search

dlsyscourse / hw1

8 stars 26 forks source link

not trace the graph for non-require-grad tensors under eager mode #12

Closed guihao-liang closed 1 year ago

guihao-liang commented 1 year ago

What this PR does is to use the make_from_op defined in Value.

I think the classmethod is a better implementation than the staticmethod since it takes care of requires_grad. If we are computing only non-require-grad tensors, they can be safely detached under the eager mode without tracing the graph.

demo

In [1]: import sys

In [2]: sys.path.append('./python/')

In [3]: import needle as ndl

In [4]: a = ndl.Tensor([1.0], requires_grad=False)

In [5]: b = ndl.Tensor([1.0], requires_grad=False)

In [6]: c = a * b

In [7]: c.is_leaf()
Out[7]: True

In [8]: c.inputs
Out[8]: []

In [9]: c
Out[9]: needle.Tensor([1.])