Installation

$ brew install automake libtorch

gem'rumale-torch'

$ bundle install

$ gem install rumale-torch

Usage

Example 1. Pendigits dataset classification

$ wget https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass/pendigits$ wget https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass/pendigits.t

require'rumale'require'rumale/torch'Torch.manual_seed(1)device=Torch.device('cpu')# Loading pendigits dataset consisting of# 16-dimensional data divided into 10 classes.x,y=Rumale::Dataset.load_libsvm_file('pendigits')# Define a neural network in torch.rb framework.classMyNet <Torch::NN::Moduledefinitializesuper@dropout=Torch::NN::Dropout.new(p:0.5)@fc1=Torch::NN::Linear.new(16,128)@fc2=Torch::NN::Linear.new(128,10)enddefforward(x)x=@fc1.call(x)x=Torch::NN::F.relu(x)x=@dropout.call(x)x=@fc2.call(x)Torch::NN::F.softmax(x)endendnet=MyNet.new.to(device)# Create a classifier with neural network model.classifier=Rumale::Torch::NeuralNetClassifier.new(model:net,device:device,batch_size:10,max_epoch:50,validation_split:0.1,verbose:true)# Learning classifier.classifier.fit(x,y)# Saving model.Torch.save(net.state_dict,'pendigits.pth')File.binwrite('pendigits.dat',Marshal.dump(classifier))

require'rumale'require'rumale/torch'# Loading neural network model.classMyNet <Torch::NN::Moduledefinitializesuper@dropout=Torch::NN::Dropout.new(p:0.5)@fc1=Torch::NN::Linear.new(16,128)@fc2=Torch::NN::Linear.new(128,10)enddefforward(x)x=@fc1.call(x)x=Torch::NN::F.relu(x)# x = @dropout.call(x)x=@fc2.call(x)Torch::NN::F.softmax(x)endendnet=MyNet.newnet.load_state_dict(Torch.load('pendigits.pth'))# Loading classifier.classifier=Marshal.load(File.binread('pendigits.dat'))classifier.model=net# Loading test dataset.x_test,y_test=Rumale::Dataset.load_libsvm_file('pendigits.t')# Predict labels of test data.p_test=classifier.predict(x_test)# Evaluate predicted result.accuracy=Rumale::EvaluationMeasure::Accuracy.new.score(y_test,p_test)puts(format("Accuracy: %2.1f%%",accuracy *100))

$ ruby train.rbepoch:  1/50 - loss: 0.2073 - accuracy: 0.3885 - val_loss: 0.2074 - val_accuracy: 0.3853epoch:  2/50 - loss: 0.1973 - accuracy: 0.4883 - val_loss: 0.1970 - val_accuracy: 0.4893epoch:  3/50 - loss: 0.1962 - accuracy: 0.4997 - val_loss: 0.1959 - val_accuracy: 0.5013...epoch: 50/50 - loss: 0.1542 - accuracy: 0.9199 - val_loss: 0.1531 - val_accuracy: 0.9293$ ruby test.rbAccuracy: 91.2%

Example 2. Cross-validation with Rumale

$ wget https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/regression/housing

require'rumale'require'rumale/torch'Torch.manual_seed(1)device=Torch.device('cpu')# Loading pendigits dataset consisting of# 13-dimensional data with single target variable.x,y=Rumale::Dataset.load_libsvm_file('housing')# Define a neural network in torch.rb framework.classMyNet <Torch::NN::Moduledefinitializesuper@fc1=Torch::NN::Linear.new(13,128)@fc2=Torch::NN::Linear.new(128,1)enddefforward(x)x=@fc1.call(x)x=Torch::NN::F.relu(x)x=@fc2.call(x)endendnet=MyNet.new.to(device)# Create a regressor with neural network model.regressor=Rumale::Torch::NeuralNetRegressor.new(model:net,device:device,batch_size:10,max_epoch:100)# Create evaluation measure, splitting strategy, and cross validation.ev=Rumale::EvaluationMeasure::R2Score.newkf=Rumale::ModelSelection::ShuffleSplit.new(n_splits:5,test_size:0.1,random_seed:1)cv=Rumale::ModelSelection::CrossValidation.new(estimator:regressor,splitter:kf,evaluator:ev)# Perform 5-cross validation.report=cv.perform(x,y)# Output result.mean_score=report[:test_score].sum /kf.n_splitsputs(format("5-CV R2-score: %.3f",mean_score))

$rubycv.rb5-CVR2-score:0.755

Contributing

License