fix autodiff

README.md

@@ -4,15 +4,24 @@
 
 **Terox is an open source tiny Deep Learning System based on Python, Cython and CUDA.**
 
+Terox is a tiny Python package that provides some features:
+- [x] Support automatic differentiation.
+- [ ] Provides convenient tensor calculation.
+- [ ] Control the parameters and the model.
+- [ ] Provides common computing functions for deep learning.
+- [ ] Provides common deep learning components.
+- [ ] Provides deep learning model optimizer.
+- [ ] Accelerate computing on CPU and GPU.
+- [ ] Support distributed computing.
+
 ---
 
 ## Setup
 
 Terox requires **Python 3.8** or higher. To check your version of Python, run either:
 
 ```Shell
-python --version
-python3 --version
+python --version # expect python version >= 3.8
 ```
 
 The next step is to install packages. There are several packages used throughout Terox, and you can install them in your enviroment by running:
@@ -31,5 +40,4 @@ Make sure that everything is installed by running python and then checking. If y
 
 ```Python
 import terox
-print(terox.__version__)
-```
+print(terox.__version__) # expect output: "Terox v0.1 by Tokisakix."

README_cn.md

@@ -4,15 +4,24 @@
 
 **Terox 是一个基于 Python、Cython 和 CUDA 的开源微型深度学习系统。**
 
+Terox 是一个很精简的 Python 包，它提供了一些特性:
+- [x] 支持自动微分。
+- [ ] 提供方便的张量计算。
+- [ ] 便捷控制参数和模型。
+- [ ] 提供深度学习常用的计算函数。
+- [ ] 提供常用的深度学习组件。
+- [ ] 提供深度学习模型优化器。
+- [ ] 提高在 CPU 和 GPU 上的计算速度。
+- [ ] 支持分布式计算。
+
 ---
 
 ## 设置
 
 Terox 要求 **Python 3.8** 或更高版本。要检查你的 Python 的版本，请运行:
 
 ```Shell
-python --version
-python3 --version
+python --version # 期望 python 版本 >= 3.8
 ```
 
 下一步是安装第三方软件包。在 Terox 项目中使用了几个包，您可以通过运行以下命令将它们安装到您的环境中:
@@ -31,5 +40,5 @@ python -m pip install -Ue .
 
 ```Python
 import terox
-print(terox.__version__)
+print(terox.__version__) # 期望输出: "Terox v0.1 by Tokisakix."
 ```

terox/autodiff/scalar.py

@@ -1,14 +1,20 @@
-from typing import Optional
+from typing import Iterable, Optional
 
 from .variable import Variable, VarHistory
 from .scalar_opts import ScalarOptsBackend
 
+scalar_count:int = 0
+
 class Scalar(Variable):
 
+    _id: int
     _item: float
 
     def __init__(self, _item:float=None, _history:Optional[VarHistory]=None, _gradient:Optional["Scalar"]=None, _backend:ScalarOptsBackend=ScalarOptsBackend()) -> None:
         super().__init__(_history, _gradient)
+        global scalar_count
+        self._id = scalar_count
+        scalar_count += 1
         if _item == None:
             _item = 0.0
         self._item = float(_item)

terox/autodiff/variable.py

@@ -25,6 +25,7 @@ def __init__(self, _func:VarFunction, _args:Iterable["Variable"]) -> None:
 
 class Variable():
 
+    _id:int
     _history: Optional[VarHistory]
     _gradient: Optional[object]
 
@@ -55,12 +56,11 @@ def _chainRule(self) -> None:
             arg._gradient += grad
         return
 
-    def backward(self, first=True):
-        if first:
-            self._oneGrad()
-        self._chainRule()
-        for parent in self._parent():
-            parent.backward(first=False)
+    def backward(self):
+        self._oneGrad()
+        TopoList = getTopoList(self)
+        for var in TopoList:
+            var._chainRule()
         return
 
     def new(self) -> "Variable":
@@ -82,4 +82,36 @@ def detach(self) -> "Variable":
         raise NotImplementedError
 
     def item(self) -> object:
-        raise NotImplementedError
+        raise NotImplementedError
+
+def _getTopoChain(var:"Variable") -> Iterable["Variable"]:
+    topoChainId = []
+    topoChainVar = []
+    for parent in var._parent():
+        topoChainId.append((var._id, parent._id))
+        topoChainVar.append(parent)
+        temp = _getTopoChain(parent)
+        topoChainId += temp[0]
+        topoChainVar += temp[1]
+    return topoChainId, topoChainVar
+
+def getTopoList(var:"Variable") -> Iterable["Variable"]:
+    topoChainId, topoChainVar = _getTopoChain(var)
+    topoId2Var = {var._id:var}
+    topoId2Degree = {var._id:1}
+    for (_, temp_id), temp_var in zip(topoChainId, topoChainVar):
+        topoId2Var[temp_id] = temp_var
+    topoChainId = list(set(topoChainId))
+    for _, parent_id in topoChainId:
+        if not parent_id in topoId2Degree:
+            topoId2Degree[parent_id] = 0
+        topoId2Degree[parent_id] += 1
+    topoList, queue = [], [var]
+    while len(queue) > 0:
+        variable = queue[0]
+        queue += variable._parent()
+        topoId2Degree[variable._id] -= 1
+        if topoId2Degree[variable._id] == 0:
+            topoList.append(variable)
+        del queue[0]
+    return topoList

test/autodiff/test_backward.py

@@ -152,15 +152,15 @@ def test_tanh(a:float=-2.0) -> None:
     return
 
 @pytest.mark.test_scalar_overload
-def test_add_backward() -> None:
+def test_complex_backward() -> None:
     A = Scalar(1.0, None, None)
     B = Scalar(2.0, None, None)
     C = A + B
     D = A + C
     E = (D - C) * (D - C)
     E.backward()
-    assert A._gradient == 5.0
-    assert B._gradient == 2.0
-    assert C._gradient == 1.0
+    assert A._gradient == 2.0
+    assert B._gradient == 0.0
+    assert C._gradient == 0.0
     assert D._gradient == 2.0
     return