-
Notifications
You must be signed in to change notification settings - Fork 11
/
train_pcm_lora_sd15.sh
109 lines (101 loc) · 4.06 KB
/
train_pcm_lora_sd15.sh
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
PREFIX=fuyun_PCM_4phases
MODEL_DIR=./stable-diffusion-v1-5
OUTPUT_DIR="outputs_formal/lora_64_$PREFIX"
PROJ_NAME="lora_64_formal_$PREFIX"
accelerate launch --main_process_port 29500 train_pcm_lora_sd15.py \
--pretrained_teacher_model=$MODEL_DIR \
--output_dir=$OUTPUT_DIR \
--tracker_project_nam=$PROJ_NAME \
--mixed_precision=fp16 \
--resolution=512 \
--lora_rank=64 \
--learning_rate=5e-6 --loss_type="huber" --adam_weight_decay=1e-3 \
--max_train_steps=5000 \
--max_train_samples=4000000 \
--dataloader_num_workers=16 \
--w_min=4 \
--w_max=5 \
--validation_steps=500 \
--checkpointing_steps=1000 --checkpoints_total_limit=10 \
--train_batch_size=20 \
--enable_xformers_memory_efficient_attention \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--report_to=wandb \
--resume_from_checkpoint=latest \
--seed=453645634 \
--num_ddim_timesteps=50 \
--multiphase=4 \
--gradient_checkpointing
# multiphase: The number of sub-trajectories that we hope to split the PF-ODE into.
# w_min and w_max: We set a larger value of CFG in our official weights. But we find it would be better to set it a bit smaller.
# You will find from the validation that using larger CFGs generate overexpourse results.
PREFIX="fuyun_PCM_4phases_*"
MODEL_DIR=./stable-diffusion-v1-5
OUTPUT_DIR="outputs_formal/lora_64_$PREFIX"
PROJ_NAME="lora_64_formal_$PREFIX"
accelerate launch --main_process_port 29500 train_pcm_lora_sd15.py \
--pretrained_teacher_model=$MODEL_DIR \
--output_dir=$OUTPUT_DIR \
--tracker_project_nam=$PROJ_NAME \
--mixed_precision=fp16 \
--resolution=512 \
--lora_rank=64 \
--learning_rate=5e-6 --loss_type="huber" --adam_weight_decay=1e-3 \
--max_train_steps=5000 \
--max_train_samples=4000000 \
--dataloader_num_workers=16 \
--w_min=4 \
--w_max=5 \
--validation_steps=500 \
--checkpointing_steps=1000 --checkpoints_total_limit=10 \
--train_batch_size=20 \
--enable_xformers_memory_efficient_attention \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--report_to=wandb \
--resume_from_checkpoint=latest \
--seed=453645634 \
--num_ddim_timesteps=50 \
--multiphase=4 \
--gradient_checkpointing \
--not_apply_cfg_solver \
--proportion_empty_prompts=0.1
# multiphase: The number of sub-trajectories that we hope to split the PF-ODE into.
# w_min and w_max: We set a larger value of CFG in our official weights. But we find it would be better to set it a bit smaller.
# not_apply_cfg_solver and set proportion_empty_prompts to larger than 0 enbale us to use larger CFG for inference.
PREFIX=fuyun_PCM_2phases_adv
MODEL_DIR=./stable-diffusion-v1-5
OUTPUT_DIR="outputs/lora_64_$PREFIX"
PROJ_NAME="lora_64_formal_$PREFIX"
accelerate launch --main_process_port 29500 train_pcm_lora_sd15_adv.py \
--pretrained_teacher_model=$MODEL_DIR \
--output_dir=$OUTPUT_DIR \
--tracker_project_nam=$PROJ_NAME \
--mixed_precision=fp16 \
--resolution=512 \
--lora_rank=64 \
--learning_rate=5e-6 --loss_type="huber" --adam_weight_decay=1e-3 \
--max_train_steps=10000 \
--max_train_samples=4000000 \
--dataloader_num_workers=16 \
--w_min=4 \
--w_max=5 \
--validation_steps=500 \
--checkpointing_steps=1000 --checkpoints_total_limit=10 \
--train_batch_size=20 \
--enable_xformers_memory_efficient_attention \
--gradient_accumulation_steps=1 \
--use_8bit_adam \
--resume_from_checkpoint=latest \
--seed=453645634 \
--report_to=wandb \
--num_ddim_timesteps=50 \
--multiphase=2 \
--gradient_checkpointing \
--adv_weight=0.1 \
--adv_lr=1e-5
# multiphase: The number of sub-trajectories that we hope to split the PF-ODE into.
# w_min and w_max: We set a larger value of CFG in our official weights. But we find it would be better to set it a bit smaller.
# set adv weight adn adv lr for proper training configure
# you will see the results change a bit slower than not using the adv loss but not too slow. Ttraining for 1k iterations is enough for obvious improvement.