- Requires a linux machine to reproduce. Reproduced in Ubuntu 20.04 desktop and SMR's server which is also Ubuntu 20.04.
- 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.
- Needed to install
runsolver
. Here is how.git clone https://github.com/utpalbora/runsolver
- Follow the README of runsolver to install
sudo apt install libnuma-dev
- Add
runsolver/src
to PATH
pip3 install sentencepiece
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.
- The synthesis time often varied a lot. For some the original was faster and for some mine was faster.
permute
took arguments in different order. But the end result is same, so it does not matter.- Example:
x.permute(0,2,1)
vsx.permute(0,1,2)
- Example:
- The following differences were also observed
- alexnet
- line 15: has an additional
x.permute(0,1,2,3)
. - last line:
Softmax(dim=-1)
instead ofSoftmax(dim=1)
- line 15: has an additional
- conv_for_text
- line 15:
MaxPool1d(2,stride=3,padding=1)
instead ofMaxPool1d(64,stride=3,padding=32)
- line 16:
flatten(x,-2)
instead offlatten(x,1)
- line 15:
- conv_pool_softmax
Softmax(dim=None)
instead ofSoftmax(dim=1)
- img_autoencoder
- line 11:
Conv2d(32,64,(3,3),stride=(3,3),padding=(0,0))
instead ofConv2d(32,64,(3,3),stride=(1,1),padding=(0,0))
- line 11:
- vgg19
- line 66:
Softmax(dim=-1)
instead ofSoftmax(dim=1)
- line 66:
- line 10:
Embedding(2,1000)
instead ofEmbedding(1,1000)
- alexnet