Input images with more than 3 channels.

[NDKoehler]

Is there a possibility to concatenate mutliple 3 channel images to a set of 9 channel input images while still using dynamic image augmentation (and if possible recordios)?

gz.

[piiswrong]

I believe you can create 3 data sets and use concat symbol.

[NDKoehler]

I also thought about this possibility, but the model.fit() routine just takes on dataiter as input. I guess its not possible atm.

[antinucleon]

When we design the model we want to cover a general need but not long tail, so I think there are some possible solution.

1. if dataset is not large, use numpy iterator
2. use low level symbolic api
3. write a new iterator (in Python)

You may find this notebook of low level API is useful: https://github.com/dmlc/mxnet/blob/master/example/notebooks/simple_bind.ipynb if you want to do 1 or 2

On Thu, Oct 29, 2015 at 03:19 Liquidburner notifications@github.com wrote:

I also thought about this possibility, but the model.fit() routine just takes on dataiter as input. I guess its not possible atm.

— Reply to this email directly or view it on GitHub https://github.com/dmlc/mxnet/issues/401#issuecomment-152123852.

[pluskid]

I would like to advocate the Julia binding's current way of handling data and label here.

• the user name the output of data iter, and the name should match some input in the network symbol
• the model.py assumes all the provided name by the user's data iter is input (either data or label or both)

so there is no need to rely on the name having a _data postfix and supporting multiple input will be straightforward. The current feed forward model then should be able to handle pretty general case like the three input data stream here or auto encoder where data itself is just labels.

I could try to port this to the python side later if needed.

(Sorry was on my phone and accidentally closed the issue...)

[tqchen]

@pluskid Your proposal seems general enough to be used in cases such as multiple inputs and etc. This is definitely an interesting direction we should go

[pluskid]

OK, I'm currently doing some refactoring to simplify the Julia side IO interface. I will try to push changes for the the Python side maybe next week if nobody has come up with a better idea by then. But feel free to comment if there is any caveat about that choice that I might missed @tqchen @antinucleon

[NDKoehler]

Thanks for your answers. Multi image stream support for the python interface would be great, since many problems (especially medical ones) consist of many different views of the same object. The Julia side already supports multiple input images streams? @pluskid

[pluskid]

It supports slightly more general notion of inputs to a network. So for example multiple inputs could be feeder into a network. But since the underlying imageio is the same, I guess you will still have to create three streams of images each with 3 channels.

[tqchen]

now that both julia and python data interface should support multiple input data, please feel free to reopen this or open a new issue if there is any questions