New Training Options

The MVE-DFL fork introduces new training options, both to help you better see training progress (via new Preview Panel that can include more pictures, and has pictures' file names shown), and to help you train better models via new augmentation options. It also allows for the use of configuration files to save / re-use / manually edit settings without the need to go through all the options every time you run training.

---

New utility options

• Session name
  you can name this training session, and it will be saved in summary.txt so you can later review / remind yourself of the options you selected in this session
• Maximum N backups
  If your disk space is limited, or you simply do not want too many backups, you can limit the maximum amount of backups that are saved (latest ones will replace the earliest one)
• Number of samples to preview
  Number of pictures you wanna preview during training. Too big s number will easily result in preview panel going out of your screen.
• Use old preview panel
  If you would prefer to use the old preview panel, instead of the new one.

New trainer optimizations

• Retrain high loss samples
  Periodically retrains the last 16 "high-loss" samples - the ones that are the most problematic for the model.
• Use fp16
  Experimental. It should increase training/inference speed and reduce model size, but it can crash the model or cause other issues.
• Max cpu cores to use
  If you need or want to limit how many CPU cores will be used. Recommended not to go over the number of physical cores. 8 is considered a good value
• Loss function
  You can change the loss function that is used for image quality assessment. Options are SSIM, MS-SSIM, or MS-SSIM+L1
• Learning rate
  The rate at which the model learns. Lower value can reduce the chance for model collapse and increase the amount of learned details, but it will slow down the training. Default value is 5e-05 (0.00005)

New data augmentations

All of these challenge the model which results in better model learning, at the cost of training time.

• Enable random downsample of samples
  Challenges model by making some samples smaller.
• Enable random noise added to samples
  Challenges model by adding noise to some samples.
• Enable random blur of samples
  Challenges model by adding blur to some samples.
• Enable random jpeg compression of samples
  Challenges model by applying jpeg compression's quality degradation to some samples.
• Enable random shadows and highlights of samples
  Helps to create dark light areas in dataset. Use src if you src dataset has a lack of shadows / different lighting situations; dst to help generalization; or all for both reasons.
• Random color
  Samples are randomly rotated around the L axis in LAB colorspace, helps generalize training