some questions about the irregular mask？

[knazeri]

some questions about the irregular mask？

[knazeri]

@ljjcoder Partial convolution is doing the same thing, only in their paper, they declare the mask M (0 for holes), which is 1-M in our code!

Judging by your results, it looks like you have downloaded their Training Mask Dataset and that's why your input is mostly white! I'm not sure if that's the dataset they used to train their model because it obviously is different than what we see in their paper. We used their Testing Irregular Mask Dataset for our training which contains 12,000 irregular masks, and by augmentation we increased the size to 60,000!

[dulibubai]

Hello！Can you help me with the augmented train masks dataset which originally made by Guilin Liu ?

[knazeri]

@dulibubai You can try rotating each image by 4 multiples of 90 degrees, assuming img is an ndaaray:

from scipy.ndimage import rotate
rotate(img, 90)

You can also reflect the image about its y axis:

img[:, ::-1, ...]

If you conbine both methods, it gives you 4 x 2 size of the original dataset.

[dulibubai]

But the original masks dataset is not the binary images,which don't like the form of test masks dataset?
How do you address it?

[knazeri]

@dulibubai Like I said, we used their Testing Irregular Mask Dataset, and they are binary mask images. You can easily augment the dataset to be almost 96,000, which is a good figure!
You can also download Quick Draw Irregular Mask Dataset by Karim Iskakov which contains lots of binary strokes drawn by human hand.

[dulibubai]

I See,Thanks again!

[ljjcoder]

@knazeri Thanks for your reply！I also some question about training。there are three stage for trainning the model，1) training the edge model, 2) training the inpaint model and 3) training the joint model。you say the celebA need about 40 epoch to train。Did the 40 epoch is need for every stage or only for 3 stage and how many epoch does it need for every stage？

[knazeri]

@jtetc Technically you need to train each one separately from scratch! However, if you want to speed up the training, you can start by pre-training your model with a smaller input size (say 128). This especially helps with edge model (mode=1)