Skip to content

Add option to normalize gradient while training #84

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add option to normalize gradient while training #84

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 18 additions & 6 deletions experiments/dmrg.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,21 +1,26 @@
using TensorTrains, TensorTrains.MatrixProductStates
using Random
using Plots

rng = Xoshiro(1)
d_original = 3
L = 5
q = rand_mps(ComplexF64, d_original, L, 4)
F = Float64
q = rand_mps(rng, F, d_original, L, 4)
normalize!(q)
nsamples = 5*10^3
X = [sample(q)[1] for _ in 1:nsamples]
X = [sample(rng, q)[1] for _ in 1:nsamples]
nll = -loglikelihood(q, X)
println("Negative Log-Likelihood according to generating distribution q = $nll\n")
mq = marginals(q)
pmq = twovar_marginals(q)

p = rand_mps(ComplexF64, 2, length(q), 4)
p = rand_mps(rng, F, 2, length(q), 4)

function CB()
nlls = zeros(0)
diff_marg = zeros(0)
diff_pair_marg = zeros(0)
dots = zeros(0)
function cb(sweep, k, p, nll)
p_cp = deepcopy(p)
Expand All @@ -24,26 +29,33 @@ function CB()
mbd = maximum(bond_dims(p.ψ))
mp = marginals(p)
d_m = maximum(maximum.(abs, mp-mq))
pmp = twovar_marginals(p)
d_pm = maximum(maximum.(abs, pmp-pmq))
println("# Sweep $sweep, site k=$k")
println("Negative LogLikelihood = $nll.\ndmax=$mbd")
println("|<p|q>|=$d")
println("Max diff marginals = $d_m\n")
println("Max diff marginals = $d_m")
println("Max diff pair marginals = $d_pm\n")
push!(nlls, -loglikelihood(p, X))
push!(diff_marg, d_m)
push!(diff_pair_marg, d_pm)
push!(dots, d)
end
end

callback = CB()
nsweeps = 40
normalize_gradient = true
two_site_dmrg!(p, X, nsweeps;
η=1e-4, ndesc=10, svd_trunc=TruncBond(d_original+2), callback)
normalize_gradient,
η=1e-4, ndesc=10, svd_trunc=TruncBond(d_original), callback)

pl1 = plot(callback.nlls, xlabel="it", ylabel="NLL", label="")
hline!(pl1, [nll], ls=:dash, c=:gray, label="NLL according to generative model")
# vline!(pl1, (N-1):(N-1):length(callback.nlls), label="Ends of sweeps", ls=:dash, c=:gray)
pl2 = plot(callback.diff_marg, xlabel="it", ylabel="Max error on marginals", label="")
# vline!(pl2, (N-1):(N-1):length(callback.nlls), label="Ends of sweeps", ls=:dash, c=:gray)
pl2b = plot(callback.diff_pair_marg, xlabel="it", ylabel="Max error on pair marginals", label="")
pl3 = plot(callback.dots, xlabel="it", ylabel="<p|q>", label="")
# vline!(pl3, (N-1):(N-1):length(callback.nlls), label="Ends of sweeps", ls=:dash, c=:gray)
plot(pl1, pl2, pl3, layout=(3,1), size=(500,800), margin=5Plots.mm)
plot(pl1, pl2, pl2b, pl3, layout=(4,1), size=(500,900), margin=5Plots.mm)
4 changes: 4 additions & 0 deletions src/dmrg.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,7 @@ function _two_site_dmrg_sweep_generic!(
η = 1e-3, # learning rate for gradient descent
ndesc = 100, # number of gradient descent steps
optimizer = Optim.Adam(; alpha=η),
normalize_gradient = false, # normalize gradient to go quicker on flat regions
callback = (sweep, k, p, loss_val) -> nothing,
prodA_left = [precompute_left_environments(p, x) for x in data[1]],
prodA_right = [precompute_right_environments(p, x) for x in data[1]],
Expand All @@ -87,6 +88,9 @@ function _two_site_dmrg_sweep_generic!(
)
if G !== nothing
G .= reshape(grad, length(G))
if normalize_gradient
G ./= norm(G)
end
end
if F !== nothing
return val
Expand Down
Loading