Benchmark 1d convnet keras vs flux

flux_vs_keras.jl

using PyCall
using Flux

function doit_keras(cfg)
    keras = pyimport("tensorflow.keras")
    inp = keras.layers.Input((nothing, 1))
    x = inp
    x = keras.layers.Conv1D(kernel_size=51, filters=50)(x)
    x = keras.layers.Conv1D(kernel_size=1, filters=1)(x)
    out = x
    m = keras.Model(inputs=inp, outputs=out)
    m.compile(optimizer="adam", loss="mse")

    x = randn(Float32, cfg.nexamples, cfg.nw, cfg.nc)
    y = randn(Float32, cfg.nexamples, cfg.nw-50, 1)
    for _ in 1:cfg.nepochs
        @time m.fit(x,y, batch_size=cfg.batch_size)
    end
end

function doit_flux(cfg)
    
    dims_x = (cfg.nw   , cfg.nc, cfg.nexamples)
    dims_y = (cfg.nw-50, cfg.nc, cfg.nexamples)
    x = randn(Float32, dims_x)
    y = randn(Float32, dims_y)
    data = Flux.Data.DataLoader((x,y), batchsize=cfg.batch_size)
    opt = ADAM()
    net = Chain(
        Conv((51,), 1 => 50, Flux.relu),
        Conv((1,), 50 => 1, Flux.relu),
    )
    the_loss = let net=net
        function _loss(x,y)
            y_pred = net(x)
            @assert size(y) == size(y_pred)
            Flux.mse(y_pred, y)
        end
    end
    for _ in 1:cfg.nepochs
        @time Flux.train!(the_loss, params(net), data, opt)
    end
end

cfg = (nexamples=1000, nw=10000, nc=1, nepochs=3, batch_size=10)
doit_keras(cfg)
doit_flux(cfg)