diffusers 源码待理解之处

一、训练DreamBooth时，相关代码的细节小计

**

class_labels = timesteps 时，模型的前向传播怎么走？待深入去看

**

利用class_prompt去生成数据，而不是instance_prompt

class DreamBoothDataset(Dataset):
    """
    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
    It pre-processes the images and the tokenizes prompts.
    """

    def __init__(
        self,
        instance_data_root,
        instance_prompt,
        tokenizer,
        class_data_root=None,
        class_prompt=None,
        class_num=None,
        size=512,
        center_crop=False,
        encoder_hidden_states=None,
        class_prompt_encoder_hidden_states=None,
        tokenizer_max_length=None,
    ):
        self.size = size
        self.center_crop = center_crop
        self.tokenizer = tokenizer
        self.encoder_hidden_states = encoder_hidden_states
        self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states
        self.tokenizer_max_length = tokenizer_max_length

        self.instance_data_root = Path(instance_data_root)
        if not self.instance_data_root.exists():
            raise ValueError(f"Instance {self.instance_data_root} images root doesn't exists.")

        self.instance_images_path = list(Path(instance_data_root).iterdir())
        self.num_instance_images = len(self.instance_images_path)
        self.instance_prompt = instance_prompt
        self._length = self.num_instance_images

        if class_data_root is not None:
            self.class_data_root = Path(class_data_root)
            self.class_data_root.mkdir(parents=True, exist_ok=True)
            self.class_images_path = list(self.class_data_root.iterdir())
            if class_num is not None:
                self.num_class_images = min(len(self.class_images_path), class_num)
            else:
                self.num_class_images = len(self.class_images_path)
            self._length = max(self.num_class_images, self.num_instance_images)
            self.class_prompt = class_prompt
        else:
            self.class_data_root = None

        self.image_transforms = transforms.Compose(
            [
                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
                transforms.ToTensor(),
                transforms.Normalize([0.5], [0.5]),
            ]
        )

    def __len__(self):
        return self._length

    def __getitem__(self, index):
        example = {}
        instance_image = Image.open(self.instance_images_path[index % self.num_instance_images])
        instance_image = exif_transpose(instance_image)

        if not instance_image.mode == "RGB":
            instance_image = instance_image.convert("RGB")
        example["instance_images"] = self.image_transforms(instance_image)

        if self.encoder_hidden_states is not None:
            example["instance_prompt_ids"] = self.encoder_hidden_states
        else:
            text_inputs = tokenize_prompt(
                self.tokenizer, self.instance_prompt, tokenizer_max_length=self.tokenizer_max_length
            )
            example["instance_prompt_ids"] = text_inputs.input_ids
            example["instance_attention_mask"] = text_inputs.attention_mask

        if self.class_data_root:
            class_image = Image.open(self.class_images_path[index % self.num_class_images])
            class_image = exif_transpose(class_image)

            if not class_image.mode == "RGB":
                class_image = class_image.convert("RGB")
            example["class_images"] = self.image_transforms(class_image)

            if self.class_prompt_encoder_hidden_states is not None:
                example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states
            else:
                class_text_inputs = tokenize_prompt(
                    self.tokenizer, self.class_prompt, tokenizer_max_length=self.tokenizer_max_length
                )
                example["class_prompt_ids"] = class_text_inputs.input_ids
                example["class_attention_mask"] = class_text_inputs.attention_mask

        return example

def tokenize_prompt(tokenizer, prompt, tokenizer_max_length=None):
    if tokenizer_max_length is not None:
        max_length = tokenizer_max_length
    else:
        max_length = tokenizer.model_max_length

    text_inputs = tokenizer(
        prompt,
        truncation=True,
        padding="max_length",
        max_length=max_length,
        return_tensors="pt",
    )

    return text_inputs

def collate_fn(examples, with_prior_preservation=False):
    has_attention_mask = "instance_attention_mask" in examples[0]

    input_ids = [example["instance_prompt_ids"] for example in examples]
    pixel_values = [example["instance_images"] for example in examples]

    if has_attention_mask:
        attention_mask = [example["instance_attention_mask"] for example in examples]

    # Concat class and instance examples for prior preservation.
    # We do this to avoid doing two forward passes.
    if with_prior_preservation:
        input_ids += [example["class_prompt_ids"] for example in examples]
        pixel_values += [example["class_images"] for example in examples]

        if has_attention_mask:
            attention_mask += [example["class_attention_mask"] for example in examples]

    pixel_values = torch.stack(pixel_values)
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()

    input_ids = torch.cat(input_ids, dim=0)

    batch = {
        "input_ids": input_ids,
        "pixel_values": pixel_values,
    }

    if has_attention_mask:
        attention_mask = torch.cat(attention_mask, dim=0)
        batch["attention_mask"] = attention_mask

    return batch

Dataset和Dataloader的构成

为了避免模型过拟合或者是说语言漂移的情况，需要用模型去用一个普通的prompt先生成样本。

fine-tune text-encoder，但是对显存要求更高

二、训练text to image，相关代码的细节小计

**

1、Dataloader的构建如下，但是为啥没有attention_mask呢？训练DreamBooth时有
2、训练或者微调模型时需要图文数据对，如果没有文本数据，可以用BLIP去生成图像描述的文本，但是文本描述不一定可靠
**

 # Get the datasets: you can either provide your own training and evaluation files (see below)
    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).

    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            args.dataset_name,
            args.dataset_config_name,
            cache_dir=args.cache_dir,
            data_dir=args.train_data_dir,
        )
    else:
        data_files = {}
        if args.train_data_dir is not None:
            data_files["train"] = os.path.join(args.train_data_dir, "**")
        dataset = load_dataset(
            "imagefolder",
            data_files=data_files,
            cache_dir=args.cache_dir,
        )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    column_names = dataset["train"].column_names

    # 6. Get the column names for input/target.
    dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
    if args.image_column is None:
        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
    else:
        image_column = args.image_column
        if image_column not in column_names:
            raise ValueError(
                f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
            )
    if args.caption_column is None:
        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
    else:
        caption_column = args.caption_column
        if caption_column not in column_names:
            raise ValueError(
                f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
            )

    # Preprocessing the datasets.
    # We need to tokenize input captions and transform the images.
    def tokenize_captions(examples, is_train=True):
        captions = []
        for caption in examples[caption_column]:
            if isinstance(caption, str):
                captions.append(caption)
            elif isinstance(caption, (list, np.ndarray)):
                # take a random caption if there are multiple
                captions.append(random.choice(caption) if is_train else caption[0])
            else:
                raise ValueError(
                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
                )
        inputs = tokenizer(
            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
        )
        return inputs.input_ids

    # Preprocessing the datasets.
    train_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    def preprocess_train(examples):
        images = [image.convert("RGB") for image in examples[image_column]]
        examples["pixel_values"] = [train_transforms(image) for image in images]
        examples["input_ids"] = tokenize_captions(examples)
        # images text pixel_values input_ids 4种key
        return examples

    with accelerator.main_process_first():
        if args.max_train_samples is not None:
            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
        # Set the training transforms
        train_dataset = dataset["train"].with_transform(preprocess_train)

    def collate_fn(examples):
        pixel_values = torch.stack([example["pixel_values"] for example in examples])
        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
        input_ids = torch.stack([example["input_ids"] for example in examples])
        return {"pixel_values": pixel_values, "input_ids": input_ids}

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        num_workers=args.dataloader_num_workers,
    )

三、训ControlNet

Dataloader的搭建的代码如下：

---

1、新增conditioning_pixel_values图像数据，用于做可控的生成
2、输入中依旧没有attention-mask，待思考

---

def make_train_dataset(args, tokenizer, accelerator):
    # Get the datasets: you can either provide your own training and evaluation files (see below)
    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).

    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            args.dataset_name,
            args.dataset_config_name,
            cache_dir=args.cache_dir,
        )
    else:
        if args.train_data_dir is not None:
            dataset = load_dataset(
                args.train_data_dir,
                cache_dir=args.cache_dir,
            )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    column_names = dataset["train"].column_names

    # 6. Get the column names for input/target.
    if args.image_column is None:
        image_column = column_names[0]
        logger.info(f"image column defaulting to {image_column}")
    else:
        image_column = args.image_column
        if image_column not in column_names:
            raise ValueError(
                f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
            )

    if args.caption_column is None:
        caption_column = column_names[1]
        logger.info(f"caption column defaulting to {caption_column}")
    else:
        caption_column = args.caption_column
        if caption_column not in column_names:
            raise ValueError(
                f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
            )

    if args.conditioning_image_column is None:
        conditioning_image_column = column_names[2]
        logger.info(f"conditioning image column defaulting to {conditioning_image_column}")
    else:
        conditioning_image_column = args.conditioning_image_column
        if conditioning_image_column not in column_names:
            raise ValueError(
                f"`--conditioning_image_column` value '{args.conditioning_image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
            )

    def tokenize_captions(examples, is_train=True):
        captions = []
        for caption in examples[caption_column]:
            if random.random() < args.proportion_empty_prompts:
                captions.append("")
            elif isinstance(caption, str):
                captions.append(caption)
            elif isinstance(caption, (list, np.ndarray)):
                # take a random caption if there are multiple
                captions.append(random.choice(caption) if is_train else caption[0])
            else:
                raise ValueError(
                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
                )
        inputs = tokenizer(
            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
        )
        return inputs.input_ids

    image_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution),
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    conditioning_image_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution),
            transforms.ToTensor(),
        ]
    )

    def preprocess_train(examples):
        images = [image.convert("RGB") for image in examples[image_column]]
        images = [image_transforms(image) for image in images]

        conditioning_images = [image.convert("RGB") for image in examples[conditioning_image_column]]
        conditioning_images = [conditioning_image_transforms(image) for image in conditioning_images]

        examples["pixel_values"] = images
        examples["conditioning_pixel_values"] = conditioning_images
        examples["input_ids"] = tokenize_captions(examples)

        return examples

    with accelerator.main_process_first():
        if args.max_train_samples is not None:
            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
        # Set the training transforms
        train_dataset = dataset["train"].with_transform(preprocess_train)

    return train_dataset


def collate_fn(examples):
    pixel_values = torch.stack([example["pixel_values"] for example in examples])
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()

    conditioning_pixel_values = torch.stack([example["conditioning_pixel_values"] for example in examples])
    conditioning_pixel_values = conditioning_pixel_values.to(memory_format=torch.contiguous_format).float()

    input_ids = torch.stack([example["input_ids"] for example in examples])

    return {
        "pixel_values": pixel_values,
        "conditioning_pixel_values": conditioning_pixel_values,
        "input_ids": input_ids,
    }