This repository contains applications for TT tensor learning using GPU tensor cores, including TT/TR tensor decompostion, TT tensor layer and DMRG algorithm.

cuTensor_TT_TR_Learning

ttdecomposition

third_order_GPU -------- third order tt decomposition

python -------- python implementations
unopt -------- unoptimized on single GPU
opt -------- optimized on single GPU

high_order_GPUs -------- high order tt decomposition

python -------- python implementations
ttdec6_singleGPU.cu -------- optimized single GPU
ttdec6_multiGPUs_2.cu -------- optimized on two GPUs
ttdec6_multiGPUs_4.cu -------- optimized on four GPUs
ttdec6_multiGPUs_8.cu -------- optimized on eight GPUs

trdecomposition

third_order_GPU -------- third order tr decomposition

python -------- python implementations
unopt -------- unoptimized on single GPU
opt -------- optimized on single GPU

high_order_GPUs -------- high order tr decomposition

python -------- python implementations
trdec6_singleGPU.cu -------- optimized on single GPU
trdec6_multiGPUs_2.cu -------- optimized on two GPUs
trdec6_multiGPUs_4.cu -------- optimized on four GPUs
trdec6_multiGPUs_8.cu -------- optimized on eight GPUs

ablation -------- ablition experiments

edrsvd_GPU -------- only EDrSVD
edrsvd_pipeline_GPU -------- EDrSVD using piepline
edrsvd_rensor_contractions_GPU -------- EDrSVD using tensor contractions

TT_tensor_layer

mnist.py -------- fully connect
mnist_tt.py -------- TT tensor layer
mnist_tt_half.py -------- TT tensor layer using apex

DMRG

python -------- TensorNetwork-JAX
baseline ------- unoptimized
optimized -------- optimized

Tensor and Deep Learning Lab LLC.

E-Mail:** [email protected]

github:**https://github.com/Dimo0316/cuTensor_TT_TR_Learning

Notice: both US and China patents are filed for this project.