script = "finetune-ldm-lora.py" # not used for now

[wandb]
mode = "offline" # "online", "offline", "disabled"
entity = "acme"
project = "test-lora-training"

[models]
unet = {checkpoint = "/path/to/stable-diffusion-1-5/unet.safetensors"}
text_encoder = {checkpoint = "/path/to/stable-diffusion-1-5/CLIPTextEncoderL.safetensors"}
lda = {checkpoint = "/path/to/stable-diffusion-1-5/lda.safetensors"}

[latent_diffusion]
unconditional_sampling_probability = 0.05
offset_noise = 0.1

[lora]
rank = 16
trigger_phrase = "a spsh photo,"
use_only_trigger_probability = 1.0
unet_targets = ["CrossAttentionBlock2d"]
text_encoder_targets = ["TransformerLayer"]
lda_targets = []

[training]
duration = "1000:epoch"
seed = 0
gpu_index = 0
batch_size = 4
gradient_accumulation = "4:step"
clip_grad_norm = 1.0
# clip_grad_value = 1.0
evaluation_interval = "5:epoch"
evaluation_seed = 1


[optimizer]
optimizer = "Prodigy" # "SGD", "Adam", "AdamW", "AdamW8bit", "Lion8bit"
learning_rate = 1
betas = [0.9, 0.999]
eps = 1e-8
weight_decay = 1e-2

[scheduler]
scheduler_type = "ConstantLR"
update_interval = "1:step"
warmup = "500:step"


[dropout]
dropout_probability = 0.2
use_gyro_dropout = false

[dataset]
hf_repo = "acme/images"
revision = "main"

[checkpointing]
# save_folder = "/path/to/ckpts"
save_interval = "1:step"

[test_diffusion]
num_inference_steps = 30
use_short_prompts = false
prompts = [
        "a cute cat",
        "a cute dog",
        "a cute bird",
        "a cute horse",
]