CODEWITHSUNDEEP
tensorflowconv-neural-network
M.SALHI
M.SALHI

Reputation: 74

depthwise_Conv2D_native Tensorflow

What is the difference between the depthwise_conv2d and depthwise_conv2d_native in TensorFlow ?

Upvotes: 1

Views: 626

Answers (1)

Alex Meiburg
Alex Meiburg

Reputation: 656

As noted in the following unit test, https://github.com/tensorflow/tensorflow/blob/b0ecc7d2c1486367ec65d297e372f8935ee3ddfe/tensorflow/python/kernel_tests/depthwise_conv_op_test.py#L87

.. they are indeed supposed to be identical (when no stride is present). depthwise_conv2d_native is the simpler implementation that doesn't have strides, and goes straight to a C++/CUDA operation. depthwise_conv2d takes an optional stride argument; it passes this into with_space_as_batch to do the stride, and then that calls depthwise_conv2d_native as the backend.

So, bottom line: there isn't a different in what they compute. In terms of performance, depthwise_conv2d could add marrrrrrginally more complexity to your model, but adds the flexibility that you can have variable strides.

Upvotes: 2

Related Questions