Tensorflow CIFAR10 代码分析

Question

I'm trying to figure out, a Tensorflow CIFAR10 tutorial and currently I can't understand line #245 , namely why is the shape for weight [dim, 384]? Is 384 a hyperparameter or is it somehow calculated?

Answer 1

Basically it was an arbitrary choice that worked out with their batch size and knowledge about the dataset.

So cifar images are 32 * 32 * 3 and by the convolution now they have 32 * 32 * 64 features and just before that they had 64 filters but they just max pooled it so now it's half the size so now its 16 * 16 * 64. They resized the images to the batch size = 128 so now its 128 * 128. Then they use weights to bring it up to 384.

Feel free to use another number but make sure you change the next layers as well. It's just an example CNN.