In the previous section, we've seen how to use the CTF format with MinibatchSource to feed data to the CNTK trainer. But most datasets don't come in this format. So, you can't really use this format unless you create your own dataset or convert the original dataset to the CTF format.
CNTK currently supports a limited set of deserializers for images, text, and speech. You can't extend the deserializers at the moment, which limits what you can do with the standard MinibatchSource. You can create your own UserMinibatchSource, but this is a complicated process. So, instead of showing you how to build a custom MinibatchSource, let's look at how to feed data into the CNTK trainer manually.
Let's first recreate the model we used to classify Iris flowers:
from cntk import default_options, input_variable
from cntk.layers...
