How to test a whole dataset? about rpg_asynet

[MessikommerNico]

The file sparse_VGG_test.py is a unittest to ensure that the output of SSC (facebook implementation) is identical to our asynchronous asynet. It does not test the prediction accuracy.
In order to avoid having too many files, we excluded many additional files including the file for the test evaluation.
To obtain the test accuracy, a quick and dirty hack would be to change the validation dataset path in the datalooder to the test dataset path and start the training script.

[978326187]

i wanna know how many epoch i need to achieve the result of the paper, i have trained 100 epoch using fb_sparse_vgg model, and the accuracy is just 0.6 in N-Caltech101.

[MessikommerNico]

The accuracy after 100 epochs should be around 0.65. The test accuracy reported in the paper was obtained with a model trained for 1500 epochs.

[978326187]

ok, thanks for your reply~
i found the paper said that the inference time using asynet is longer than sparseNet, will the asynet be faster than sparseNet in FPGA or some hardware whose computing power is not strong?
If not, I doubt that it's still not a good choice industrious application since the size of the model is the same as SparseNet.

[MessikommerNico]

It depends on the specific hardware implementation. With a smart memory access design, the asynchronous sparse convolutions can be much faster as the computational complexity is much lower compared to normal convolutions.

[978326187]

It would be fantastic if the specific hardware is created.

[978326187]

I just remembered what I wanted to ask is How to test a whole dataset using asynet . we can only using SSC in training script

[MessikommerNico]

The output of SSC and asynet is identical. Thus, the accuracy achieved by SSC is the same as achieved by asynet