Upload rita-v15-rembg-upscaled-8gpu/training_config.toml with huggingface_hub
Browse files
rita-v15-rembg-upscaled-8gpu/training_config.toml
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
pretrained_model_name_or_path = "/home/ubuntu/sv-training-texas/models/flux1-dev.safetensors"
|
2 |
+
ae = "/home/ubuntu/sv-training-texas/models/ae.safetensors"
|
3 |
+
t5xxl = "/home/ubuntu/sv-training-texas/models/t5xxl_fp16.safetensors"
|
4 |
+
clip_l = "/home/ubuntu/sv-training-texas/models/clip_l.safetensors"
|
5 |
+
output_dir = "/home/ubuntu/sv-training-texas/kohya_models/rita-kohya-tests/rita-v15-rembg-upscaled-8gpu"
|
6 |
+
dataset_repeats = 10
|
7 |
+
resolution = "512,512"
|
8 |
+
train_batch_size = 2
|
9 |
+
network_dim = 16
|
10 |
+
network_alpha = 16
|
11 |
+
optimizer_type = "adamwschedulefree"
|
12 |
+
unet_lr = 0.0005
|
13 |
+
epoch = 12
|
14 |
+
max_train_steps = 100
|
15 |
+
apply_t5_attn_mask = false
|
16 |
+
cache_latents = true
|
17 |
+
cache_latents_to_disk = true
|
18 |
+
cache_text_encoder_outputs = true
|
19 |
+
cache_text_encoder_outputs_to_disk = true
|
20 |
+
clip_skip = 1
|
21 |
+
discrete_flow_shift = 3.1582
|
22 |
+
full_bf16 = true
|
23 |
+
mixed_precision = "bf16"
|
24 |
+
gradient_accumulation_steps = 1
|
25 |
+
gradient_checkpointing = true
|
26 |
+
guidance_scale = 1.0
|
27 |
+
highvram = true
|
28 |
+
huber_c = 0.1
|
29 |
+
huber_schedule = "snr"
|
30 |
+
loss_type = "l2"
|
31 |
+
lr_scheduler = "cosine_with_restarts"
|
32 |
+
lr_scheduler_args = []
|
33 |
+
lr_scheduler_num_cycles = 3
|
34 |
+
lr_scheduler_power = 1
|
35 |
+
max_data_loader_n_workers = 0
|
36 |
+
max_grad_norm = 1
|
37 |
+
max_timestep = 1000
|
38 |
+
min_snr_gamma = 5
|
39 |
+
model_prediction_type = "raw"
|
40 |
+
network_module = "networks.lora_flux"
|
41 |
+
network_train_unet_only = true
|
42 |
+
noise_offset = 0.1
|
43 |
+
noise_offset_type = "Original"
|
44 |
+
optimizer_args = []
|
45 |
+
prior_loss_weight = 1
|
46 |
+
sdpa = true
|
47 |
+
seed = 42
|
48 |
+
t5xxl_max_token_length = 512
|
49 |
+
text_encoder_lr = []
|
50 |
+
timestep_sampling = "sigmoid"
|
51 |
+
output_name = "lora"
|
52 |
+
save_state_to_huggingface = true
|
53 |
+
save_model_as = "safetensors"
|
54 |
+
save_every_n_epochs = 2
|
55 |
+
save_precision = "fp16"
|
56 |
+
caption_extension = ".txt"
|
57 |
+
xformers = "sdpa"
|
58 |
+
enable_bucket = false
|
59 |
+
dataset_config = "/home/ubuntu/sv-training-texas/kohya_models/rita-kohya-tests/rita-v15-rembg-upscaled-8gpu/dataset_config.json"
|
60 |
+
sample_sampler = "euler"
|
61 |
+
sample_prompts = "kohya_configs/sample_prompts.txt"
|
62 |
+
wandb_api_key = ""
|