Transparency-Segmentation/config.yaml at main · xycoord/Transparency-Segmentation · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
model:
  base_model_path:
    default: 'stabilityai/stable-diffusion-3.5-medium'
    type: str
    help: Base model for transfer learning
  data_type:
    default: 'bf16'
    type: str
    help: Data type for training

project_config:
  output_dir:
    default: '/workspace/output-e2e'
    type: str
    help: Output directory

dataset:
  dataset_path:
    default: '/workspace/Trans10k/'
    type: str
    help: Path to dataset
  dataset_name:
    default: 'trans10k'
    type: str
    help: Name of dataset
  val_difficulty:
    default: 'mix'
    type: str
    help: Difficulty level for validation dataset ['easy', 'hard', 'mix']
  test_difficulty:
    default: 'hard'
    type: str
    help: Difficulty level for test dataset ['easy', 'hard', 'mix']
  dataloader_num_workers:
    default: 4
    type: int
    help: Number of workers for dataloader

prompt_embeds:
  prompt_embeds_path:
    default: './precomputed_prompt_embeddings/'
    type: str
    help: Path to prompt embeds

training:
  resume_from_checkpoint:
    default: checkpoint-5966
    type: str
    help: Path to checkpoint to resume from
  epochs:
    default: 43
    type: int
    help: Number of epochs to train
  max_train_steps:
    default: null
    type: int
    help: Maximum number of training steps
  train_batch_size:
    default: 4
    type: int
    help: Batch size for training
  gradient_accumulation_steps:
    default: 2 #Change to 4 for 2 GPUs
    type: int
    help: Gradient accumulation steps
  save_checkpoint_steps:
    default: 200
    type: int
    help: Save checkpoint every n steps
  save_checkpoint_epochs:
    default: 1
    type: int
    help: Save checkpoint every n epochs (overrides save_checkpoint_steps)
  checkpoint_offset_warmup:
    default: true
    type: bool
    help: Offset the checkpoint step by the number of warmup steps
  val_steps:
    default: 157
    type: int
    help: Validate every n steps
  num_vals_train:
    default: null
    type: int
    help: Number of validations per val_steps

validation:
  load_checkpoint:
    default: checkpoint-5966
    type: str
    help: Path to checkpoint to load for validation
  num_vals_test:
    default: null
    type: int
    help: Number of validations for test dataset


learning_rate_scheduler:
  lr:
    default: 2e-6 #5e-6 #2.5e-5
    type: float
    help: Learning rate for training
  min_lr_ratio:
    default: 0.1
    type: float
    help: Ratio of the minimum lr to the initial lr
  lr_scheduler:
    default: 'cosine'
    type: str
    help: Learning rate scheduler
  lr_cycles:
    default: 3.5
    type: float
    help: Number of cycles for learning rate scheduler
  lr_warmup_steps:
    default: null
    type: int
    help: Number of warmup steps for learning rate scheduler
  lr_warmup_epochs:
    default: 1
    type: int
    help: Number of warmup epochs for learning rate scheduler (overrides lr_warmup_steps)

optimizer:
  adam_beta1:
    default: 0.9
    type: float
    help: Beta1 for Adam optimizer
  adam_beta2:
    default: 0.999
    type: float
    help: Beta2 for Adam optimizer
  adam_weight_decay:
    default: 0.0
    type: float
    help: Weight decay for Adam optimizer
  adam_epsilon:
    default: 1e-15
    type: float
    help: Epsilon for Adam optimizer

noise:
  weighting_scheme:
    default: 'logit_normal'
    type: str
    help: Weighting scheme for noise ("sigma_sqrt", "logit_normal", "mode", "cosmap")
  logit_mean:
    default: 0.0
    type: float
    help: Mean for logit_normal weighting scheme
  logit_std:
    default: 1.0
    type: float
    help: Std for logit_normal weighting scheme
  mode_scale:
    default: 1.29
    type: float
    help: Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.

gradient_clipping: # Ignore this if using DeepSpeed
  max_grad_norm:
    default: 1.0 #0.1
    type: float
    help: Maximum gradient norm