Merge pull request #1 from Yesifan/resume

Support resume train

Author: Yesifan

.gitignore

@@ -1,5 +1,6 @@
 # Project ignores
 darkit/tmp
+test/test_datasets_temp
 
 
 # Created by .ignore support plugin (hsz.mobi)

README.md

@@ -34,7 +34,7 @@ pip install -r requirements.txt
   ```
 #### Notes
 - When running `npm run dev`, pay attention to the `server.proxy` configuration in `vite.config.ts`, it should match the address of the `FastAPI` service you started.
-- If you encounter errors like *Couldn't import the plugin "https://cdn.jsdelivr.net/npm/@inlang/message-lint-rule-without-source@latest/dist/index.js"*, it may be that your network cannot access `cdn.jsdelivr.net`. Please find a proxy or acceleration node and replace it in `project.inlang/settings.json`.
+- If you encounter errors like *Couldn't import the plugin "https://xxxxxxxxx.xxx/xxxxxxxxxxxx"*, it may be that your network cannot access `cdn.jsdelivr.net`. Please find a proxy or acceleration node and replace it in `project.inlang/settings.json`.
 
 ### Unit Testing
 #### Writing Unit Tests

README.zh.md

@@ -32,7 +32,7 @@ pip install -r requirements.txt
   ```
 #### 注意事项
 - 运行 `npm run dev` 时需要注意 `vite.config.ts` 中的 `server.proxy` 配置， 应该确保与你启动的 `FastAPI` 服务地址一致。
-- 如果出现 *Couldn't import the plugin "https://cdn.jsdelivr.net/npm/@inlang/message-lint-rule-without-source@latest/dist/index.js"* 类似的错误，可能是你的网络无法访问 `cdn.jsdelivr.net`，请自行寻找代理或者加速节点并在 `project.inlang/settings.json` 中进行替换。
+- 如果出现 *Couldn't import the plugin "https://xxxxxxxxx.xxx/xxxxxxxxxxxx"* 类似的错误，可能是你的网络无法访问 `cdn.jsdelivr.net`，请自行寻找代理或者加速节点并在 `project.inlang/settings.json` 中进行替换。
 
 ### 单元测试
 #### 编写单元测试

darkit/__init__.py

@@ -1 +1 @@
-__version__ = "0.1.10"
+__version__ = "0.1.11"

darkit/core/README.md

@@ -1,11 +1,12 @@
 from . import utils
-from .trainer import Trainer, TrainerConfig, LogFieldnames
+from .trainer import Trainer, FabricTrainer, TrainerConfig, LogFieldnames
 from .predicter import Predicter
 
 
 __all__ = (
     "utils",
     "Trainer",
+    "FabricTrainer",
     "Predicter",
     "TrainerConfig",
     "LogFieldnames",

darkit/core/__init__.py

@@ -3,7 +3,8 @@
 import torch.nn as nn
 from typing import Optional, Union
 from pathlib import Path
-from .utils import MODEL_PATH
+from .lib.inject import inject_script
+from .utils import MODEL_PATH, model as model_utils
 
 
 class Predicter:
@@ -47,54 +48,38 @@ def __new__(
             # 否则返回父类的实例
             return super().__new__(cls)
 
+    @classmethod
+    def get_root(cls) -> Path:
+        return MODEL_PATH
+
     @classmethod
     def get_save_directory(cls, name: str) -> Path:
-        save_directory = MODEL_PATH / name
+        save_directory = cls.get_root() / name
         return save_directory
 
+    @classmethod
+    def get_fork_directory(cls, fork: str) -> Optional[Path]:
+        return model_utils.get_fork_directory(cls.get_root(), fork)
+
     @classmethod
     def get_checkpoint(cls, name: str, checkpoint: Optional[str] = None) -> Path:
         save_directory = cls.get_save_directory(name)
-
-        # 寻找文件夹下的最新的 checkpoint 的 name
-        if checkpoint:
-            # check if the checkpoint exists
-            if not (save_directory / f"{checkpoint}.pth").exists():
-                raise FileNotFoundError(f"checkpoint {checkpoint} not found")
-            return save_directory / f"{checkpoint}.pth"
-        try:
-            checkpoint_path = max(
-                save_directory.glob("*.pth"), key=lambda x: x.stat().st_ctime
-            )
-            # 去掉后缀
-            return checkpoint_path
-        except ValueError:
-            raise FileNotFoundError(f"checkpoint not found in {save_directory}")
+        return model_utils.get_checkpoint(save_directory, checkpoint)
 
     @classmethod
     def get_model_config_json(cls, name: str) -> dict:
         save_directory = cls.get_save_directory(name)
-        with open(save_directory / "config.json", "r") as f:
-            config_dict = json.load(f)
-        return config_dict
+        return model_utils.get_model_config_json(save_directory)
 
     @classmethod
     def get_external_config_json(cls, name: str) -> Optional[dict]:
         save_directory = cls.get_save_directory(name)
-        external_config_path = save_directory / "external_config.json"
-        if external_config_path.exists():
-            with open(external_config_path, "r") as f:
-                config_dict = json.load(f)
-            return config_dict
-        else:
-            return None
+        return model_utils.get_external_config_json(save_directory)
 
     @classmethod
     def get_trainer_config_json(cls, name: str) -> dict:
         save_directory = cls.get_save_directory(name)
-        with open(save_directory / "trainer_config.json", "r") as f:
-            config_dict = json.load(f)
-        return config_dict
+        return model_utils.get_trainer_config_json(save_directory)
 
     @classmethod
     def get_model(cls, name: str, checkpoint: Optional[str] = None) -> nn.Module:
@@ -114,6 +99,16 @@ def get_model(cls, name: str, checkpoint: Optional[str] = None) -> nn.Module:
         print(f"Loading model {sub_class_name} from {checkpoint_path}")
         return sub_class.get_model(name, checkpoint)
 
+    @classmethod
+    def inject_script(cls, model, name: str):
+        external_config = cls.get_external_config_json(name)
+        if external_config:
+            fork = external_config.get("fork", None)
+            fork_directory = cls.get_fork_directory(fork)
+            if fork_directory is not None:
+                model = inject_script(model, fork_directory)
+        return model
+
     @classmethod
     def from_pretrained(
         cls,
@@ -124,6 +119,7 @@ def from_pretrained(
         trainer_config = cls.get_trainer_config_json(name)
         device = device if device else trainer_config.get("device", "cuda")
         model = cls.get_model(name, checkpoint).to(device)
+        model = cls.inject_script(model, name)
         return cls(name, model, device=device)
 
     def _predict(self, ctx):