Update tensor.py

tinygrad/tensor.py

@@ -91,7 +91,7 @@ def __init__(self, data, gpu=None):
       self.cuda_()
 
     # internal variables used for autograd graph construction
-    # Tensor列表
+    # _ctx是一个Function，当前环境上下文是在哪个函数内部
     self._ctx = None
 
   def __repr__(self):
@@ -147,15 +147,19 @@ def deepwalk(node):
 
     for t0 in reversed(nodes):
       assert (t0.grad is not None)
+      # __class__.__name__ 类名
+      # 在调试环境下进行反向传播（梯度计算）
       with ProfileOp(t0._ctx.__class__.__name__, [t0.grad], backward=True):
         grads = t0._ctx.backward(t0._ctx, t0.grad.data)
+      # grads放到列表中，以便下面应用zip方法
       if len(t0._ctx.parents) == 1:
         grads = [grads]
       for t,g in zip(t0._ctx.parents, grads):
         if g is None:
           continue
         assert g.shape == t.data.shape, \
           "grad shape must match tensor shape in %r, %r != %r" % (self._ctx, g.shape, t.data.shape)
+        # 在backward中作了链式的乘法运算，这里将不同parents对该变量的梯度（偏导数）再求和
         t.grad = Tensor(g) if t.grad is None else (t.grad + Tensor(g))
 
   # ***** tinygrad supports CPU and GPU *****
@@ -202,6 +206,7 @@ def detach(self):
   opsgpu = {}
 
   # ***** non first class ops *****
+  # ***** 一些稍复杂的算子 ****
 
   def mean(self):
     div = Tensor(np.array([1/np.prod(self.shape)], dtype=self.data.dtype), gpu=self.gpu)
@@ -229,7 +234,7 @@ def __init__(self, *tensors):
     self.parents = tensors
     self.saved_tensors = []
 
-  # 保存x到saved_tensors列表中
+  # 保存变量x（参数）到saved_tensors列表中
   def save_for_backward(self, *x):
     self.saved_tensors.extend(x)
 
@@ -245,6 +250,7 @@ def apply(self, *x, **kwargs):
     # overwrite with passed params
     for k, v in kwargs.items():
       setattr(ctx, k, v)
+    # 前向运算
     with ProfileOp(ctx.__class__.__name__, x):
       ret = Tensor(op.forward(ctx, *[t.data for t in x], **kwargs))
     ret._ctx = ctx