Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2302
Star 20931
Fork 5422
P
Paddle

refine reshape_op shape error message !22480

!22480 已合并 由 saxon_zh@saxon_zh 创建

Created by: luotao1

reshape op增强维度类报错信息

input = np.random.random([2, 25]).astype("float32")
x = fluid.layers.data(name="x", shape=[2, 25], append_batch_size=False, dtype="float32")
shape = fluid.layers.data(name="shape", shape=[1, 4], append_batch_size=False, dtype="float32")
fluid.layers.reshape(x, shape=shape)

cpu = fluid.core.CPUPlace()
exe = fluid.Executor(cpu)

1.6版本修复PR：#20099 测试脚本：test.py.zip

1.shape参数中元素不止一个为-1时

exe.run(fluid.default_main_program(), 
           feed={"x": input, 
                      "shape": np.array([-1, -1, 5, 1]).astype("int32")})
  • 1.6修改前报错：报错内容表述不清楚，没有打印维度信息
Only one input dimension of Attr(shape) can be unknown.
  • 1.6修改后报错（如1.6版本没有修改，不需要写）
ShapeError: Only one dimension value of 'shape' in ReshapeOp can be -1. But received shape = [-1, -1, 5, 1], shape[1] is also -1.
  • 本PR修改后报错
InvalidArgumentError: Only one dimension value of 'shape' in ReshapeOp can be -1. But received shape = [-1, -1, 5, 1], shape[1] is also -1.
  [Hint: Expected unk_dim_idx == -1, but received unk_dim_idx:0 != -1:-1.] at (/home/luotao/Paddle/paddle/fluid/operators/reshape_op.cc:149)
  [operator < reshape2 > error]

2.shape参数中元素为0时，0所对应的维度出错

exe.run(fluid.default_main_program(), 
           feed={"x": input, 
                      "shape": np.array([2, 5, 5, 0]).astype("int32")})
  • 1.6 修改前报错：报错内容表述不清楚，没有打印维度信息
The index of dimension to copy from input shape must be less than the size of input shape.
  • 1.6修改后报错（如1.6版本没有修改，不需要写）
ShapeError: The index of 0 in `shape` must be less than the input tensor X's dimensions. But received shape = [2, 5, 5, 0], shape[3] = 0, X's shape = [2, 25], X's dimensions = 2.
  • 本PR修改后报错
InvalidArgumentError: The index of 0 in `shape` must be less than the input tensor X's dimensions. But received shape = [2, 5, 5, 0], shape[3] = 0, X's shape = [2, 25], X's dimensions = 2.
  [Hint: Expected static_cast<int>(i) < in_dims.size(), but received static_cast<int>(i):3 >= in_dims.size():2.] at (/home/luotao/Paddle/paddle/fluid/operators/reshape_op.cc:159)
  [operator < reshape2 > error]

3.shape参数有-1以外的负值

exe.run(fluid.default_main_program(), 
           feed={"x": input, 
                      "shape": np.array([-1, -5, 5, 1]).astype("int32")})
  • 1.6 修改前报错：
Each input dimension of Attr(shape) must not be negtive except one unknown dimension.
  • 1.6修改后报错（如1.6版本没有修改，不需要写）
ShapeError: Each dimension value of 'shape' in ReshapeOp must not be negtive except one unknown dimension. But received  shape = [-1, -5, 5, 1], shape[1] = -5.
  • 本PR修改后报错
InvalidArgumentError: Each dimension value of 'shape' in ReshapeOp must not be negtive except one unknown dimension. But received  shape = [-1, -5, 5, 1], shape[1] = -5.
  [Hint: Expected shape[i] > 0, but received shape[i]:-5 <= 0:0.] at (/home/luotao/Paddle/paddle/fluid/operators/reshape_op.cc:167)
  [operator < reshape2 > error]

4.传入无效的shape值，无法正确推断出-1对应的实际值

exe.run(fluid.default_main_program(), 
           feed={"x": input, 
                      "shape": np.array([-1, 3, 5, 1]).astype("int32")})
  • 1.6修改前报错：
Invalid shape is given.
  • 1.6修改后报错（如1.6版本没有修改，不需要写）
ShapeError: The 'shape' in ReshapeOp is invalid. The input tensor X'size must be divisible by known capacity of 'shape'. But received X's shape = [2, 25], X's size = 50, 'shape' is [-1, 3, 5, 1], known capacity of 'shape' is -15.
  • 本PR修改后报错
InvalidArgumentError: The 'shape' attribute in ReshapeOp is invalid. The input tensor X'size must be divisible by known capacity of 'shape'. But received X's shape = [2, 25], X's size = 50, 'shape' is [-1, 3, 5, 1], known capacity of 'shape' is -15.
  [Hint: Expected output_shape[unk_dim_idx] * capacity == -in_size, but received output_shape[unk_dim_idx] * capacity:-45 != -in_size:-50.] at (/home/luotao/Paddle/paddle/fluid/operators/reshape_op.cc:190)
  [operator < reshape2 > error]

5.传入无效的shape值，输入x的size与 shape不匹配

exe.run(fluid.default_main_program(), 
           feed={"x": input, 
                      "shape": np.array([1, 3, 5, 1]).astype("int32")})
  • 1.6修改前报错：
Invalid shape is given.
  • 1.6修改后报错（如1.6版本没有修改，不需要写）
ShapeError: 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, 25], X's size = 50, 'shape' is [1, 3, 5, 1], the capacity of 'shape' is 15.
  • 本PR修改后报错
InvalidArgumentError: 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, 25], X's size = 50, 'shape' is [1, 3, 5, 1], the capacity of 'shape' is 15.
  [Hint: Expected capacity == in_size, but received capacity:15 != in_size:50.] at (/home/luotao/Paddle/paddle/fluid/operators/reshape_op.cc:204)
  [operator < reshape2 > error]