Skip to content

Latest commit

 

History

History
31 lines (29 loc) · 1.73 KB

reproduction.md

File metadata and controls

31 lines (29 loc) · 1.73 KB

What is undocumented

  1. Requires a linux machine to reproduce. Reproduced in Ubuntu 20.04 desktop and SMR's server which is also Ubuntu 20.04.
  2. The file http://nlp.stanford.edu/data/glove.6B.zip does not exist. I only tried TFIDF so did not face any problem regarding this.
  3. Needed to install runsolver. Here is how.
    1. git clone https://github.com/utpalbora/runsolver
    2. Follow the README of runsolver to install
    3. sudo apt install libnuma-dev
    4. Add runsolver/src to PATH
  4. pip3 install sentencepiece

Results

I only tried the TF-IDF version of SOAR. I used 1 hour timeout and 64GB memory limit as the original authors. The results are identical to the original one except for the following minor differences.

  1. The synthesis time often varied a lot. For some the original was faster and for some mine was faster.
  2. permute took arguments in different order. But the end result is same, so it does not matter.
    1. Example: x.permute(0,2,1) vs x.permute(0,1,2)
  3. The following differences were also observed
    1. alexnet
      1. line 15: has an additional x.permute(0,1,2,3).
      2. last line: Softmax(dim=-1) instead of Softmax(dim=1)
    2. conv_for_text
      1. line 15: MaxPool1d(2,stride=3,padding=1) instead of MaxPool1d(64,stride=3,padding=32)
      2. line 16: flatten(x,-2) instead of flatten(x,1)
    3. conv_pool_softmax
      1. Softmax(dim=None) instead of Softmax(dim=1)
    4. img_autoencoder
      1. line 11: Conv2d(32,64,(3,3),stride=(3,3),padding=(0,0)) instead of Conv2d(32,64,(3,3),stride=(1,1),padding=(0,0))
    5. vgg19
      1. line 66: Softmax(dim=-1) instead of Softmax(dim=1)
    6. line 10: Embedding(2,1000) instead of Embedding(1,1000)