Add Sherlock scripts (#62)

* Add sherlock scripts

* Allow setting hyperparams for synthetic comparison via command-line args.

* Add birdsong run scripts for sherlock

* Add scripts for speech runs on sherlock

* fix speech.jl script for sherlock

figures/songbird.jl

@@ -0,0 +1,42 @@
+
+using Revise
+import JLD
+import PyCall
+import CMF
+
+scipy_io = PyCall.pyimport("scipy.io")
+matfile = scipy_io.loadmat("../../cmf_data/MackeviciusData.mat")
+song = matfile["SONG"]
+
+L = 50
+K = 3
+
+algorithms = Dict(
+ "HALS" => :hals,
+ "Mult" => :mult,
+ "ANLS" => :anls
+)
+max_time = 60
+seed = sum([Int(c) for c in "INITIALIZE"]);
+
+res = Dict()
+
+# Warmstart algorithms
+for alg in keys(algorithms)
+ CMF.fit_cnmf(
+ song, K=K, L=L,
+ alg=algorithms[alg], max_itr=1, max_time=Inf
+ )
+end
+println("Finished warmstart")
+
+for alg in keys(algorithms)
+ res[alg] = CMF.fit_cnmf(
+ song, K=K, L=L,
+ alg=algorithms[alg], max_itr=Inf, max_time=max_time, seed=seed
+ )
+
+ println("Finished ", alg)
+end
+
+JLD.save("songbird.jld", res)

figures/songbird_run.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+#
+#SBATCH --job-name=test
+#
+#SBATCH --time=200:00
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=2
+#SBATCH --mem-per-cpu=16G
+
+ml julia/1.0.0
+ml viz
+ml py-matplotlib
+srun julia birdsong.jl

figures/speech.jl

@@ -0,0 +1,52 @@
+
+using CMF
+using PyPlot
+using DSP
+using WAV
+using JLD
+using ArgParse
+
+path = "../../cmf_data/ira_glass.wav"
+s, fs = WAV.wavread(path);
+
+# Downsample to 8 Khz
+fs_new = 8e3
+p = Int(round(fs / fs_new))
+s = s[1:p:end];
+
+
+# Log transform spectrogram
+start_idx = 400
+end_idx = 800
+S = spectrogram(s[:,1], 512, 384; window=hanning)
+t = time(S)
+f = freq(S)
+data = log10.(S.power)
+data = data .+ abs(minimum(data))
+
+# Parse commandline arguments to allow specifying model size
+s = ArgParseSettings()
+@add_arg_table s begin
+ "--alg"
+ help = "Alg to use for speech fit"
+ arg_type = String
+ default = "mult"
+end
+parsed_args = parse_args(ARGS, s)
+
+alg = Symbol(parsed_args["alg"])
+
+# fit once to compile
+CMF.fit_cnmf(data; L=12, K=20,
+ alg=alg, max_itr=1, max_time=6000,
+ check_convergence=false
+ )
+
+results = CMF.fit_cnmf(data; L=12, K=20,
+ alg=alg, max_itr=Inf, max_time=6000,
+ check_convergence=false
+ )
+
+
+alg_str = parsed_args["alg"]
+JLD.save("speech_run_$alg_str.jld", "results", results)

figures/speech_run.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+#
+#SBATCH --job-name=test
+#
+#SBATCH --time=200:00
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=2
+#SBATCH --mem-per-cpu=16G
+
+ml julia/1.0.0
+ml viz
+ml py-matplotlib
+sbatch julia speech.jl --alg "mult"
+sbatch julia speech.jl --alg "anls"
+sbatch julia speech.jl --alg "hals"

figures/synthetic_comparison.jl

@@ -0,0 +1,108 @@
+
+import CMF
+import PyPlot; plt = PyPlot;
+import Random; Random.seed!(0);
+import JLD
+using ArgParse
+using Dates
+using Random
+include("../datasets/synthetic.jl")
+
+
+# Parse commandline arguments to allow specifying model size
+s = ArgParseSettings()
+@add_arg_table s begin
+ "--N"
+ help = "Number of features"
+ arg_type = Int
+ default = 250
+ "--K"
+ help = "Number of components"
+ arg_type = Int
+ default = 5
+ "--L"
+ help = "Lag"
+ arg_type = Int
+ default = 20
+ "--alpha"
+ help = "Dirichlet param alpha for generating data"
+ arg_type = Float64
+ default = 0.1
+ "--sigma"
+ help = "Dirichlet param sigma for generating data"
+ arg_type = Float64
+ default = 0.2
+ "--p_h"
+ help = "Probability of nonzero h"
+ arg_type = Float64
+ default = 0.1
+ "--noise_scale"
+ help = "Noise scale for synthetic data"
+ arg_type = Float64
+ default = 0.1
+end
+parsed_args = parse_args(ARGS, s)
+
+
+# Model params
+N = parsed_args["N"]
+K = parsed_args["K"]
+L = parsed_args["L"]
+
+# Synthetic data params
+alpha = parsed_args["alpha"]
+p_h = parsed_args["p_h"]
+sigma = parsed_args["sigma"]
+noise_scale = parsed_args["noise_scale"]
+
+T_list = [500, 2500, 10_000, 50_000]
+runtimes = Dict(
+ 500 => 60,
+ 2500 => 120,
+ 10_000 => 400,
+ 50_000 => 1000
+)
+
+alg_list = [:hals, :mult, :anls]
+labels = Dict(:hals => "HALS", :mult => "MULT", :anls => "ANLS")
+
+results = Dict()
+results["args"] = parsed_args
+
+data_list = Dict()
+
+for T in T_list
+ data, trueW, trueH = synthetic_sequences(
+ K=K, N=N, L=L, T=T,
+ alpha=alpha, 
+ p_h=p_h, 
+ sigma=sigma, 
+ noise_scale=noise_scale
+ )
+ data_list[T] = (data, trueW, trueH)
+end
+
+# Run once to warmup algorithms
+for T in [500]
+ for alg in alg_list
+ CMF.fit_cnmf(data_list[T][1], alg=alg, max_itr=1)
+ end
+end
+
+for T in T_list
+ results[T] = Dict()
+
+ for alg in alg_list
+ # use same initialization for all algs
+ Random.seed!(0)
+
+ results[T][alg] = CMF.fit_cnmf(
+ data_list[T][1], alg=alg,
+ K=K, L=L, max_time=runtimes[T], max_itr=Inf
+ )
+ end
+end
+
+time_now = string(now())
+JLD.save("./synthetic_comparison_$time_now.jld", "results", results)
+

figures/synthetic_run.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+#
+#SBATCH --job-name=test
+#
+#SBATCH --time=300:00
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=2
+#SBATCH --mem-per-cpu=16G
+
+ml julia/1.0.0
+ml viz
+ml py-matplotlib
+srun julia synthetic_comparison.jl