GitHub - Ratbuyer/dlmc: Deep Learning Matrix Collection from google research

Branches Tags

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
rn50		rn50
transformer		transformer
README		README
dlmc.csv		dlmc.csv
rn50_batchsizes.txt		rn50_batchsizes.txt
rn50_matrices.txt		rn50_matrices.txt
transformer_matrices.txt		transformer_matrices.txt

Repository files navigation

Dataset of sparse matrices from https://arxiv.org/abs/2006.10901. Matrices collected from models from https://arxiv.org/abs/1902.09574.

The sparse matrices associated with each model are stored in their respective directories ("transformer" and "rn50"). The matrices are organized by the sparsification technique used to produce them and the final sparsity level. The complete list of matrices for each model can be found in `transformer_matrices.txt` and `rn50_matrices.txt`. Statistics for all matrices are included in dlmc.csv.

Each matrix is stored in its own file. The storage format starts with a 3-integer header "nrows, ncols, nnz" that describes the number of rows in the matrix, the number of columns in the matrix, and the number of nonzeros in the matrix. Each line that follows contains the integer column indices for nonzero values in a row, separated by spaces. All sparse matrices are laid out s.t. the forward pass is a left-multiplication by a sparse matrix, i.e. A*B=C, where A is sparse, B is dense and C is dense. Convolutional weight matrices are stored as the 2D matrices that would be needed to perform the matrix multiplication as part of a explicit-GEMM convolution (im2col + matmul). For these problems the product of the output spatial dimensions (i.e., the n-dimension of the MxKxN matmul with a single image) is included in `rn50_batchsizes.txt`.