Add bincount op #36317
Add bincount op #36317
Conversation
|
Thanks for your contribution! |
python/paddle/tensor/linalg.py
Outdated
|
|
||
| Args: | ||
| x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor. | ||
| weights (Tensor, optional): Weight for each value in the input tensor. Should have the same shape as input. |
python/paddle/tensor/linalg.py
Outdated
| Args: | ||
| x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor. | ||
| weights (Tensor, optional): Weight for each value in the input tensor. Should have the same shape as input. | ||
| minlength (int): Minimum number of bins. Should be non-negative integer. |
There was a problem hiding this comment.
int -> int, optional
Default is 0.
| static_cast<double>(0)); | ||
| for (int64_t i = 0; i < input_numel; i++) { | ||
| output_data[input_data[i]] += static_cast<double>(weights_data[i]); | ||
| } |
There was a problem hiding this comment.
这个float和double的分支代码是否可以使用T合并，也和int类型的Weights更适配
There was a problem hiding this comment.
这边是为了保证与竞品一致。weights可以是任意数据类型，只有当weights float32时，output的类型才是float32。其他三种情况下output都是float64。
|
|
||
| KernelBincount<T, InputT, double><<<GET_BLOCKS(input_numel), | ||
| PADDLE_CUDA_NUM_THREADS, 0, stream>>>( | ||
| input_data, input_numel, has_weights, weights_data, output_data); |
There was a problem hiding this comment.
同CPU，float和double的分支代码是否可以使用T合并
|
|
||
| Args: | ||
| x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor. | ||
| weights (Tensor, optional): Weight for each value in the input tensor. Should have the same shape as input. Default is None. |
| @@ -44,6 +44,7 @@ | |||
| from .linalg import cholesky # noqa: F401 | |||
| from .linalg import bmm # noqa: F401 | |||
| from .linalg import histogram # noqa: F401 | |||
| from .linalg import bincount # noqa: F401 | |||
There was a problem hiding this comment.
we shall also add bincount in tensor_method_func list below to get paddle.Tensor.bincount
| PADDLE_ENFORCE_GE( | ||
| input_min, static_cast<InputT>(0), | ||
| platform::errors::InvalidArgument( | ||
| "The elements in input tensor must be non-negative ints")); |
There was a problem hiding this comment.
the check of PADDLE_ENFORCE* should in InferShape rather than in Compute, so we can detect illegal input earlier
| PADDLE_ENFORCE_EQ(input_type_match, true, | ||
| platform::errors::InvalidArgument( | ||
| "Input(X) holds the wrong type, it holds %s, but " | ||
| "desires to be %s or %s", | ||
| paddle::framework::DataTypeToString(input_type), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT32), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT64))); |
There was a problem hiding this comment.
the check of PADDLE_ENFORCE* should in InferShape rather than in Compute, so we can detect illegal input earlier
| PADDLE_ENFORCE_EQ( | ||
| platform::is_gpu_place(context.GetPlace()), true, | ||
| platform::errors::InvalidArgument("It must use CUDAPlace.")); |
There was a problem hiding this comment.
the check of PADDLE_ENFORCE* should in InferShape rather than in Compute, so we can detect illegal input earlier. but in this case, because register BincountCUDAKernel in cuda, no need this check.
| PADDLE_ENFORCE_GE( | ||
| *std::min_element(input_data, input_data + input_numel), | ||
| static_cast<InputT>(0), | ||
| platform::errors::InvalidArgument( | ||
| "The elements in input tensor must be non-negative ints")); |
There was a problem hiding this comment.
the check of PADDLE_ENFORCE* should in InferShape rather than in Compute, so we can detect illegal input earlier
paddle/fluid/operators/bincount_op.h
Outdated
| PADDLE_ENFORCE_EQ(input_type_match, true, | ||
| platform::errors::InvalidArgument( | ||
| "Input(X) holds the wrong type, it holds %s, but " | ||
| "desires to be %s or %s", | ||
| paddle::framework::DataTypeToString(input_type), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT32), | ||
| paddle::framework::DataTypeToString( | ||
| framework::proto::VarType::INT64))); |
There was a problem hiding this comment.
the check of PADDLE_ENFORCE* should in InferShape rather than in Compute, so we can detect illegal input earlier
python/paddle/tensor/linalg.py
Outdated
| if paddle.max(x) < 0: | ||
| raise ValueError("Elements in Input(x) should all be non-negative") |
There was a problem hiding this comment.
x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor.
shall we must check:
- paddle.min(x) >= 0:
- x.ndim == 1;
- x.numel() != 0
| Args: | ||
| x (Tensor): A Tensor with non-negative integer. Should be 1-D tensor. | ||
| weights (Tensor, optional): Weight for each value in the input tensor. Should have the same shape as input. Default is None. | ||
| minlength (int, optional): Minimum number of bins. Should be non-negative integer. Default is 0. |
smallv0221
dismissed stale reviews from XiaoguangHu01 and jeff41404
via
647aa0d
* Add bincount op * upload cpu version * fix unitest * fix unittest * fix unittest * fix en doc * add more test * fix en doc * add more test case * fix test * fix input vailidation * fix input check * fix unittest * fix test * fix en doc cherry-pick
PR types
New features
PR changes
OPs
Describe
Add bincount op
中文文档pr：PaddlePaddle/docs#3959
英文文档：