Update README.md

README.md

@@ -4,7 +4,9 @@ license: mit
 
 Model convert from [https://github.com/KichangKim/DeepDanbooru](https://github.com/KichangKim/DeepDanbooru)
 
-Usage:
+## Usage:
+
+### Basic use
 
 ```python
 import cv2
@@ -41,3 +43,107 @@ img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
 tags = tagger_predict(img, 0.5)
 print(tags)
 ```
+
+### Multi-gpu batch process
+
+
+```python
+import cv2
+import torch
+import os
+import numpy as np
+import onnxruntime as rt
+from huggingface_hub import hf_hub_download
+from torch.utils.data import DataLoader, Dataset
+from PIL import Image
+from tqdm import tqdm
+from threading import Thread
+
+
+class MyDataset(Dataset):
+    def __init__(self, image_list):
+        self.image_list = image_list
+
+    def __len__(self):
+        length = len(self.image_list)
+        return length
+
+    def __getitem__(self, index):
+        image = Image.open(self.image_list[index]).convert("RGB")
+        image = np.asarray(image)
+        s = 512
+        h, w = image.shape[:-1]
+        h, w = (s, int(s * w / h)) if h > w else (int(s * h / w), s)
+        ph, pw = s - h, s - w
+        image = cv2.resize(image, (w, h), interpolation=cv2.INTER_AREA)
+        image = cv2.copyMakeBorder(image, ph // 2, ph - ph // 2, pw // 2, pw - pw // 2, cv2.BORDER_REPLICATE)
+        image = image.astype(np.float32) / 255
+        image = torch.from_numpy(image)
+        idx = torch.tensor([index], dtype=torch.int32)
+        return image, idx
+
+
+def get_images(path):
+    def file_ext(fname):
+        return os.path.splitext(fname)[1].lower()
+
+    all_files = {
+        os.path.relpath(os.path.join(root, fname), path)
+        for root, _dirs, files in os.walk(path)
+        for fname in files
+    }
+    all_images = sorted(
+        os.path.join(path, fname) for fname in all_files if file_ext(fname) in [".png", ".jpg", ".jpeg"]
+    )
+    print(len(all_images))
+    return all_images
+
+
+def process(all_images, batch_size=8, score_threshold=0.35):
+    predictions = {}
+
+    def work_fn(images, device_id):
+        dataset = MyDataset(images)
+        dataloader = DataLoader(
+            dataset,
+            batch_size=batch_size,
+            shuffle=False,
+            persistent_workers=True,
+            num_workers=4,
+            pin_memory=True,
+        )
+        for data in tqdm(dataloader):
+            image, idxs = data
+            image = image.numpy()
+            probs = tagger_model[device_id].run(None, {"input_1": image})[0]
+            probs = probs.astype(np.float32)
+            bs = probs.shape[0]
+            for i in range(bs):
+                tags = []
+                for prob, label in zip(probs[i].tolist(), tagger_tags):
+                    if prob > score_threshold:
+                        tags.append((label.replace("_", " "), prob))
+                predictions[images[idxs[i].item()]] = tags
+
+    gpu_num = len(tagger_model)
+    image_num = (len(all_images) // gpu_num) + 1
+    ts = [Thread(target=work_fn, args=(all_images[i * image_num:(i + 1) * image_num], i)) for i in range(gpu_num)]
+    for t in ts:
+        t.start()
+    for t in ts:
+        t.join()
+    return predictions
+
+
+gpu_num = 4
+batch_size = 8
+tagger_model_path = hf_hub_download(repo_id="skytnt/deepdanbooru_onnx", filename="deepdanbooru.onnx")
+tagger_model = [
+    rt.InferenceSession(tagger_model_path, providers=['CUDAExecutionProvider'], provider_options=[{'device_id': i}]) for
+    i in range(4)]
+tagger_model_meta = tagger_model[0].get_modelmeta().custom_metadata_map
+tagger_tags = eval(tagger_model_meta['tags'])
+
+all_images = get_images("./data")
+predictions = process(all_images, batch_size)
+```