How to prepare for my own data? about graph_attention_pool HOT 5 CLOSED

Hi, it is for graph classification = image classification. But in principle it can also be used for node classification, which will make it equivalent to the semantic segmentation problem.

from graph_attention_pool.

To get the mn and sd values I used this function to go over the entire dataset and collect global statistics:
https://github.com/bknyaz/graph_attention_pool/blob/master/utils.py#L188. Note that I'm using sd[sd < 1e-2] = 1 to avoid dividing by small numbers and blowing up node features.
Then during training and testing I use this mn and sd as here: https://github.com/bknyaz/graph_attention_pool/blob/master/train_test.py#L113

So there is no need to calculate statistics for each image. The steps are:

1. compute global statistics mn, sd (this is done once before training)
2. during training/testing for each node (superpixel) with features x (sp_coord, sp_intensity) you compute (x-mn)/sd

But you can try computing statistics mn, sd for each image instead of using global ones. It can result in better performance in some cases. I haven't tried this in this project.

from graph_attention_pool.

Sorry I didn't notice this issue. Are you still having this problem?
To create mnist_75sp_train I use the script extract_superpixels.py: python extract_superpixels.py -s train -t 4.
Let me know if you need futher help.

from graph_attention_pool.

@bknyaz
Hi, I wonder that super-pixels dataset preparation is founded for node classification or graph classification ?

from graph_attention_pool.

@bknyaz
Thanks. I am preparing my own data with superpixel scheme and following the repo homepage instruction.
I notice these lines:
# mean and std computed for superpixel features in the training set
mn = torch.tensor([0.11225057, 0.11225057, 0.11225057, 0.44206527, 0.43950436]).view(1, 1, -1)
sd = torch.tensor([0.2721889, 0.2721889, 0.2721889, 0.2987583, 0.30080357]).view(1, 1, -1)

How to calculate mean and std for node features? Which step should I start to work these out?
While the sp_intensity and sp_coord in single image has been calculated,
for the whole dataset (B, C, W, H) I calculated mean and std via np.mean/np.std(data, axis=(0, 2, 3)) thus I got:
[0.47861066 0.49431583 0.43597046] / [0.21912035 0.22233568 0.2661024]
Can I adopt these results for sp_coord and sp_intensity directly or should I re-calculate for each image superpixel features?

from graph_attention_pool.