Try out InceptionV4 on a dataset


#1

Hi everyone,

Is there anyone who already wrote a main for InceptionV4? As you know, it’s available on the repository but no tutorial or example shows any usage of that model. I really need to get InceptionV4 running so if there is anyone who could help me with it, I’d appreciate it. Here is where the model can be found in the repository.

Now what remains is a way to create the network and feed the images in. Although there are examples of that for resnet, they don’t work with InceptionV4 due to difference in the parameters.

Could someone perhaps share his main with InceptionV4 running?


#2

For example using the latest mxnet installed using pip install mxnet --pre

You could load the symbols into Gluon.

net = get_symbol()
net.save('inception-v4-symbol.json')
gluon_net = mx.gluon.nn.SymbolBlock.imports('inception-v4-symbol.json', ['data'])
gluon_net.initialize()
gluon_net(mx.nd.ones((1,3,299,299)))

And then use a normal training loop to load your data. Make sure they are big enough though, 299x299 seems to do the trick.

However it looks like you would need to be quite careful with how you initialize the parameters:
Have a look at this question to get an idea of how to do that granularly.