search

baidu-research / DeepBench

Benchmarking Deep Learning operations on different hardware
Apache License 2.0
1.07k stars 239 forks source link

Typo in convolution configs? #42

Closed Kirpich30000 closed 7 years ago

Kirpich30000 commented 7 years ago

Hi, Looks like there is a typo in spreadsheets: image

Output of the first DeepSpeech convolution does not fit the second layer: out_W = (W + 2 pad_w - filter_w + 1) / stride_w = (700 + 0 - 5 + 1) / 2 = 348 out_H = (H + 2 pad_h - filter_h + 1) / stride_h = (161 + 0 - 20 + 1) / 2 = 71

I guess R should be filter height and S should be filter width. In that case DeepSpeech layers fit perfectly: out_W = (700 + 0 - 20 + 1) / 2 = 341 out_H = (161 + 0 - 5 + 1) / 2 = 79

Please also chech KWS case for the same issue.

sharannarang commented 7 years ago

@Kirpich30000 , Thanks for pointing out this issue!

Turns out the typo existed in all the inference results and reference spreadsheets. I've created a PR to fix it. Please take a look.