Could you add some information about data augmentation in the readme?

[andreruas]

What are your command line arguments?: python3 main.py --num_epochs 200 --batch_size 1 --h_flip true --brightness 10 --rotation 30 --model DeepLabV3-Res152 Have you written any custom code?: No TensorFlow version?: 1.8

Describe the problem

Hello, I've been using your library for a week or so now and it has been excellent overall. Thanks for putting together such a comprehensive and accessible package for deep learning.

I had a question about data augmentation though, and wasn't able to find an answer when looking through the readme or the code. I'm not completely sure what the units/limits are for the brightness augmentation (is it a percentage from 0 to 100, or 0 to 255)? Also, I have the same question with the rotation flag. Is it in degrees?

What have you done to try and solve this issue?: I'm asking this question because I've trained DeepLabv3-Res152 on the CamVid dataset a few times using your library with various settings (successfully), but last night I set the --brightness 10 and --rotation 30 flags for the first time. Unfortunately, the loss grew extremely high after a few epochs, and the model never converged. I figured the brightness would change by 10% and the rotation by 30 degrees: am I wrong about that?

Could you add a sentence or two explaining this in the readme? Again, thanks so much for putting together such an excellent package.

Source code / logs

[2018-06-26 08:44:25] Epoch = 499 Count = 400 Current_Loss = 10795887616.0000 Time = 2.12 [2018-06-26 08:44:27] Epoch = 499 Count = 420 Current_Loss = 22598909952.0000 Time = 2.13

[GeorgeSeif]

Hi there @andreruas, thanks for the feedback and thorough description of the issue!

Ah I see that it can be a bit confusing. So for the brightness, it's simply an exponential factor. You can see in this line:

table = np.array([((i / 255.0) ** factor) * 255 for i in np.arange(0, 256)]).astype(np.uint8)

that "factor" is applied as an exponent. I admit it is a bit confusing and it's fairly old code. I can probably change it to something a bit clearer and easier to use! I would honestly avoid using it for now as it should really be changed. I do like your idea/thought of having be a percentage though.

The rotation is simply in degrees :).

I'll finally be getting a new computer soon and be able to make more updates to the repo!

Let me know if you have any more questions

[GeorgeSeif]

Ah I actually decided to do a quick update. So I changed the brightness to be like you said as a percentage, but it's between 0 and 1 (so 10% max change is 0.1). I documented it in the command line args

[andreruas]

Thanks for the quick update, I appreciate it. After updating my code, however, I think you may have made a small mistake: using self.brightness instead of args.brightness. This is the error I got after running the main.py (the same command as above).

Traceback (most recent call last): File "main.py", line 288, in input_image, output_image = data_augmentation(input_image, output_image) File "main.py", line 109, in data_augmentation factor = 1.0 + random.uniform(-1.0*self.brightness, self.brightness) NameError: name 'self' is not defined

[GeorgeSeif]

Are you're right! Sorry, I don't have a GPU right now so I haven't been able to test anything. I'll be getting one soon though!. I'll change that one