batched_inference.py

import torch.multiprocessing as multiprocessing
import torchvision.transforms as transforms
from torch import autocast
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import torch
from torchvision.transforms import InterpolationMode
from tqdm import tqdm
import json
import os

torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.autograd.set_detect_anomaly(False)
torch.autograd.profiler.emit_nvtx(enabled=False)
torch.autograd.profiler.profile(enabled=False)
torch.backends.cudnn.benchmark = True


class ImageDataset(Dataset):
    def __init__(self, image_folder_path, allowed_extensions):
        self.allowed_extensions = allowed_extensions
        self.all_image_paths, self.all_image_names, self.image_base_paths = self.get_image_paths(image_folder_path)
        self.train_size = len(self.all_image_paths)
        print(f"Number of images to be tagged: {self.train_size}")
        self.thin_transform = transforms.Compose([
            transforms.Resize(448, interpolation=InterpolationMode.BICUBIC),
            transforms.CenterCrop(448),
            transforms.ToTensor(),
            # Normalize image
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
        self.normal_transform = transforms.Compose([
            transforms.Resize((448, 448), interpolation=InterpolationMode.BICUBIC),
            transforms.ToTensor(),
            # Normalize image
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

        ])

    def get_image_paths(self, folder_path):
        image_paths = []
        image_file_names = []
        image_base_paths = []
        for root, dirs, files in os.walk(folder_path):
            for file in files:
                if file.lower().split(".")[-1] in self.allowed_extensions:
                    image_paths.append((os.path.abspath(os.path.join(root, file))))
                    image_file_names.append(file.split(".")[0])
                    image_base_paths.append(root)
        return image_paths, image_file_names, image_base_paths

    def __len__(self):
        return len(self.all_image_paths)

    def __getitem__(self, index):
        image = Image.open(self.all_image_paths[index]).convert("RGB")
        ratio = image.height / image.width
        if ratio > 2.0 or ratio < 0.5:
            image = self.thin_transform(image)
        else:
            image = self.normal_transform(image)

        return {
            'image': image,
            "image_name": self.all_image_names[index],
            "image_root": self.image_base_paths[index]
        }


def prepare_model(model_path: str):
    model = torch.load(model_path)
    model.to(memory_format=torch.channels_last)
    model = model.eval()
    return model


def train(tagging_is_running, model, dataloader, train_data, output_queue):
    print('Begin tagging')
    model.eval()
    counter = 0

    with torch.no_grad():
        for i, data in tqdm(enumerate(dataloader), total=int(len(train_data) / dataloader.batch_size)):
            this_data = data['image'].to("cuda")
            with autocast(device_type='cuda', dtype=torch.bfloat16):
                outputs = model(this_data)

                probabilities = torch.nn.functional.sigmoid(outputs)
                output_queue.put((probabilities.to("cpu"), data["image_name"], data["image_root"]))

            counter += 1
    _ = tagging_is_running.get()
    print("Tagging finished!")


def tag_writer(tagging_is_running, output_queue, threshold):
    with open("tags_8034.json", "r") as f:
        tags = json.load(f)
    tags.append("placeholder0")
    tags = sorted(tags)
    tag_count = len(tags)
    assert tag_count == 8035, f"The length of tag list is not correct. Correct: 8035, current: {tag_count}"

    while not (tagging_is_running.qsize() > 0 and output_queue.qsize() > 0):
        tag_probabilities, image_names, image_roots = output_queue.get()
        tag_probabilities = tag_probabilities.tolist()

        for per_image_tag_probabilities, image_name, image_root in zip(tag_probabilities, image_names, image_roots,
                                                                       strict=True):
            this_image_tags = []
            this_image_tag_probabilities = []
            for index, per_tag_probability in enumerate(per_image_tag_probabilities):
                if per_tag_probability > threshold:
                    tag = allowed_tags[index]
                    if "placeholder" not in tag:
                        this_image_tags.append(tag)
                        this_image_tag_probabilities.append(str(int(round(per_tag_probability, 3) * 1000)))
            output_file = os.path.join(image_root, os.path.splitext(image_name)[0] + ".txt")
            with open(output_file, "w", encoding="utf-8") as this_output:
                # set this to true if you want tags separated with commas instead of spaces (will output "tag0, tag1...")
                use_comma_sep = True
                sep = " "
                if use_comma_sep:
                    sep = ", "
                # set this to true if you want to replace underscores with spaces
                remove_underscores = True
                if remove_underscores:
                    this_image_tags = map(lambda e: e.replace('_', ' '), this_image_tags)
                this_output.write(sep.join(this_image_tags))
                # change output_probabilities to True if you want probabilities
                output_probabilities = False
                if output_probabilities:
                    this_output.write("\n")
                    this_output.write(sep.join(this_image_tag_probabilities))


def main():
    image_folder_path = "/path/to/img/folder"
    # all images should be in this folder and/or its subfolders.
    # I will generate a text file for every image.
    model_path = "/path/to/your/model.pth"
    allowed_extensions = {"jpg", "jpeg", "png", "webp"}
    batch_size = 64
    # if you have a 24GB card, you can try 256
    threshold = 0.3

    multiprocessing.set_start_method('spawn')
    output_queue = multiprocessing.Queue()
    tagging_is_running = multiprocessing.Queue(maxsize=5)
    tagging_is_running.put("Running!")

    if not torch.cuda.is_available():
        raise RuntimeError("CUDA is not available!")

    model = prepare_model(model_path).to("cuda")

    dataset = ImageDataset(image_folder_path, allowed_extensions)

    batched_loader = DataLoader(
        dataset,
        batch_size=batch_size,
        shuffle=False,
        num_workers=12,  # if you have a big batch size, a good cpu, and enough cpu memory, try 12
        pin_memory=True,
        drop_last=False,
    )
    process_writer = multiprocessing.Process(target=tag_writer,
                                             args=(tagging_is_running, output_queue, threshold))
    process_writer.start()
    process_tagger = multiprocessing.Process(target=train,
                                             args=(tagging_is_running, model, batched_loader, dataset, output_queue,))
    process_tagger.start()
    process_writer.join()
    process_tagger.join()


if __name__ == "__main__":
    main()