Flux: Hard to use train! and DataLoader for minibatched NamedTuple dataset with GPU

I want to write the following neural networks machine learning code on GPU but I need to write complicated and dirty something within dataloader’s for loop to pass data to train!:

# This GPU based program works but very slow than CPU

using Flux
using Flux.Data: DataLoader
using Printf

xpu = gpu
# xpu = cpu

n = 100
dataset = (
    input = (
        data1 = rand(1, n),
        data2 = rand(1, n),
        data3 = rand(1, n),
        data4 = rand(5, 10, n)
    output = rand(1, n)
) |> xpu
dataloader = DataLoader(dataset.input, dataset.output, batchsize = 4)
model = Chain(
    (input) -> cat(dims=1, input.data1, input.data2), # Here I ignore data3 and data4 for simplicity. In real model, I'll use these data.
    Dense(2, 10, relu),
    Dense(10, 1)
) |> xpu

loss(input, output) = Flux.mse(model(input), output)
optimizer = ADAM()
epoch_length = 100
for epoch in 1:epoch_length
    for (input, output) in dataloader
        # complicated and maybe slow, I need to reconstruct minibatched dataset
        data = [((data1 = input.data1[:, i], data2 = input.data2[:, i], data3 = input.data3[:, i], data4 = input.data4[:, :, i]), output[:, i]) for i in size(output)[2]]
        Flux.train!(loss, params(model), data, optimizer)
    loss(dataset.input, dataset.output) |> println

This code is very slower than CPU. How should I do to run it fast? What is the good practice?

If the batchsize is too small, it is hard to gain performance from GPU.

1 Like

The batch size is small and the complexity of the model is too small, and it seems there is another working GPU process by another user. Thanks!