Reputation: 111
Tensorflow and cudnn convolution have different padding. But CPU/GPU convolutions are similar for user. How tensorflow uses cudnn to make "tensoflow style" convolution?
Upvotes: 2
Views: 496
Reputation: 5206
I suggest you look at the implementation of tf.nn.convolution (in tensorflow/python/ops/nn_ops.py) to see how it's done from python, and the implementation of the conv2d kernel in tensorflow/core/kernels/conv_ops.cc for the C++ bit (which calls into cudnn).
Upvotes: 1