src/eval.py

import os

# External libraries
import torch
from accelerate import Accelerator
from accelerate.logging import get_logger
from diffusers import AutoencoderKL, DDIMScheduler
from diffusers.utils import check_min_version
from diffusers.utils.import_utils import is_xformers_available
from transformers import CLIPTextModel, CLIPTokenizer

# Custom imports
from src.datasets.dresscode import DressCodeDataset
from src.datasets.vitonhd import VitonHDDataset
from src.mgd_pipelines.mgd_pipe import MGDPipe
from src.mgd_pipelines.mgd_pipe_disentangled import MGDPipeDisentangled
from src.utils.image_from_pipe import generate_images_from_mgd_pipe
from src.utils.set_seeds import set_seed

# Ensure the minimum version of diffusers is installed
check_min_version("0.10.0.dev0")

logger = get_logger(__name__, log_level="INFO")
os.environ["TOKENIZERS_PARALLELISM"] = "true"
os.environ["WANDB_START_METHOD"] = "thread"


def main(args):
    # Initialize Accelerator
    accelerator = Accelerator(mixed_precision=args.get("mixed_precision", "fp16"))
    device = accelerator.device

    # Set the training seed
    if args.get("seed") is not None:
        set_seed(args["seed"])

    # Load scheduler, tokenizer, and models
    val_scheduler = DDIMScheduler.from_pretrained(args["pretrained_model_name_or_path"], subfolder="scheduler")
    val_scheduler.set_timesteps(50, device=device)

    tokenizer = CLIPTokenizer.from_pretrained(
        args["pretrained_model_name_or_path"], subfolder="tokenizer", revision=args.get("revision", None)
    )
    text_encoder = CLIPTextModel.from_pretrained(
        args["pretrained_model_name_or_path"], subfolder="text_encoder", revision=args.get("revision", None)
    )
    vae = AutoencoderKL.from_pretrained(args["pretrained_model_name_or_path"], subfolder="vae", revision=args.get("revision", None))

    # Load UNet
    unet = torch.hub.load(
        repo_or_dir="aimagelab/multimodal-garment-designer",
        source="github",
        model="mgd",
        pretrained=True,
    )

    # Freeze models
    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)

    # Enable memory efficient attention if requested
    if args.get("enable_xformers_memory_efficient_attention", False):
        if is_xformers_available():
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError("xformers is not available. Install it to enable memory-efficient attention.")

    # Set dataset category
    category = [args.get("category", "dresses")]

    # Load dataset
    if args["dataset"] == "dresscode":
        test_dataset = DressCodeDataset(
            dataroot_path=args["dataset_path"],
            phase="test",
            order=args.get("test_order", 0),
            radius=5,
            sketch_threshold_range=(20, 20),
            tokenizer=tokenizer,
            category=category,
            size=(512, 384),
        )
    elif args["dataset"] == "vitonhd":
        test_dataset = VitonHDDataset(
            dataroot_path=args["dataset_path"],
            phase="test",
            order=args.get("test_order", 0),
            sketch_threshold_range=(20, 20),
            radius=5,
            tokenizer=tokenizer,
            size=(512, 384),
        )
    else:
        raise NotImplementedError(f"Dataset {args['dataset']} is not supported.")

    # Prepare dataloader
    test_dataloader = torch.utils.data.DataLoader(
        test_dataset,
        shuffle=False,
        batch_size=args.get("batch_size", 1),
        num_workers=args.get("num_workers_test", 4),
    )

    # Cast models to appropriate precision
    weight_dtype = torch.float32 if args.get("mixed_precision") != "fp16" else torch.float16
    text_encoder.to(device, dtype=weight_dtype)
    vae.to(device, dtype=weight_dtype)
    unet.eval()

    # Select pipeline
    with torch.inference_mode():
        pipeline_class = MGDPipeDisentangled if args.get("disentagle", False) else MGDPipe
        val_pipe = pipeline_class(
            text_encoder=text_encoder,
            vae=vae,
            unet=unet.to(vae.dtype),
            tokenizer=tokenizer,
            scheduler=val_scheduler,
        ).to(device)

        val_pipe.enable_attention_slicing()

        # Prepare dataloader with accelerator
        test_dataloader = accelerator.prepare(test_dataloader)

        # Generate images
        output_path = os.path.join(args["output_dir"], args.get("save_name", "generated_image.png"))
        generate_images_from_mgd_pipe(
            test_order=args.get("test_order", 0),
            pipe=val_pipe,
            test_dataloader=test_dataloader,
            save_name=args.get("save_name", "generated_image"),
            dataset=args["dataset"],
            output_dir=args["output_dir"],
            guidance_scale=args.get("guidance_scale", 7.5),
            guidance_scale_pose=args.get("guidance_scale_pose", 0.5),
            guidance_scale_sketch=args.get("guidance_scale_sketch", 7.5),
            sketch_cond_rate=args.get("sketch_cond_rate", 1.0),
            start_cond_rate=args.get("start_cond_rate", 0.0),
            no_pose=False,
            disentagle=args.get("disentagle", False),
            seed=args.get("seed", None),
        )

    # Return the output image path for verification
    return output_path


if __name__ == "__main__":
    # Example usage for debugging
    example_args = {
        "pretrained_model_name_or_path": "./models",
        "dataset": "dresscode",
        "dataset_path": "./datasets/dresscode",
        "output_dir": "./outputs",
        "guidance_scale": 7.5,
        "guidance_scale_sketch": 7.5,
        "mixed_precision": "fp16",
        "batch_size": 1,
        "seed": 42,
    }
    output_image = main(example_args)
    print(f"Image generated at: {output_image}")