diffusers中sd的微调和lora微调_train_text_to_image_lora.py

train_text_to_image.py

代码：

accelerator = Accelerator()->
 
noise_sheduler = DDPMScheduler.from_pretrained(,"scheduler")-> 
tokenizer = CLIPTokenizer.from_pretrained(,"tokenizer")->
text_encoder = CLIPTokenizer.form_pretrained(,"text_encoder")->
vae = AutoencoderKL.from_pretrained(,"vae")->
unet = UNet2DConditionModel(,'unet')->
 
vae.requires_grad_(False)->
text_encoder.requires_grad_(False)->
 
unet.enable_gradient_checkpoint()->
optimizer_cls = torch.optim.AdamW->
optimizer = optimizer_cls(unet.parameters(),lr,betas,weight_decay,eps)->
 
dataset = load_dataset(dataset_name,dataset_config_name,cache_dir)->
train_transforms = transforms.Compose([])->
train_dataset = dataset['train'].with_transformer(preprocess_train)->
- images = [image.convert("RGB") for image in examples[image_c