[跑代码]BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion

Installation(下载代码-装环境)

conda create -n bk-sdm python=3.8
conda activate bk-sdm
git clone https://github.com/Nota-NetsPresso/BK-SDM.git
cd BK-SDM
pip install -r requirements.txt

Note on the torch versions we've used

• torch 1.13.1 for MS-COCO evaluation & DreamBooth finetuning on a single 24GB RTX3090

• torch 2.0.1 for KD pretraining on a single 80GB A10

  火炬2.0.1在单个80GB A100上进行KD预训练

  • 如果A100上总批大小为256的预训练导致gpu内存不足，请检查torch版本并考虑升级到torch>2.0.0。
    我的版本也是torch2.0.1 单个A100（80G）理论上吃的下256batch

  • 

小的例子

PNDM采样器 50步去噪声

等效代码(仅修改SD-v1.4的U-Net，同时保留其文本编码器和图像解码器):

Distillation Pretraining

Our code was based on train_text_to_image.py of Diffusers 0.15.0.dev0. To access the latest version, use this link.

BK-SDM的diffusers版本0.15
我的diffusers版本比较高0.24.0

检测是否能够训练（先下载数据集get_laion_data.sh再运行代码kd_train_toy.sh）

1 一个玩具数据集(11K的img-txt对)下载到。

bash scripts/get_laion_data.sh preprocessed_11k

/data/laion_aes/preprocessed_11k (1.7GB in tar.gz;1.8GB数据文件夹)。
get_laion_data.sh

需要修改，实际就是下载这三个数据集，我自行下载

# https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/data/improved_aesthetics_6.5plus/preprocessed_11k.tar.gz
# https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/data/improved_aesthetics_6.5plus/preprocessed_212k.tar.gz
# https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/data/improved_aesthetics_6.5plus/preprocessed_2256k.tar.gz

我修改后下载文件名 https://... .../preprocessed_11k.tar.gz直接粘贴到网址里面也可以下载
wget S3_URL -0 FILe_PATH

S3_URL 就是这个网址 FILe_PATH 就是下载路径./data/laion_aes/preprocessed_11k

DATA_TYPE=$"preprocessed_11k"  # {preprocessed_11k, preprocessed_212k, preprocessed_2256k}
FILE_NAME="${DATA_TYPE}.tar.gz"
 

DATA_DIR="./data/laion_aes/"
FILE_UNZIP_DIR="${DATA_DIR}${DATA_TYPE}"
FILE_PATH="${DATA_DIR}${FILE_NAME}"

if [ "$DATA_TYPE" = "preprocessed_11k" ] || [ "$DATA_TYPE" = "preprocessed_212k" ]; then
    echo "-> preprocessed_11k or 212k"
    S3_URL="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/data/improved_aesthetics_6.5plus/${FILE_NAME}"
elif [ "$DATA_TYPE" = "preprocessed_2256k" ]; then
    S3_URL="https://netspresso-research-code-release.s3.us-east-2.amazonaws.com/data/improved_aesthetics_6.25plus/${FILE_NAME}"
else
    echo "Something wrong in data folder name"
    exit
fi

wget $S3_URL -O $FILE_PATH
tar -xvzf $FILE_PATH -C $DATA_DIR
echo "downloaded to ${FILE_UNZIP_DIR}"

2 一个小脚本可以用来验证代码的可执行性，并找到与你的GPU匹配的批处理大小。

批量大小为8 (=4×2)，训练BK-SDM-Base 20次迭代大约需要5分钟和22GB的GPU内存。

bash scripts/kd_train_toy.sh

MODEL_NAME="CompVis/stable-diffusion-v1-4"
TRAIN_DATA_DIR="./data/laion_aes/preprocessed_11k" # please adjust it if needed
UNET_CONFIG_PATH="./src/unet_config"

UNET_NAME="bk_small" # option: ["bk_base", "bk_small", "bk_tiny"]
OUTPUT_DIR="./results/toy_"$UNET_NAME # please adjust it if needed

BATCH_SIZE=2
GRAD_ACCUMULATION=4

StartTime=$(date +%s)

CUDA_VISIBLE_DEVICES=1 accelerate launch src/kd_train_text_to_image.py \
  --pretrained_model_name_or_path $MODEL_NAME \
  --train_data_dir $TRAIN_DATA_DIR\
  --use_ema \
  --resolution 512 --center_crop --random_flip \
  --train_batch_size $BATCH_SIZE \
  --gradient_checkpointing \
  --mixed_precision="fp16" \
  --learning_rate 5e-05 \
  --max_grad_norm 1 \
  --lr_scheduler="constant" --lr_warmup_steps=0 \
  --report_to="all" \
  --max_train_steps=20 \
  --seed 1234 \
  --gradient_accumulation_steps $GRAD_ACCUMULATION \
  --checkpointing_steps 5 \
  --valid_steps 5 \
  --lambda_sd 1.0 --lambda_kd_output 1.0 --lambda_kd_feat 1.0 \
  --use_copy_weight_from_teacher \
  --unet_config_path $UNET_CONFIG_PATH --unet_config_name $UNET_NAME \
  --output_dir $OUTPUT_DIR


EndTime=$(date +%s)
echo "** KD training takes $(($EndTime - $StartTime)) seconds."

单GPU训练BK-SDM{Base, Small, Tiny}-0.22M数据训练

bash scripts/get_laion_data.sh preprocessed_212k
bash scripts/kd_train.sh

1 下载数据集preprocessed_212k

2 训练kd_train.sh

(256batch 训练BD-SM-Base 50K轮次需要300hours/53G单卡)

(64batch 训练BD-SM-Base 50K轮次需要60hours/28G单卡) 不理解？

单GPU训练BK-SDM{Base, Small, Tiny}-2.3M数据训练