mirror of
https://gitcode.com/gh_mirrors/ope/OpenFace.git
synced 2026-05-12 18:32:30 +00:00
54 lines
2.6 KiB
Matlab
54 lines
2.6 KiB
Matlab
function [ patch_expert, corr, rms_error] = CreatePatchExpert( samples, labels, unnormed_samples, normalisation_options)
|
|
%CREATEPATCHEXPERT Summary of this function goes here
|
|
% Detailed explanation goes here
|
|
|
|
num_examples = size(samples, 1);
|
|
|
|
region_length = normalisation_options.normalisationRegion - normalisation_options.patchSize + 1;
|
|
region_length = region_length(1) * region_length(2);
|
|
|
|
% this part sets the split boundaries for SVR training, logit training and test subsets
|
|
train_SVR_start = 1;
|
|
train_SVR_end = int32(normalisation_options.svmRatio * num_examples - 1);
|
|
train_SVR_end = train_SVR_end - mod(train_SVR_end, region_length);
|
|
|
|
train_logit_start = train_SVR_end + 1;
|
|
train_logit_end = train_logit_start + int32(normalisation_options.logitRatio * num_examples) - 1;
|
|
|
|
% make sure it ends within same area of interest (region)
|
|
train_logit_end = train_logit_end - mod(train_logit_end, region_length);
|
|
|
|
test_start = train_logit_end + 1;
|
|
test_end = size(samples,1);
|
|
|
|
% picking training data for SVR
|
|
examples_train_SVR = samples(train_SVR_start:train_SVR_end, :);
|
|
labels_train_SVR = labels(train_SVR_start:train_SVR_end);
|
|
|
|
% Train the SVR using liblinear (older version used libSVM,
|
|
% but liblinear is much faster)
|
|
[w, b] = Train_SVR(examples_train_SVR, labels_train_SVR);
|
|
|
|
% Now create the test dataset
|
|
examples_test = samples(test_start:test_end, :);
|
|
labels_test = labels(test_start:test_end);
|
|
|
|
% Training the logistic regressor now
|
|
examples_train_logit = samples(train_logit_start:train_logit_end, :);
|
|
labels_train_logit = labels(train_logit_start:train_logit_end, :);
|
|
|
|
patch_expert = CreateLogisticRegressor(examples_train_logit, labels_train_logit, w, b, normalisation_options);
|
|
|
|
% Evaluate the patch expert now
|
|
[rms_error, corr, ~] = EvaluatePatchExpert(examples_test, labels_test, patch_expert, false);
|
|
|
|
fprintf('Rms error %.3f, correlation %.3f\n', rms_error, corr);
|
|
|
|
% Assert that our implementation and the convolution based one are equivalent
|
|
[~, ~, responses_svm] = EvaluatePatchExpert(samples(1:size(unnormed_samples,1)*region_length,:), labels(1:size(unnormed_samples,1)*region_length), patch_expert, false);
|
|
[responses_ncc] = SVR_expert_ncc_response(unnormed_samples, patch_expert, normalisation_options, normalisation_options.normalisationRegion, region_length);
|
|
assert(mean(abs(responses_svm-responses_ncc))< 1e-2);
|
|
|
|
end
|
|
|