[SINGA-380] Fix bugs from Reshape

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Assignee: Unassigned
• Reporter: wangwei
• Created: 11/Jul/18 07:14
• Updated: 11/Jul/18 08:29

Description

The current implementation of reshape operation has memory leak caused by the dangling Tensor::block_ pointer.

This ticket is going to fix the bug and update the Reshape API in C++ and Python.

For C++,

Class Tensor {
 public:
  Tensor& Reshape(const Shape & s); // change the internal shape and return a the tensor itself
 ...  
};

Tensor Reshape(const Tensor& t, const Shape& s); // create a new tensor with the given shape; share the memory if possible.

For Python, the reshape operation from autograd.py always return a new pytensor which shares the memory with the original tensor if possible.

We change the API for transpose for C++ and Python in the same way.

C++, Tensor method Transpose changes the internal fields and return the tensor itself; the global method Transpose returns a new tensor which shares the memory with the original tensor.

Python, the transpose operation always returns a new tensor that shares memory with the original tensor.