Introduction

This is the code of MELD-ED, including a training of a deep Locality Sensitive Multi-Embedding (LSME) functionLSME and a neighbor search procedureneighbor_index_experiments.

A framework that can train an (d1, d2)-LSME function with customized parameters: N, length of sequence; k, number of embedding; m dimensions, and through δ. You can train the model using the simulation dataset provided at zenodo, with N = 20 or N = 100, or generate simulation sequences of the length you need with the code at simulation.

Usage

There is an example of using our code:

• Installation: python vision >= 3.9; conda --version 24.11.3; torch (2.7.1+cu128)

• Train/valid/test data loading by data_reader_pn_.py:

from data_reader_pn_ import data_load_bd2 

df_tr, df_v, df_test = data_load_bd2(rate, d1, d2, data_path, num_test, num_train_valid)

which require the following parameters: (d1, d2); rate of train and valid (like rate = 0.9means 90% of whole set for training and 10% for valdiation); num_test is the number you need for each edit distance for testing; num_train_valid is the number you need for each edit distance for training and validation.

• Model training: the structure of Siamese model, hinge loss, and train functions are saved in model_loss_train.py. Run python runner.py to train the model and evaluate its performance by displaying the loss per episode and accuracy. A quick example show how to train a (1, 3)-LSME function for sequence length N = 20, k = 20, m = 40, δ = 10:

def main():
    # run example 
    N_len = 20  # sequence length N
    d1, d2 = [1, 3] 
    m_dim = 40  # dimension of embedding vectors m
    batch_size = 5000
    num_b = 20  # number of embedding vectors K
    delta = 10
    a_ = torch.rand(100, 1, 4, N_len).to(device)
    rate = 0.9  # train/valid = 9:1

    data_path = '~/data/seq20n/'
    path = '~/'

    num_test, num_train_valid = [20000, 100000]  
    # e.g., 20000 per edit pairs for test, 100000 for train+valid

    Training_Evaluation_Parameter_Set(d1, d2, rate, a_, data_path, path, num_test, num_train_valid, batch_size, delta, m_dim, num_b)

The structure and parameter of trained model will be saved at ~/models/SiamNNL1_nm_Inp_Model_2_20k_40m_(1-3)s_10delta.pt, or you can directly use the model we trained which are saved at zenodo. The performance of the model is saved at ~/results/

• Embedding generation: After finishing the step of training or loading the models, run python embedding_generator.py to get embeddings of sequences. With the generated embedding, you can move to the step of neighbor search procedure by neighbor_index_experiments.