Reproduced results of ViG for different graph types is confusing.

[tdzdog]

I have investigated the influence of graph types for ViG. However, my experiment results are confusing and different from the data in paper. As shown in Table 6, EdgeConv has highest FLOPS and accuracy. However, according to my results, the results of MaxRelative, GraphSage and EdgeConv are 74.42, 74.46 and 74.24 (ViG-ti). It shows EdgeConv has higher computation but lower accuracy. I also tried PrymaidViG-ti and ViG-s, all results show the high-computation graph (EdgeConv or GraphSage) is worse than the low-computation graph (MaxRelative). This is confusing and I am wondering why. This makes me doubt if the graph architecture necessary enough? Can you release the pretrained models for other graph structures than MaxRelative?

[iamhankai]

In my experience, MaxRelative is also the best option. If you can propose better alternative than MaxRelative, it would be exciting!