This repository contains testers for SparseTrain. SparsTrain only supports machines with AVX-512 support, such as the Intel Skylake-X generation. Please modify the Makefile to suit your environment. Compiling the testers requires SparseTrain and Intel VTune Amplifier being installed. You can run the testers with or without VTune. After running make, you will get 2 executables from each source file. The one with _verify suffix verifies the output from SparseTrain against a (very slow) reference convolution. Because the verification takes very long time, we suggest you only verify with small convolution configureations. For example, N=16 C=128 W=7 H=7 K=128 O=1 P=1 R=3 S=3. The one with _noverify as suffix does not perform such verification. The executables to test SparseTrain are:
./fwd_trans_test_verify
./fwd_trans_test_noverify
./bwd_trans_test_verify
./bwd_trans_test_noverify
./bww_trans_test_verify
./bww_trans_test_noverify
The others are for internal tests. fwd is forward propagation; bwd is backward propagation by input; bww is backward propagation by weights. To execute the tests, please provide the arguments as:
./fwd_trans_test_noverify [sparsity percentage] [iterations] [batch size] [input channel] [input width] [input height] [output channel] [horizontal stride] [vertical stride] [filter width] [filter height]
For example
./fwd_trans_test_noverify 70 10000 16 256 14 14 512 1 1 3 3
runs the forward propagation at 70% sparsity for 10000 iterations. The convolution size is set to N=16 C=256 W=14 H=14 K=512 O=1 P=1 R=3 S=3.
The testing programs output the execution time in cycles measured with the RDTSCP instruction to standard out.
The 3 spreadsheets contain profiled sparsity of layers from VGG16, ResNet-34, and ResNet-50 during their training from beginning to convergence.