_make_data_gen() and start_enqueueing_threads()

[villanuevab]

I'm not sure I understand what is happening in inputs/kitti_seg_input.py here: https://github.com/MarvinTeichmann/KittiSeg/blob/master/inputs/kitti_seg_input.py#L169-L173.

elif phase == 'train':
    yield jitter_input(hypes, image, gt_image)
    yield jitter_input(hypes, np.fliplr(image), np.fliplr(gt_image))

So, we are yielding 2 generators to the variable, gen, in L359:

gen = _make_data_gen(hypes, phase, data_dir)
gen.next() # but then why do we just iterate over it? 

But then why do we do the gen.next() in the following line, without making use of the returned value? What is the purpose of this portion of the code, in start_enqueueing_threads()?

[MarvinTeichmann]

Your assumption, that the second statement @L173 will never be reached is wrong.

The generator created by yield will jump back to the last yield keyword and execute the remainder of the function from there, pausing at the next yield command. You can test this behaviour using the minimal example below.

In [3]: def mygen():
   ...:     while True:
   ...:         yield 0
   ...:         yield 1
   ...:         
In [4]: gen = mygen()
In [5]: gen.next()
Out[5]: 0
In [6]: gen.next()
Out[6]: 1
In [7]: gen.next()
Out[7]: 0
In [8]: gen.next()
Out[8]: 1

[villanuevab]

Ah, yes. A quick review of yield statements and your response showed me that my starting assumption was wrong. Thank you!

[villanuevab]

Does this mean that the dataset is effectively doubled during training? In particular, that there are two versions of the input image in the training data:

1. the input image itself, and,
2. a horizontally flipped copy of the input image

[MarvinTeichmann]

Yes, data augmentation like random flipping is used to increasing the amount of effectively available training data.

Other forms of data augmentation applied is random aspect ratio augmentation (random resize, random crop) and color augmentation (random brightness, random contrast, random saturation and random hue).

[wenouyang]

Hi Marvin, related to random resize or random crop, does that mean you the training set will have different image size when feeding to the network input?

[MarvinTeichmann]

If you use batch_size = 1 (default) you can train the network with variable image sizes (default and recommended).

For batch_sizes > 1 all inputs need the same size. That can be achieved by setting reseize_image or crop_patch to true. Resize_image will perform bilinear resize after the augmentation (if you use both random resize and random crop this means that still meaningful augmentation is done) . crop_patch will crop a random patch of fixed size on the image.