Inception_v3是否使用dilations？适配TensorFlow 1.3推理兼容性咨询

First off, let me give you a clear, straightforward answer: the official Inception_v3 implementation bundled with TensorFlow 1.3 does NOT use dilated convolutions (the dilations/rate parameter). So you can safely run inference on your TF 1.3 server without any compatibility issues related to this feature.

Let me break this down with more context:

• The core Inception_v3 architecture, as shipped in TF 1.3, relies on standard convolutions, pooling layers, and the classic Inception modules (combining 1x1, 3x3, 5x5 convolutions) — dilated convolutions weren't part of its original design when this TF version was released.
• If you peek into the source code of slim.nets.inception_v3 in TF 1.3, you'll notice all convolution layers use regular settings with no dilation configured. For example, the first convolution layer looks like this:

  slim.conv2d(inputs, 64, [7, 7], stride=2, scope='conv1')
  

  There’s no rate parameter (TensorFlow’s early name for dilation control) specified anywhere in the official implementation.

A quick sanity check for your use case: if you’re using the pre-trained Inception_v3 weights provided by Google for TF 1.3, just load the checkpoint and run your inference pipeline as usual. Even if you built the model from scratch using TF 1.3’s slim API, as long as you didn’t manually add dilated convolution logic, everything will work smoothly.

内容的提问来源于stack exchange，提问作者Tujamo