pymaxflow speed

[Yuzhu-Tao]

Hello, I have been using the pymaxflow library recently and found that it can only run the data on the CPU, which makes the running speed very slow. Is there any way to speed it up?

[pmneila]

Hi @Tylertyz

Unfortunately, parallelizing maximum flow algorithms is tricky, as they rely on sequential decisions and have rather irregular memory access patterns. There is some research on GPU maxflow algorithms, but they’re usually designed for specific types of graphs and aren’t always general-purpose or easy to integrate. Integrating parallel/GPU support into PyMaxflow would require a considerable effort and might introduce backward-incompatible changes, and I don't have much time to devote to that at the moment...

Just out of curiosity, what kind of graph are you working with? How large is it, how long does it take to run, and what kind of speedup are you hoping for?

[Yuzhu-Tao]

I'm trying to add it to the network for training, but it keeps going from GPU to CPU and back again, which is time-consuming. If pymaxflow doesn't work on the GPU, do you have any other suggestions to quickly implement the functionality of the code below? Or do you know of other CUDA libraries or other ways to do this?? I would appreciate it if you could help me with this problem.
g = maxflow. Graphfloat
nodes = g.add_grid_nodes((overlap_y_list[b], overlap_x_list[b]))
source,sink = (math.inf, 0), (0, math.inf)
g.add_grid_tedges(nodes[bomask1 == 1], *sink)
g.add_grid_edges(nodes, weights=(norm + np.roll(norm, -1, dims=1)), structure=np.tensor([[0, 0, 1]]) , symmetric=True)
flow = g.maxflow()
segments = g.get_grid_segments(nodes)

[Yuzhu-Tao]

If I want to implement graphcut, do you have any suggestions for using the GPU approach? I've been working on this piece lately, and I'd appreciate it if you could give me some advice.