# This file was generated by PredictMD version 0.34.21
# For help, please visit https://predictmd.net
using PredictMDExtra
PredictMDExtra.import_all()
using PredictMD
PredictMD.import_all()
### Begin project-specific settings
DIRECTORY_CONTAINING_THIS_FILE = @__DIR__
PROJECT_DIRECTORY = dirname(
joinpath(splitpath(DIRECTORY_CONTAINING_THIS_FILE)...)
)
PROJECT_OUTPUT_DIRECTORY = joinpath(
PROJECT_DIRECTORY,
"output",
)
mkpath(PROJECT_OUTPUT_DIRECTORY)
mkpath(joinpath(PROJECT_OUTPUT_DIRECTORY, "data"))
mkpath(joinpath(PROJECT_OUTPUT_DIRECTORY, "models"))
mkpath(joinpath(PROJECT_OUTPUT_DIRECTORY, "plots"))
### End project-specific settings
### Begin Knet neural network classifier code
Random.seed!(999)
trainingandtuning_features_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"trainingandtuning_features_df.csv",
)
trainingandtuning_labels_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"trainingandtuning_labels_df.csv",
)
testing_features_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"testing_features_df.csv",
)
testing_labels_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"testing_labels_df.csv",
)
training_features_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"training_features_df.csv",
)
training_labels_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"training_labels_df.csv",
)
tuning_features_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"tuning_features_df.csv",
)
tuning_labels_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"tuning_labels_df.csv",
)
trainingandtuning_features_df = DataFrames.DataFrame(
FileIO.load(
trainingandtuning_features_df_filename;
type_detect_rows = 100,
)
)
trainingandtuning_labels_df = DataFrames.DataFrame(
FileIO.load(
trainingandtuning_labels_df_filename;
type_detect_rows = 100,
)
)
testing_features_df = DataFrames.DataFrame(
FileIO.load(
testing_features_df_filename;
type_detect_rows = 100,
)
)
testing_labels_df = DataFrames.DataFrame(
FileIO.load(
testing_labels_df_filename;
type_detect_rows = 100,
)
)
training_features_df = DataFrames.DataFrame(
FileIO.load(
training_features_df_filename;
type_detect_rows = 100,
)
)
training_labels_df = DataFrames.DataFrame(
FileIO.load(
training_labels_df_filename;
type_detect_rows = 100,
)
)
tuning_features_df = DataFrames.DataFrame(
FileIO.load(
tuning_features_df_filename;
type_detect_rows = 100,
)
)
tuning_labels_df = DataFrames.DataFrame(
FileIO.load(
tuning_labels_df_filename;
type_detect_rows = 100,
)
)
smoted_training_features_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"smoted_training_features_df.csv",
)
smoted_training_labels_df_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"smoted_training_labels_df.csv",
)
smoted_training_features_df = DataFrames.DataFrame(
FileIO.load(
smoted_training_features_df_filename;
type_detect_rows = 100,
)
)
smoted_training_labels_df = DataFrames.DataFrame(
FileIO.load(
smoted_training_labels_df_filename;
type_detect_rows = 100,
)
)
categorical_feature_names_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"categorical_feature_names.jld2",
)
continuous_feature_names_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"data",
"continuous_feature_names.jld2",
)
categorical_feature_names = FileIO.load(
categorical_feature_names_filename,
"categorical_feature_names",
)
continuous_feature_names = FileIO.load(
continuous_feature_names_filename,
"continuous_feature_names",
)
feature_names = vcat(categorical_feature_names, continuous_feature_names)
single_label_name = :Class
negative_class = "benign"
positive_class = "malignant"
single_label_levels = [negative_class, positive_class]
categorical_label_names = Symbol[single_label_name]
continuous_label_names = Symbol[]
label_names = vcat(categorical_label_names, continuous_label_names)
knet_mlp_predict_function_source = """
function knetmlp_predict(
w,
x0::AbstractArray;
probabilities::Bool = true,
)
x1 = Knet.relu.( w[1]*x0 .+ w[2] )
x2 = Knet.relu.( w[3]*x1 .+ w[4] )
x3 = w[5]*x2 .+ w[6]
unnormalizedlogprobs = x3
if probabilities
normalizedlogprobs = Knet.logp(unnormalizedlogprobs; dims = 1)
normalizedprobs = exp.(normalizedlogprobs)
return normalizedprobs
else
return unnormalizedlogprobs
end
end
"""
knet_mlp_loss_function_source = """
function knetmlp_loss(
predict::Function,
modelweights,
x::AbstractArray,
ytrue::AbstractArray;
L1::Real = Float64(0),
L2::Real = Float64(0),
)
loss = Knet.nll(
predict(
modelweights,
x;
probabilities = false,
),
ytrue;
dims = 1,
)
if L1 != 0
loss += L1 * sum(sum(abs, w_i) for w_i in modelweights[1:2:end])
end
if L2 != 0
loss += L2 * sum(sum(abs2, w_i) for w_i in modelweights[1:2:end])
end
return loss
end
"""
feature_contrasts = PredictMD.generate_feature_contrasts(
smoted_training_features_df,
feature_names,
)
knetmlp_modelweights = Any[
Float64.(
0.1f0*randn(Float64,64,feature_contrasts.num_array_columns_without_intercept)
),
Float64.(
fill(Float64(0),64,1)
),
Float64.(
0.1f0*randn(Float64,32,64)
),
Float64.(
fill(Float64(0),32,1)
),
Float64.(
0.1f0*randn(Float64,2,32)
),
Float64.(
fill(Float64(0),2,1)
),
]
knetmlp_losshyperparameters = Dict()
knetmlp_losshyperparameters[:L1] = Float64(0.0)
knetmlp_losshyperparameters[:L2] = Float64(0.0)
knetmlp_optimizationalgorithm = :Momentum
knetmlp_optimizerhyperparameters = Dict()
knetmlp_minibatchsize = 48
knet_mlp_classifier =
PredictMD.single_labelmulticlassdataframeknetclassifier(
feature_names,
single_label_name,
single_label_levels;
package = :Knet,
name = "Knet MLP",
predict_function_source = knet_mlp_predict_function_source,
loss_function_source = knet_mlp_loss_function_source,
losshyperparameters = knetmlp_losshyperparameters,
optimizationalgorithm = knetmlp_optimizationalgorithm,
optimizerhyperparameters = knetmlp_optimizerhyperparameters,
minibatchsize = knetmlp_minibatchsize,
modelweights = knetmlp_modelweights,
printlosseverynepochs = 1,
maxepochs = 50,
feature_contrasts = feature_contrasts,
)
PredictMD.parse_functions!(knet_mlp_classifier)
PredictMD.fit!(
knet_mlp_classifier,
smoted_training_features_df,
smoted_training_labels_df,
tuning_features_df,
tuning_labels_df,
)
PredictMD.set_max_epochs!(knet_mlp_classifier, 100)
PredictMD.fit!(
knet_mlp_classifier,
smoted_training_features_df,
smoted_training_labels_df,
tuning_features_df,
tuning_labels_df,
)
knet_learningcurve_lossvsepoch = PredictMD.plotlearningcurve(
knet_mlp_classifier,
:loss_vs_epoch;
);
display(knet_learningcurve_lossvsepoch)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_learningcurve_lossvsepoch.pdf",
),
knet_learningcurve_lossvsepoch,
)
knet_learningcurve_lossvsepoch_skip10epochs = PredictMD.plotlearningcurve(
knet_mlp_classifier,
:loss_vs_epoch;
startat = 10,
endat = :end,
);
display(knet_learningcurve_lossvsepoch_skip10epochs)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_learningcurve_lossvsepoch_skip10epochs.pdf",
),
knet_learningcurve_lossvsepoch_skip10epochs,
)
knet_learningcurve_lossvsiteration = PredictMD.plotlearningcurve(
knet_mlp_classifier,
:loss_vs_iteration;
window = 50,
sampleevery = 10,
);
display(knet_learningcurve_lossvsiteration)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_learningcurve_lossvsiteration.pdf",
),
knet_learningcurve_lossvsiteration,
)
knet_learningcurve_lossvsiteration_skip100iterations =
PredictMD.plotlearningcurve(
knet_mlp_classifier,
:loss_vs_iteration;
window = 50,
sampleevery = 10,
startat = 100,
endat = :end,
);
display(knet_learningcurve_lossvsiteration_skip100iterations)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_learningcurve_lossvsiteration_skip100iterations.pdf",
),
knet_learningcurve_lossvsiteration_skip100iterations,
)
knet_mlp_classifier_hist_training =
PredictMD.plotsinglelabelbinaryclassifierhistogram(
knet_mlp_classifier,
smoted_training_features_df,
smoted_training_labels_df,
single_label_name,
single_label_levels,
);
display(knet_mlp_classifier_hist_training)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_mlp_classifier_hist_training.pdf",
),
knet_mlp_classifier_hist_training,
)
knet_mlp_classifier_hist_testing =
PredictMD.plotsinglelabelbinaryclassifierhistogram(
knet_mlp_classifier,
testing_features_df,
testing_labels_df,
single_label_name,
single_label_levels,
);
display(knet_mlp_classifier_hist_testing)
PredictMD.save_plot(
joinpath(
PROJECT_OUTPUT_DIRECTORY,
"plots",
"knet_mlp_classifier_hist_testing.pdf",
),
knet_mlp_classifier_hist_testing,
)
show(
PredictMD.singlelabelbinaryclassificationmetrics(
knet_mlp_classifier,
smoted_training_features_df,
smoted_training_labels_df,
single_label_name,
positive_class;
sensitivity = 0.95,
);
allrows = true,
allcols = true,
splitcols = false,
)
show(
PredictMD.singlelabelbinaryclassificationmetrics(
knet_mlp_classifier,
testing_features_df,
testing_labels_df,
single_label_name,
positive_class;
sensitivity = 0.95,
);
allrows = true,
allcols = true,
splitcols = false,
)
knet_mlp_classifier_filename = joinpath(
PROJECT_OUTPUT_DIRECTORY,
"models",
"knet_mlp_classifier.jld2",
)
PredictMD.save_model(knet_mlp_classifier_filename, knet_mlp_classifier)
### End Knet neural network classifier code
# This file was generated by PredictMD version 0.34.21
# For help, please visit https://predictmd.netThis page was generated using Literate.jl.