search

weizhou-geek / VGCN-PyTorch

PyTorch Implementation of TCSVT 2020 "Blind Omnidirectional Image Quality Assessment with Viewport Oriented Graph Convolutional Networks"
MIT License
22 stars 9 forks source link

About GCN layer #3

Open hannaljv opened 3 years ago

hannaljv commented 3 years ago

Dear sir, Thank you for your wonderful work. I misunderstanding about how to calculate the output of the GCN layer, for example, "GraphConvolution(512, 256)". (In case I want to change the other input size). Can you explain it? Thank you.

weizhou-geek commented 3 years ago

@hannaljv For the GCN layer, the input feature size is (batch size, node, input channel) and the output size is (batch size, node, output channel). GraphConvolution(input channel, output channel) only changes the feature channel dimension, the batch size and the number of nodes remain unchanged.

hannaljv commented 3 years ago

@weizhou-geek thank you so much. I wonder that if I have nxn nodes (can be considered as a gray image (nxn)), n,n Affinity matrix, and each node is represented by 1 feature (a number), so can I feed it in GCN layer

weizhou-geek commented 3 years ago

@weizhou-geek thank you so much. I wonder that if I have nxn nodes (can be considered as a gray image (nxn)), n,n Affinity matrix, and each node is represented by 1 feature (a number), so can I feed it in GCN layer

Yes, you can feed it into the GCN layer.

hannaljv commented 3 years ago

I try to put it in form of (batch, node = nxn, feature = 1) (1) and (batch, node = n, feature = n) (2), but It seen doesn't work.

weizhou-geek commented 3 years ago

I try to put it in form of (batch, node = nxn, feature = 1) (1) and (batch, node = n, feature = n) (2), but It seen doesn't work.

You may first refer to the adjacent matrix whose size should be (n \times n, n \times n).