.mtx to .npz

[CunyangWei]

I found that the method you provided for downloading the dataset no longer seems to be available. I'm curious whether using my own data and converting it to the .npz format required by FlashSparse in this way would be reasonable. Or do you have a better script for converting from .mtx to .npz?

Besides, FlashSparse doesn't seem to support non-square matrix?

import argparse
import numpy as np
from scipy.io import mmread
from scipy.sparse import coo_matrix

coo = mmread(xxx.mtx)
row = coo.row
col = coo.col
num_edges = coo.nnz
num_nodes_src = coo.shape[0]
num_nodes_dst = coo.shape[1]

data_dict = {
    'num_nodes_src': np.array(num_nodes_src, dtype=np.int32),
    'num_nodes_dst': np.array(num_nodes_dst, dtype=np.int32),
    'num_edges':     np.array(num_edges,     dtype=np.int32),
    'src_li':        row.astype(np.int32),
    'dst_li':        col.astype(np.int32)
}

np.savez(xxx.npz, **data_dict)

[JinliangShi]

The sparse matrices can be downloaded from the SuiteSparse repository, while GNN-related datasets are available on the official websites of DGL and PyG. After downloading, the matrices can be transformed from .tar.gz to .npz format using the provided script. Subsequently, the matrices can be partitioned and preprocessed using the FS_Block_gpu.preprocess_gpu_fs() function. An example implementation of this function is available in FlashSparse/eva/kernel/spmm/fs_fp16/mdataset2.py. The data conversion process is included in the /dataset.txt .

[JinliangShi]

Flashsparse can support non-square matrix