---

Fast Siamese Recurrent Neural Network Approximation for the Triangular Global Alignment Kernel

Global Alignment Kernel

1. Copy compute_gram_matrix.json_sample to any name what you want. $cd experiments
   $cp compute_gram_matrix.json_sample (new_json_file_name)
2. Specify

• dataset_type: UCIauslan, UCIcharacter or upperChar
• dataset_location: the directory which holds the dataset
• output_dir: directory to save the results
• output_filename_format: format of the output file
• sigma and triangular: gak variables
• data_augmentation_size: number of times to augment

3. Run the experiment $python3 compute_gram_matrix.py with (new_json_file_name)

GRAM Matrix-based approach

Matrix completion

1. Copy complete_matrix_(algorithm).json_sample to any name what you want. $cd experiments $cp compute_gram_matrix.json_sample (new_json_file_name)
2. Modify the parameters "fast_rnn" requires a pretrained model.
3. Run the experiment $python3 complete_matrix.py with (new_json_file_name)

Classification error

1. Run the experiment $python3 compute_classification_errors.py with (parameter)=(new_value)

Feature-based approach

Feature mapping approximation

1. Copy linear_svm.json_sample to any name what you want. $cd experiments $cp linear_svm.json_sample (new_json_file_name)
2. Modify the parameters This method requires a pretrained model.
3. Run the experiment $python3 linear_svm.py with (new_json_file_name)