中文 (Chinese)

• 版本、环境信息

Paddle With CUDA: True

OS: debian stretch/sid
Python version: 3.7.4

CUDA version: 9.2.148
cuDNN version: 7.3.1
Nvidia driver version: 418.67

• 复现信息：

动态图版本

import paddle
import paddle.fluid as fluid
from paddle.fluid import layers, dygraph as dg

place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
fluid.enable_dygraph(place)

inp = layers.randn([5,4,128,128])
slc = inp > 0

# inp[layers.where(slc)] = 0

inp[slc] = 0

print(inp[layers.where(slc)])

静态图版本

import paddle
import paddle.fluid as fluid
from paddle.fluid import layers
import numpy as np

exe = fluid.Executor(fluid.CUDAPlace(0))

x = fluid.data(name='x', shape=[5,4,128,128], dtype='float32')
slc = x > 0
x[slc] = 0
y = x * 1.0

prog = fluid.default_main_program()
each_ret, = exe.run(prog, feed={'x':np.random.rand(5,4,128,128).astype(np.float32)}, fetch_list=[y])
print(each_ret)

• 建议描述：希望能实现如上的操作，正如numpy所能做到的那样，这样比起index_select会方便很多

英文 (English)

Version/Environment Infomation

Paddle With CUDA: True

OS: debian stretch/sid
Python version: 3.7.4

CUDA version: 9.2.148
cuDNN version: 7.3.1
Nvidia driver version: 418.67

To Reproduce

Dynamic Graph Version

import paddle
import paddle.fluid as fluid
from paddle.fluid import layers, dygraph as dg

place = fluid.CUDAPlace(fluid.dygraph.ParallelEnv().dev_id)
fluid.enable_dygraph(place)

inp = layers.randn([5,4,128,128])
slc = inp > 0

# inp[layers.where(slc)] = 0

inp[slc] = 0

print(inp[layers.where(slc)])

Static Graph Version

import paddle
import paddle.fluid as fluid
from paddle.fluid import layers
import numpy as np

exe = fluid.Executor(fluid.CUDAPlace(0))

x = fluid.data(name='x', shape=[5,4,128,128], dtype='float32')
slc = x > 0
x[slc] = 0
y = x * 1.0

prog = fluid.default_main_program()
each_ret, = exe.run(prog, feed={'x':np.random.rand(5,4,128,128).astype(np.float32)}, fetch_list=[y])
print(each_ret)

Describe the feature and the current behavior/state.

Hope to achieve the above operation, as numpy can do, this will be much more convenient than index_select