predict_advanced.py

import gc
import torch
from cog import BasePredictor, Input, Path
from lora_diffusion.cli_lora_pti import train as lora_train
from lora_diffusion import (
    UNET_DEFAULT_TARGET_REPLACE,
    TEXT_ENCODER_DEFAULT_TARGET_REPLACE,
)

from common import (
    random_seed,
    clean_directories,
    extract_zip_and_flatten,
    get_output_filename,
)


class Predictor(BasePredictor):
    def predict(
        self,
        instance_data: Path = Input(
            description="A ZIP file containing your training images (JPG, PNG, etc. size not restricted). These images contain your 'subject' that you want the trained model to embed in the output domain for later generating customized scenes beyond the training images. For best results, use images without noise or unrelated objects in the background.",
        ),
        seed: int = Input(description="A seed for reproducible training", default=1337),
        resolution: int = Input(
            description="The resolution for input images. All the images in the train/validation dataset will be resized to this"
            " resolution.",
            default=512,
        ),
        train_text_encoder: bool = Input(
            description="Whether to train the text encoder",
            default=True,
        ),
        train_batch_size: int = Input(
            description="Batch size (per device) for the training dataloader.",
            default=1,
        ),
        gradient_accumulation_steps: int = Input(
            description="Number of updates steps to accumulate before performing a backward/update pass.",
            default=4,
        ),
        gradient_checkpointing: bool = Input(
            description="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
            default=False,
        ),
        scale_lr: bool = Input(
            description="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
            default=True,
        ),
        lr_scheduler: str = Input(
            description="The scheduler type to use",
            choices=[
                "linear",
                "cosine",
                "cosine_with_restarts",
                "polynomial",
                "constant",
                "constant_with_warmup",
            ],
            default="constant",
        ),
        lr_warmup_steps: int = Input(
            description="Number of steps for the warmup in the lr scheduler.",
            default=0,
        ),
        clip_ti_decay: bool = Input(
            default=True,
            description="Whether or not to perform Bayesian Learning Rule on norm of the CLIP latent.",
        ),
        color_jitter: bool = Input(
            default=True,
            description="Whether or not to use color jitter at augmentation.",
        ),
        continue_inversion: bool = Input(
            default=False,
            description="Whether or not to continue inversion.",
        ),
        continue_inversion_lr: float = Input(
            default=1e-4,
            description="The learning rate for continuing an inversion.",
        ),
        initializer_tokens: str = Input(
            default=None,
            description="The tokens to use for the initializer. If not provided, will randomly initialize from gaussian N(0,0.017^2)",
        ),
        learning_rate_text: float = Input(
            default=1e-5,
            description="The learning rate for the text encoder.",
        ),
        learning_rate_ti: float = Input(
            default=5e-4,
            description="The learning rate for the TI.",
        ),
        learning_rate_unet: float = Input(
            default=1e-4,
            description="The learning rate for the unet.",
        ),
        lora_rank: int = Input(
            default=4,
            description="Rank of the LoRA. Larger it is, more likely to capture fidelity but less likely to be editable. Larger rank will make the end result larger.",
        ),
        lora_dropout_p: float = Input(
            default=0.1,
            description="Dropout for the LoRA layer. Reference LoRA paper for more details.",
        ),
        lora_scale: float = Input(
            default=1.0,
            description="Scaling parameter at the end of the LoRA layer.",
        ),
        lr_scheduler_lora: str = Input(
            description="The scheduler type to use",
            choices=[
                "linear",
                "cosine",
                "cosine_with_restarts",
                "polynomial",
                "constant",
                "constant_with_warmup",
            ],
            default="constant",
        ),
        lr_warmup_steps_lora: int = Input(
            description="Number of steps for the warmup in the lr scheduler.",
            default=0,
        ),
        max_train_steps_ti: int = Input(
            default=500,
            description="The maximum number of training steps for the TI.",
        ),
        max_train_steps_tuning: int = Input(
            default=1000,
            description="The maximum number of training steps for the tuning.",
        ),
        placeholder_token_at_data: str = Input(
            default=None,
            description="If this value is provided as 'X|Y', it will transform target word X into Y at caption. You are required to provide caption as filename (not regarding extension), and Y has to contain placeholder token below. You are also required to set `None` for `use_template` argument to use this feature.",
        ),
        placeholder_tokens: str = Input(
            default="<s1>|<s2>",
            description="The placeholder tokens to use for the initializer. If not provided, will use the first tokens of the data.",
        ),
        use_face_segmentation_condition: bool = Input(
            default=False,
            description="Whether or not to use the face segmentation condition.",
        ),
        use_template: str = Input(
            default="object",
            description="The template to use for the inversion.",
            choices=[
                "object",
                "style",
                "none",
            ],
        ),
        weight_decay_lora: float = Input(
            default=0.001,
            description="The weight decay for the LORA loss.",
        ),
        weight_decay_ti: float = Input(
            default=0.00,
            description="The weight decay for the TI.",
        ),
    ) -> Path:
        if seed is None:
            seed = random_seed()
        print(f"Using seed: {seed}")

        assert (
            train_text_encoder
        ), "train_text_encoder must be True. This will be updated in the future."

        # check that the data is provided
        cog_instance_data = "cog_instance_data"
        cog_class_data = "cog_class_data"
        cog_output_dir = "checkpoints"
        clean_directories([cog_instance_data, cog_output_dir, cog_class_data])

        extract_zip_and_flatten(instance_data, cog_instance_data)

        if use_template == "none":
            use_template = "null"

        # some settings are fixed for the replicate model
        lora_train(
            pretrained_model_name_or_path="./stable-diffusion-v1-5-cache",
            pretrained_vae_name_or_path=None,
            revision=None,
            instance_data_dir=cog_instance_data,
            seed=seed,
            resolution=resolution,
            train_text_encoder=train_text_encoder,
            train_batch_size=train_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            gradient_checkpointing=gradient_checkpointing,
            scale_lr=scale_lr,
            lr_scheduler=lr_scheduler,
            lr_warmup_steps=lr_warmup_steps,
            use_8bit_adam=False,
            mixed_precision="fp16",
            output_dir=cog_output_dir,
            clip_ti_decay=clip_ti_decay,
            color_jitter=color_jitter,
            continue_inversion=continue_inversion,
            continue_inversion_lr=continue_inversion_lr,
            device="cuda:0",
            initializer_tokens=initializer_tokens,
            learning_rate_text=learning_rate_text,
            learning_rate_ti=learning_rate_ti,
            learning_rate_unet=learning_rate_unet,
            lora_clip_target_modules=TEXT_ENCODER_DEFAULT_TARGET_REPLACE,
            lora_rank=lora_rank,
            lora_dropout_p=lora_dropout_p,
            lora_scale=lora_scale,
            lora_unet_target_modules=UNET_DEFAULT_TARGET_REPLACE,
            lr_scheduler_lora=lr_scheduler_lora,
            lr_warmup_steps_lora=lr_warmup_steps_lora,
            max_train_steps_ti=max_train_steps_ti,
            max_train_steps_tuning=max_train_steps_tuning,
            perform_inversion=True,
            placeholder_token_at_data=placeholder_token_at_data,
            placeholder_tokens=placeholder_tokens,
            save_steps=max_train_steps_tuning,
            use_face_segmentation_condition=use_face_segmentation_condition,
            use_template=use_template,
            weight_decay_lora=weight_decay_lora,
            weight_decay_ti=weight_decay_ti,
        )

        gc.collect()
        torch.cuda.empty_cache()

        weights_path = (
            Path(cog_output_dir) / f"step_{max_train_steps_tuning}.safetensors"
        )
        output_path = Path(cog_output_dir) / get_output_filename(instance_data)
        weights_path.rename(output_path)

        return output_path