search

jamesmf / mnistCRNN

Simple TimeDistributed() wrapper Demo in Keras; sums images of MNIST digits
61 stars 24 forks source link

Change conv layer for MobileNet model #10

Closed sun9700 closed 6 years ago

sun9700 commented 6 years ago

We try to replace a simple convolution layer with a complete pretrained model (MobileNet model) and get a lower RMS.

jamesmf commented 6 years ago

This is a great addition! Here's what I'm thinking:

It's still useful to see how to add use it to wrap a single layer, so it might be worth keeping the RNN version, and explaining that the mobilenet example is essentially a (X)RNN example, where (X) is any model that honors the right intput:output mapping.

So I'd like to have two files: addMNIST_layer.py and addMNIST_model.py and explain the use case in the readme. I think that maximizes the utility of this example.

If there are two files, I'll probably do a bit of refactoring such that the data processing functions are properly shared between them.

Also if you decide to edit this PR, can you remove the .idea config xml file?