0

I am trying to train VGGNET model on a custom dataset, I have experimented with the model on colab but now I need to run the same code on local machine.

I tried to run the code with Tensorflow GPU 1.15 and CUDA 10.0 (colab I am using has Tensorflow GPU 1.15 and CUDA 10.0), but the code is giving this error

...
  (0) Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
     [[{{node block1_conv1/convolution}}]]
     [[loss/mul/_169]]
...

Things I tried

  • Tried different version of cuDNN 7.6.4 and 7.4.2
  • Tried downgrading Tensorflow GPU to 1.14 from 1.15

I have also gone through these posts post1, post2 but I could not be able to resolve the error

System: Ubuntu 18.04, CUDA 10.0, cuDNN 7.4.2, Nvidia driver 440.64, Tensorflow GPU 1.15

talonmies
  • 70,661
  • 34
  • 192
  • 269
Atinesh
  • 1,790
  • 9
  • 36
  • 57

1 Answers1

0

Seems like there were some issues with RTX Cards, I had to add few lines of code to make it work

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
tf.Session(config=config)
Atinesh
  • 1,790
  • 9
  • 36
  • 57