Run pretrained ELMo model to get single sentence perplexity.

Modified from Allenai bilm-tf

pip install tensorflow-gpu==1.2 h5py
python setup.py install

1. Data file format: each line in a file is a sentence to calculate perplexity. data

2. Split data file into pieces, one sentence per piece.

cd data
split sents.txt -d -l 1 -a 4 cs

3. Run the evaluation script.

sh evaluate.sh

4. The perplexity score is shown in stdout

...
5946: 129.57085
5947: 1412.2032
5948: 5172.711
5949: 2126.5542
...

Sentence line number followed by the perplxity (unnormalized by sentence length)

To finetune the ELMo on additional corpus, first downlaod the pretrained model to models/

The tensorflow checkpoint is available by downloading these files:

vocabulary checkpoint options 1 2 3

|
|--models
    |--vocab-2016-09-10.txt
    |--checkpoint
            |--checkpoint
            |--options.json
            |--model.ckpt-935588.meta
            |--model.ckpt-935588.index
            |--model.ckpt-935588.data-00000-of-00001

Then use the following script.

sh finetune.sh

After finetuning the model, you can run the evaluation again to see the finetune effect.