NebulaeWis commited on
Commit
9fc0aa3
1 Parent(s): 45aac4d

Upload lora_config_1.5fp16_4dim_forlib_spv_uploadver.toml

Browse files
lora_config_1.5fp16_4dim_forlib_spv_uploadver.toml ADDED
@@ -0,0 +1,152 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [Basics]
2
+ pretrained_model_name_or_path = "***"
3
+ train_data_dir = "***"
4
+ resolution = "512,768"
5
+ seed = 23
6
+ max_train_steps = 1000 # This is overwritten by max_train_epochs anyway
7
+ #max_train_epochs = 40
8
+ clip_skip = 2
9
+
10
+ [Save]
11
+ output_dir = "***"
12
+ output_name = "test"
13
+ save_precision = "fp16"
14
+ save_model_as = "safetensors"
15
+ save_every_n_epochs = 9999
16
+ save_every_n_steps = 9999
17
+ save_state = false
18
+ save_last_n_steps_state = 1 # basically saving the last + final state if save_state set to true
19
+ # save_last_n_epochs_state = 1
20
+ # save_n_epoch_ratio = 10
21
+ # save_last_n_epochs = 10
22
+ save_last_n_steps = 200
23
+
24
+ [SDv2]
25
+ v2 = false
26
+ v_parameterization = false
27
+ scale_v_pred_loss_like_noise_pred = false
28
+
29
+ [Network_setup]
30
+ network_dim = 4
31
+ network_alpha = 2
32
+ dim_from_weights = false
33
+ network_dropout = 0
34
+ network_train_unet_only = true
35
+ network_train_text_encoder_only = false
36
+ resume = false
37
+ # network_weights = 'path/to/network_weights'
38
+ # base_weights = 'path/to/base_weights'
39
+ # base_weights_multiplier = 1
40
+
41
+ [LyCORIS]
42
+ network_module = "lycoris.kohya"
43
+ network_args = [ "preset=attn-mlp", "algo=lora",]
44
+
45
+ [Optimizer]
46
+ train_batch_size = 8
47
+ gradient_checkpointing = false
48
+ gradient_accumulation_steps = 1
49
+ optimizer_type = "AdamW8bit"
50
+ unet_lr = 6e-4
51
+ text_encoder_lr = 6e-4
52
+ max_grad_norm = 1.0
53
+ optimizer_args = [ "weight_decay=0.1", "betas=0.9,0.99",]
54
+
55
+ [Lr_scheduler]
56
+ lr_scheduler_type = ""
57
+ lr_scheduler = "constant"
58
+ lr_warmup_steps = 0
59
+ lr_scheduler_num_cycles = 1
60
+ lr_scheduler_power = 1.0 # Polynomial power for polynomial scheduler
61
+ # lr_scheduler_args = ...
62
+
63
+ [Training_preciscion]
64
+ mixed_precision = "fp16"
65
+ full_fp16 = false
66
+ full_bf16= false
67
+ [Further_improvement]
68
+ min_snr_gamma = 0
69
+ # noise_offset = 0.05 # cannot be set with multires_noise
70
+ # adaptive_noise_scale = 0
71
+ multires_noise_discount = 0.3
72
+ multires_noise_iterations = 6
73
+ # scale_weight_norms = 1
74
+
75
+ [ARB]
76
+ enable_bucket = true
77
+ min_bucket_reso = 320
78
+ max_bucket_reso = 960
79
+ bucket_reso_steps = 64
80
+ bucket_no_upscale = false
81
+
82
+ [Captions]
83
+ shuffle_caption = false
84
+ caption_extension = ".txt"
85
+ keep_tokens = 0
86
+ caption_dropout_rate = 0.05
87
+ caption_dropout_every_n_epochs = 0
88
+ caption_tag_dropout_rate = 0.0
89
+ max_token_length = 150
90
+ weighted_captions = false
91
+ token_warmup_min = 1
92
+ token_warmup_step = 0
93
+
94
+ [Attention]
95
+ mem_eff_attn = false
96
+ xformers = true
97
+
98
+ [Data_augmentation]
99
+ color_aug = false
100
+ flip_aug = false
101
+ random_crop = false
102
+
103
+ [Cache_latents]
104
+ cache_latents = true
105
+ vae_batch_size = 1
106
+ cache_latents_to_disk = false
107
+
108
+ [Sampling_during_training]
109
+ sample_sampler = "ddim"
110
+ # sample_every_n_steps = 5000 # overwritten by n_epochs
111
+ # sample_every_n_epochs = 1
112
+ # sample_prompts = "sample_prompts.txt"
113
+
114
+ [Logging]
115
+ logging_dir = "logs_training"
116
+ log_with = "tensorboard"
117
+ log_prefix = "lora_"
118
+ # log_tracker_name = ?
119
+ # wandb_api_key = ?
120
+
121
+ [Dataset]
122
+ max_data_loader_n_workers = 8
123
+ persistent_data_loader_workers = true
124
+ dataset_repeats = 1 # Not sure how this is used
125
+ # dataset_class = package.module.Class
126
+ # dataset_config = ...
127
+
128
+ [Regularization]
129
+ # This is not really needed because you can do regularization by putting everything in train
130
+ # reg_data_dir = "/path/to/reg"
131
+ prior_loss_weight = 1.0
132
+
133
+ [Huggingface]
134
+ save_state_to_huggingface = false
135
+ resume_from_huggingface = false
136
+ async_upload = false
137
+ # There are more arguments
138
+
139
+ [Debugging]
140
+ debug_dataset = false
141
+
142
+ [Deprecated]
143
+ use_8bit_adam = false
144
+ use_lion_optimizer = false
145
+ learning_rate = 0.0002
146
+
147
+ [Others]
148
+ lowram = false
149
+ # in_json = "/path/to/json_metadata"
150
+ # face_crop_aug_range = 2.0
151
+ # vae = "/path/to/vae"
152
+ training_comment = "nebulae"