初步支持 macOS (Apple Silicon)

升级至PP-OCRv5
Tested with Python 3.13
感谢用户 "期望" 提供的macOS设备用于开发测试

.gitignore

@@ -373,5 +373,5 @@ test*_no_sub*.mp4
 /backend/models/big-lama/big-lama.pt
 /test/debug/
 /backend/tools/train/release_model/
-model.onnx
+inference.onnx
 /config/config.json

README.md

@@ -135,7 +135,7 @@ cd <源码所在目录>
 
 #### 4. 安装合适的运行环境
 
-本项目支持 CUDA（NVIDIA显卡加速）、CPU（无 GPU）和 DirectML（AMD、Intel等GPU/APU加速）三种运行模式。
+本项目支持 CUDA (NVIDIA显卡加速)、CPU (无 GPU)、 DirectML (AMD、Intel等GPU/APU加速) 和 macOS (Apple Silicon) 四种运行模式。
 
 ##### (1) CUDA（NVIDIA 显卡用户）
 
@@ -198,7 +198,16 @@ cd <源码所在目录>
   pip install torch==2.7.0 torchvision==0.22.0
   pip install -r requirements.txt
   ```
-
+##### (4) macOS 运行 (Apple Silicon)
+- 适用于 macOS (Apple Silicon) 设备
+- macOS (Intel) 请使用CPU, 强行使用GPU只会更慢
+- macOS (Apple Silicon)上字幕检测精准模式的模型(ch_det_fast/model.onnx)精度似乎不太理想, 推荐使用快速模式
+  ```shell
+  pip install paddlepaddle==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
+  pip install torch==2.7.0 torchvision==0.22.0
+  pip install -r requirements.txt
+  ```
+  > 基于Python3.13版本测试
 #### 4. 运行程序
 
 - 运行图形化界面

README_en.md

@@ -133,7 +133,7 @@ cd <source_code_directory>
 
 #### 4. Install the Appropriate Runtime Environment
 
-This project supports three runtime modes: CUDA (NVIDIA GPU acceleration), CPU (no GPU) and DirectML (AMD, Intel, and other GPUs/APUs).
+This project supports four running modes: CUDA (NVIDIA GPU acceleration), CPU (no GPU), DirectML (AMD, Intel and other GPU/APU acceleration), and macOS (Apple Silicon).
 
 ##### (1) CUDA (For NVIDIA GPU users)
 
@@ -199,6 +199,17 @@ This project supports three runtime modes: CUDA (NVIDIA GPU acceleration), CPU (
   pip install -r requirements.txt
   ```
 
+##### (4) Running on macOS (Apple Silicon)
+- Suitable for macOS (Apple Silicon) devices
+- For macOS (Intel), please use the CPU mode. Forcing GPU usage will only be slower.
+- On macOS (Apple Silicon), the accuracy of the subtitle detection precision mode model (ch_det_fast/model.onnx) does not seem to be ideal. It is recommended to use the fast mode instead.
+  ```shell
+  pip install paddlepaddle==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
+  pip install torch==2.7.0 torchvision==0.22.0
+  pip install -r requirements.txt
+  ```
+  > Tested with Python 3.13
+
 #### 4. Run the program
 
 - Run the graphical interface

backend/main.py

@@ -395,7 +395,7 @@ class SubtitleRemover:
             accelerator_name = self.hardware_accelerator.accelerator_name
             if accelerator_name == 'DirectML' and config.inpaintMode.value in [InpaintMode.STTN_AUTO, InpaintMode.STTN_DET]:
                 model_device = 'DirectML'
-            if self.hardware_accelerator.has_cuda():
+            if self.hardware_accelerator.has_cuda() or self.hardware_accelerator.has_mps():
                 model_device = accelerator_name
         self.append_output(tr['Main']['UseModel'].format(f"{model_friendly_name} ({model_device})"))
 
@@ -446,7 +446,7 @@ class SubtitleRemover:
     @cached_property
     def lama_inpaint(self):
         model_path = os.path.join(self.model_config.LAMA_MODEL_DIR, 'big-lama.pt')
-        device = self.hardware_accelerator.device if self.hardware_accelerator.has_cuda() else torch.device("cpu")
+        device = self.hardware_accelerator.device if self.hardware_accelerator.has_cuda() or self.hardware_accelerator.has_mps() else torch.device("cpu")
         return LamaInpaint(device, model_path)
 
     @cached_property

backend/models/V4/ch_det/fs_manifest.csv

@@ -1,4 +0,0 @@
-filename,filesize,encoding,header
-inference_1.pdiparams,50000000,,
-inference_2.pdiparams,50000000,,
-inference_3.pdiparams,13295054,,

backend/models/V5/ch_det/fs_manifest.csv

@@ -0,0 +1,3 @@
+filename,filesize,encoding,header
+
+inference_1.onnx,50000000,,

backend/tools/common_tools.py

@@ -37,9 +37,11 @@ def is_video_or_image(filename):
     # 检查扩展名是否在定义的视频或图片文件后缀集合中
     return file_extension in video_extensions or file_extension in image_extensions
 
-def merge_big_file_if_not_exists(dir, file):
+def merge_big_file_if_not_exists(dir, file, man_filename = None):
     if file not in os.listdir(dir):
         fs = Filesplit()
+        if man_filename is not None:
+            fs.man_filename = man_filename
         fs.merge(input_dir=dir)
 
 def get_readable_path(path):

backend/tools/hardware_accelerator.py

@@ -21,6 +21,7 @@ class HardwareAccelerator:
     def __init__(self):
         self.__cuda = False
         self.__dml = False
+        self.__mps = False
         self.__onnx_providers = []
         self.__enabled = True
         self.__device = None
@@ -28,6 +29,7 @@ class HardwareAccelerator:
     def initialize(self):
         self.check_directml_available()
         self.check_cuda_available()
+        self.check_mps_available()
         self.load_onnx_providers()
 
     def check_directml_available(self):
@@ -36,6 +38,9 @@ class HardwareAccelerator:
     def check_cuda_available(self):
         self.__cuda = torch.cuda.is_available()
 
+    def check_mps_available(self):
+        self.__mps = torch.backends.mps.is_available() and torch.backends.mps.is_built()
+
     def load_onnx_providers(self):
         try:
             import onnxruntime as ort
@@ -65,7 +70,7 @@ class HardwareAccelerator:
     def has_accelerator(self):
         if not self.__enabled:
             return False
-        return self.__cuda or self.__dml or len(self.__onnx_providers) > 0
+        return self.__cuda or self.__dml or self.__mps or len(self.__onnx_providers) > 0
 
     @property
     def accelerator_name(self):
@@ -75,6 +80,8 @@ class HardwareAccelerator:
             return "DirectML"
         if self.__cuda:
             return "GPU"
+        if self.__mps:
+            return "MPS"
         elif len(self.__onnx_providers) > 0:
             return ", ".join(self.__onnx_providers)
         else:
@@ -83,7 +90,7 @@ class HardwareAccelerator:
     @property
     def onnx_providers(self):
         if not self.__enabled:
-            return []
+            return ["CPUExecutionProvider"]
         return self.__onnx_providers
 
     def has_cuda(self):
@@ -91,6 +98,11 @@ class HardwareAccelerator:
             return False
         return self.__cuda
     
+    def has_mps(self):
+        if not self.__enabled:
+            return False
+        return self.__mps
+
     def set_enabled(self, enable):
         self.__enabled = enable
 
@@ -117,4 +129,6 @@ class HardwareAccelerator:
                     self.__dml = False
             if self.__cuda:
                 return torch.device("cuda:0")
+            if self.__mps:
+                return torch.device("mps")
         return torch.device("cpu")

backend/tools/model_config.py

@@ -10,54 +10,13 @@ class ModelConfig:
         self.STTN_DET_MODEL_PATH = os.path.join(BASE_DIR, 'models', 'sttn-det', 'sttn.pth')
         self.PROPAINTER_MODEL_DIR = os.path.join(BASE_DIR,'models', 'propainter')
         if config.subtitleDetectMode.value == SubtitleDetectMode.Fast:
-            self.DET_MODEL_DIR = os.path.join(BASE_DIR,'models', 'V4', 'ch_det_fast')
+            self.DET_MODEL_DIR = os.path.join(BASE_DIR,'models', 'V5', 'ch_det_fast')
         elif config.subtitleDetectMode.value == SubtitleDetectMode.Accurate:
-            self.DET_MODEL_DIR = os.path.join(BASE_DIR, 'models', 'V4', 'ch_det')
+            self.DET_MODEL_DIR = os.path.join(BASE_DIR, 'models', 'V5', 'ch_det')
         else:
             raise ValueError(f"Invalid subtitle detect mode: {config.subtitleDetectMode.value}")
 
         merge_big_file_if_not_exists(self.LAMA_MODEL_DIR, 'bit-lama.pt')
         merge_big_file_if_not_exists(self.PROPAINTER_MODEL_DIR, 'ProPainter.pth')
-        merge_big_file_if_not_exists(self.DET_MODEL_DIR, 'inference.pdiparams')
-    
-    def convertToOnnxModelIfNeeded(self, model_dir, model_filename="inference.pdmodel", params_filename="inference.pdiparams", opset_version=14):
-        """Converts a Paddle model to ONNX if ONNX providers are available and the model does not already exist."""
-        
-        onnx_model_path = os.path.join(model_dir, "model.onnx")
-
-        if os.path.exists(onnx_model_path):
-            print(f"ONNX model already exists: {onnx_model_path}. Skipping conversion.")
-            return onnx_model_path
-        
-        print(f"Converting Paddle model {model_dir} to ONNX...")
-        model_file = os.path.join(model_dir, model_filename)
-        params_file = os.path.join(model_dir, params_filename) if params_filename else ""
-
-        try:
-            import paddle2onnx
-            # Ensure the target directory exists
-            os.makedirs(os.path.dirname(onnx_model_path), exist_ok=True)
-
-            # Convert and save the model
-            onnx_model = paddle2onnx.export(
-                model_filename=model_file,
-                params_filename=params_file,
-                save_file=onnx_model_path,
-                opset_version=opset_version,
-                auto_upgrade_opset=True,
-                verbose=True,
-                enable_onnx_checker=True,
-                enable_experimental_op=True,
-                enable_optimize=True,
-                custom_op_info={},
-                deploy_backend="onnxruntime",
-                calibration_file="calibration.cache",
-                external_file=os.path.join(model_dir, "external_data"),
-                export_fp16_model=False,
-            )
-
-            print(f"Conversion successful. ONNX model saved to: {onnx_model_path}")
-            return onnx_model_path
-        except Exception as e:
-            print(f"Error during conversion: {e}")
-            return model_dir
+        merge_big_file_if_not_exists(self.DET_MODEL_DIR, 'inference.onnx')
+    

backend/tools/subtitle_detect.py

@@ -1,3 +1,4 @@
+import os
 import sys
 from functools import cached_property
 
@@ -33,8 +34,8 @@ class SubtitleDetect:
         parser = utility.init_args()
         args = parser.parse_args([])
         args.det_algorithm = 'DB'
-        args.det_model_dir = model_config.convertToOnnxModelIfNeeded(model_config.DET_MODEL_DIR) if len(onnx_providers) > 0 else model_config.DET_MODEL_DIR
-        args.use_gpu=hardware_accelerator.has_cuda()
+        args.det_model_dir = os.path.join(model_config.DET_MODEL_DIR, 'inference.onnx') if len(onnx_providers) > 0 else model_config.DET_MODEL_DIR
+        args.use_gpu=False
         args.use_onnx=len(onnx_providers) > 0
         args.onnx_providers=onnx_providers
         return TextDetector(args)

requirements.txt

@@ -18,9 +18,8 @@ protobuf==6.30.2
 av==14.3.0
 einops==0.8.1
 paddleocr==2.10.0
-paddle2onnx==1.3.1
 onnxruntime-gpu==1.20.1; sys_platform == 'linux'
 # do not upgrade, windows 10/11 compatable issues, check hardware_acclerator.py:100
 onnxruntime-directml==1.20.1; sys_platform == 'win32'
-onnxruntime-coreml==1.13.1; sys_platform == 'darwin'
+onnxruntime==1.22.0; sys_platform == 'darwin'
 je-showinfilemanager==1.1.6a4