PR Category

User Experience

PR Types

Bug fixes

Description

1. 本 pr 是修复 【Hackathon 6th No.9】Add cartesian_prod API to Paddle -part #65605 (comment) 出现的问题：静态图下
   paddle.reshape 的输入如果是 zero-size tensor，结果返回的是 None。
   修改前：

   paddle.enable_static()
   startup_program = paddle.static.Program()
   main_program = paddle.static.Program()
   executor = paddle.static.Executor()
   data_shape = [2, 1, 0]
   data_np = np.ones(d_shape, 'float32')
   with paddle.static.program_guard(main_program, startup_program):
       x = paddle.static.data(
           name="x", shape=d_shape, dtype='float32'
       )
       out = paddle.reshape(data_shape, [1, 0])
       fetch_list = [out]
       feeds = {"x": data_np}
       executor.run(startup_program)
       res = executor.run(
           main_program, feed=feeds, fetch_list=fetch_list
       )
   print(res[0])
   # None

   修改后：

   paddle.enable_static()
   startup_program = paddle.static.Program()
   main_program = paddle.static.Program()
   executor = paddle.static.Executor()
   data_shape = [2, 1, 0]
   data_np = np.ones(d_shape, 'float32')
   with paddle.static.program_guard(main_program, startup_program):
       x = paddle.static.data(
           name="x", shape=d_shape, dtype='float32'
       )
       out = paddle.reshape(data_shape, [1, 0])
       fetch_list = [out]
       feeds = {"x": data_np}
       executor.run(startup_program)
       res = executor.run(
           main_program, feed=feeds, fetch_list=fetch_list
       )
   print(res[0])
   # []
   print(res[0].shape)
   # [1, 0]

2. 本 pr 把之前对 infermeta/unary.cc 中 ValidateShape 的修改迁移到了 fluid/operator/reshape_op.cc 的 ValidateShape 中，
   之前的修改是：【Hackathon 6th No.9】Update ValidateShape to support zero sized tensor flatten -part #64715 ，相比当时修改之前增加支持的 case 是：当输入是
   zero-size tensor，且 output_shape 中 0 出现的位置小于 in_dims.size()，例如：
   修改前：

   x = paddle.rand([2, 1, 0])
   x = paddle.reshape(x, [0])
   print(x)
   # ValueError: (InvalidArgument) The 'shape' in ReshapeOp is invalid. The input tensor X'size must be equal to the capacity 
   # of 'shape'. But received X's shape = [2, 1, 0], X's size = 0, 'shape' is [0], the capacity of 'shape' is 2.
   #  [Hint: Expected capacity == in_size, but received capacity:2 != in_size:0.] (at ..\paddle\phi\infermeta\unary.cc:1967)

   修改后：

   x = paddle.rand([2, 1, 0])
   x = paddle.reshape(x, [0])
   print(x)
   # Tensor(shape=[0], dtype=float32, place=Place(cpu), stop_gradient=True,
   #  [])