3D卷积张量流float64

Question

I'm working on some convolutional calculation on the Tensorflow and got some strange errors with the Double Precision numbers.

I needed to calculate those tensors in float64 because of the precision issues.

The conv3d works fine with me on float32, but on the Double precision, it's only works on CPU and for the GPU the error is saying

InvalidArgumentError (see above for traceback): Cannot assign a device for operation 'Conv3D': Could not satisfy explicit device specification '/device:GPU:0' because no supported kernel for GPU devices is available.
Registered kernels:
  device='CPU'; T in [DT_HALF]
  device='CPU'; T in [DT_FLOAT]
  device='CPU'; T in [DT_DOUBLE]
  device='GPU'; T in [DT_HALF]
  device='GPU'; T in [DT_FLOAT]

 [[Node: Conv3D = Conv3D[T=DT_DOUBLE, data_format="NDHWC", dilations=[1, 1, 1, 1, 1], padding="SAME", strides=[1, 1, 1, 1, 1], _device="/device:GPU:0"](Const, Const_1)]]

Below is my test code.

import tensorflow as tf
import numpy as np

with tf.device('/device:GPU:0'):

    a = tf.constant(2 * np.ones([1, 200, 200, 200, 9], dtype=np.float64), dtype=tf.float64)

    b = tf.constant(np.ones([3, 3, 3, 9, 9], dtype=np.float64), dtype=tf.float64)


    sess = tf.Session()

    init = tf.global_variables_initializer()

    sess.run(init)


    c = tf.nn.conv3d(a, b, strides=[1, 1, 1, 1, 1], padding='SAME')
    print(a)
    print(b)
    print(c)
    for i in range(1000):
        print(sess.run(c))

is there any way to use gpus for these ops?

Answer 1

Support for double precision on a GPU is hardware dependent. A quick search for your make and model and "support for double precision" should get you an answer whether your GPU can support double precision math.

A quick search for my GPU (gtx 1060) and its support for double precision resulted in this link: https://devtalk.nvidia.com/default/topic/995849/does-the-gtx1060-support-double-precision-/

For the gtx 1060, the throughput for double precision math is 1/32 that of single precision math. So even if your GPU supports it, the performance many not match your expectations.