Implementing realtime augmentation

[awhitesong]

Hi, i was going through your code for using realtime_augmentation.py,
I can see the training in all try* files. I am trying to understand how minibatch is augmented in training function. xs_shared is used for training but is initialized to zero? Saw the previous issue, you said the jpegs are loaded on the fly. So , trying to go through train_norm, but its pretty messed up. So not getting how ra is used. How should i be using functions in ra for realtime augmenting if i have a preloaded dataset trani_set_x like:

train_model = theano.function(inputs=[index], outputs=cost, updates=updates,
givens={
                x: train_set_x[index * batch_size:(index + 1) * batch_size],
                y: train_set_y[index * batch_size:(index + 1) * batch_size]
})

If not, i can load the image dataset on the fly, then where its been done in the code?

[benanne]

This is the part of the code that loads up the JPEG images and perturbs them: https://github.com/benanne/kaggle-galaxies/blob/master/realtime_augmentation.py#L185-L194

you can simply replace the load_data.load_img() call by something else that returns a numpy array and then you're all set.

[awhitesong]

1. Can i include my mini_batch instead of image_index in the function?

2. load_and_process_image is being called in LoadAndProcess class.
   So if i call LoadAndProcess(img_index, ds_transforms, augmentation_params, target_sizes), it will return me the augmented data. This is being returned in any try* file as:

   augmented_data_gen = ra.realtime_augmented_data_gen(num_chunks=NUM_CHUNKS, chunk_size=CHUNK_SIZE,
                                                   augmentation_params=augmentation_params, ds_transforms=ds_transforms,
                                                   target_sizes=input_sizes, processor_class=ra.LoadAndProcessPysexCenteringRescaling)
   

   But how is this augmented data in any try* file, taken as input to CNN in the training?

3. Also the load_and_process_image calls random_perturbation_transform through perturb_and_dscrop, which does transformations. And the augmenting functions are also implemented in load_data.py. So do i need to include load_data.py as well in my code?

I am just trying to understand what all files do i need to include in my code and how should i augment and include the data in my training loop. How should i use augmented_data_gen(pt.2) in my theano training function?

[benanne]

1. sure, just modify it to suit your needs.
2. I don't understand the question.
3. I suppose so.

I should clarify that I am unable to give extensive support for this code. It's provided as-is. If you want to use it or parts of it for anything else, you will have to figure things out on your own. Thanks for your understanding.

[awhitesong]

Actually i think i am understanding the codebase now. I figured it out somehow. Thanks a lot anyway for the help and a great code. :)