If you do not set `optimizer_config` at all or set it to `null` in train.json, the defaults are `adamw8bit` with standard betas of `(0.9,0.999)`, weight decay `0.01`, and epsilon `1e-8`.
In `optimizer.json` the you can set independent optimizer settings for both the text encoder and unet. If you want shared settings, just fill out the `base` section and leave `text_encoder_overrides` properties null an they will be copied from the `base` section.
If you set the `text_encder_lr_scale` property, the text encoder will be trained with a multiple of the Unet learning rate if it the LR is being copied. If you explicitly set the text encoder LR, the `text_encder_lr_scale` is ignored. `text_encder_lr_scale` is likely to be deprecated in the future, but now is left for backwards compatibility.
For each of the `unet` and `text_encoder` sections, you can set the following properties:
`optimizer` value is the type of optimizer to use. Below are the supported optimizers.
Lucidrains' [implementation](https://github.com/lucidrains/lion-pytorch) of the [lion optimizer](https://arxiv.org/abs/2302.06675). Click links to read more. `Epsilon` is not used by lion.
Recommended settings for lion based on the paper are as follows:
"optimizer": "adamw8bit",
"lr": 1e-7,
"lr_scheduler": "constant",
"betas": [0.9, 0.999],
"epsilon": 1e-8,
"weight_decay": 0.10
The recommendations are based on "1/10th LR" but "10x the weight decay" compared to AdamW when training diffusion models. There are no known recommendations for the CLIP text encoder. Lion converges quickly, so take care with the learning rate, and even lower learning rates may be effective.
LR can be set in `optimizer.json` and excluded from the main CLI arg or train.json but if you use the main CLI arg or set it in the main train.json it will override the setting. This was done to make sure existing behavior will not break. To set LR in the `optimizer.json` make sure to delete `"lr": 1.3e-6` in your main train.json and exclude the CLI arg.
The text encoder LR can run at a different value to the Unet LR. This may help prevent over-fitting, especially if you're training from SD2 checkpoints.
If you're training SD2.1 you will likely experience great benefit from partially freezing the text encoder. You can control text encoder freezing using the `text_encoder_freezing` block in your `optimizer.json`:
```
"text_encoder_freezing": {
"freeze_embeddings": true,
"freeze_front_n_layers": -6,
"freeze_final_layer_norm": false
}
```
The SD2.1 text encoder is arranged as follows:
```
embeddings -> CLIP text encoder (23 layers) -> final layer norm
(The SD1.5 text encoder is similar but it has only 12 CLIP layers.) Typically you would apply freezing starting from the left and moving to the right, although it might be interesting to experiment with different freezing patterns. You can control this using the following parameters:
*`freeze_front_n_layers` freezes the front N layers of the CLIP text encoder. You can also pass null to leave the CLIP layers unfrozen, or negative values to count from the back. In the example above, `-6` will freeze all but the last 6 layers.
Recommended settings for SD2.1 are provided in `optimizerSD21.json`: frozen embeddings, all CLIP layers frozen except for the last 6, final layer norm unfrozen. If you want to experiment, start by trying different values for `freeze_front_n_layers`: `-2` is slower but seems to produce a higher quality model, whereas `-10` is faster but can be more difficult to control.
Betas, weight decay, and epsilon are documented in the [AdamW paper](https://arxiv.org/abs/1711.05101) and there is a wealth of information on the web, but consider those experimental to tweak.
