits and uses also ClamAV for scanning dangerous files.

---

# Pros & Cons

The pros & cons are subjective to your necessities.

✔️ PROS

❌ CONS

• All of the pros of Applio.
• Supports MRF HiFi-GAN
• Supports RefineGAN
• Has a Warmup Phase option
• Uses the RAdam optimizer
• Avg running loss
• 44.1k Sample rate support
• Mel similarity metric
• SoX resampler
• Checkpointing
• Improved inplace

• More complicated features.

---

# ‎

1. Go to the github repo here. Then find the releases tab and click it.

   ‎ ‎

2. Click on the zip file and download it. Then go into your C drive and extract it.

3. Go into the codename fork folder and run the run-install.bat file then once it's done run go-fork.bat.

---

# ‎

---

# 44.1k Sample Rate:

Under Training there is the option to use 44.1k as your trainng sample rate, however there are currently no pretrains for it.

---

# MRF HifiGAN & RefineGAN:

• In the training section you are given the option to choose your vocoder
  • HiFi-GAN: the default vocoder for RVC.
  • MRF HiFi-GAN: a version of HiFi-GAN with MRF instead of MPD and new loss functions. This has higher fidelity but only works with this fork and the latest non-precompiled version Applio.
  • RefineGAN: an entirely new GAN which is in a experimental phase.

---

# Warmup Phase:

In the training section there is an option to enable a warmup phase and a slider to choose how long it lasts. Do not use this with RAdam since RAdam does this on its own.

• The warmup phase is where the learning rate (lr) linearly increased for a certain amount of epochs, this can be used to prevent large destabilizing updates in the early stages of training.
  • There isn't much testing on what using a warmup in RVC does so expect varying results.

---

# Avg Running Loss:

While training it logs the average loss per epoch as the standard loss and rolling average loss over 5 epochs to evaluate general trends and the model's performance over time.

---

# RAdam Optimizer:

This fork uses the RAdam (Rectified Adam) optim as the default optim instead of the AdamW optim.

• Some of its key features are:
  • Automatically adapts the learning rate and warmup.
  • Requires less hyperparameter tuning.
  • More stable training.
  • Better generalization.
  • Minimal computational overhead.

---

# Checkpointing:

Checkpointing reduces the vram usage, requirement of computation and training speed by 20-30 percent. Enable it If you're an user of a 4GB GPU or if you intend to use a bigger batch size than your gpu can handle.

---

# Upcoming Features:

• More / different configurable optimizers.
• Adjustable hop length for RMVPE.
• Custom initial learning rate per Generator and Discriminator.
• Custom gradient norm value ( from the ui level )
• Ability to delay / headstart the Generator or Discriminator.
• More warmup options ( Cosine anneal and so on ).
• And more...

---

# ‎