peterzhang2029/adaptivesoftmax_torch

adaptive softmax implemented by torch

Word-level language modeling RNN Using adaptive softmax

By default you need download the preprocessed wikitext-2 dataset wikitext-2.zip and put the zip file under data folder.

Of course you can use your own text dataset as well.

cd data 
unzip wikitext-2.zip
cd ..
python main.py --epochs 6   # Train a LSTM on Wikitext-2 with CUDA
python main.py --epochs 6 --model LSTM_adaptive    # Train a LSTM with adaptive softmax on Wikitext-2 with CUDA

You may add --cuda option to accelerate the training process if you have a gpu.

Result:

For regular softmax:

Train:

epoch	epoch time	valid loss	valid ppl
end of epoch 1	time: 48.67s	valid loss 5.52	valid ppl 250.68
end of epoch 2	time: 48.61s	valid loss 5.28	valid ppl 196.87
end of epoch 3	time: 48.66s	valid loss 5.17	valid ppl 175.51
end of epoch 4	time: 48.90s	valid loss 5.09	valid ppl 162.27
end of epoch 5	time: 49.76s	valid loss 5.08	valid ppl 161.23
end of epoch 6	time: 49.63s	valid loss 5.01	valid ppl 149.77

Test:

test loss 4.94s
test ppl 139.35

all Time_cost: 296.55s

For adaptive softmax:

Train:

epoch	epoch time	valid loss	valid ppl
end of epoch 1	time: 28.68s	valid loss 5.56	valid ppl 259.24
end of epoch 2	time: 30.67s	valid loss 5.34	valid ppl 208.12
end of epoch 3	time: 30.43s	valid loss 5.21	valid ppl 183.21
end of epoch 4	time: 30.32s	valid loss 5.10	valid ppl 164.43
end of epoch 5	time: 28.89s	valid loss 5.09	valid ppl 162.26
end of epoch 6	time: 29.30s	valid loss 5.04	valid ppl 154.52

Test:

test loss 4.97s
test ppl 143.32

all Time_cost: 179.856s

Loss comparison

PPL comparison

How to get wikitext-2.zip

由于文件太大，所以放到网盘中了

欢迎关注公众号：机器学习爱好者

或者扫码：

关注后回复 adaptive 就可以获取下载地址啦