Bindings for custom training dataset generation

[cacophrene]

Hi,

I would like to generate my own image dataset for training of a vgg19 CNN. I may have missed it, but I could not find dataset, iterator and batch functions that were introduced in r1.4 (I am using the git version of tensorflow-ocaml together with tensorflow library r1.5). Is there a binding to these functions or should I use another method to build my own training dataset?

Many thanks for your help.

[LaurentMazare]

Indeed dataset ops are not handled currently: the ocaml bindings are automatically generated but don't work properly on some operation types and this is the case for the dataset functions. I've been using placeholders to be able to feed any data to my models. The advantage is that the dataset processing is done on the ocaml side rather than in the tensorflow graph.
That being said I understand that dataset primitives may be interesting for some use cases, would you have an idea which primitives would be helpful to you ?

[cacophrene]

Thanks for your fast reply.
So I guess I should have a closer look at your mnist_conv.ml example for placeholders. At the moment I have a list of tensors corresponding to normalized image data (indeed it is convenient to load them with camlimages). I would like to use them as xs and ys values for Model.fit. What I have not figured out yet is how to achieve a packing function that would convert this list into a suitable tensor. I am still in the process of browsing the binding though.

[cacophrene]

Hi,
I think I can give you a more precise answer now. What I would like to achieve is a so-called minibatch that I could use to feed Model.fit (that is, getting something similar to the output of Mnist_helper.read_files, for instance). I can probably do that by using placeholders and Session.run, but as far as I know there is currently no feed_dic option available. Could you help with that? Thanks already for the time you spent on my issue.

[LaurentMazare]

Hi,
Actually if you're using Model.fit from the Fnn api, this would handle mini-batching for you. You can see an example here, xs and ys are your input tensors which get split according to batch_size before being fed to the model.

If you want to use the TensorFlow api directly rather than through Fnn (which will give you some greater control over what your model can do), you should indeed give placeholders to Session.run via Session.Input (this is similar to using feed_dic on the python side). An example of doing mini-batches with this can be found here.

Hope this helps, and don't hesitate if you have further questions.

[cacophrene]

Thanks a lot for the detailed answer, I will give a try.