virtualenv -p python3 <your_virtualenv_path>
source <your_virtualenv_path>/bin/activate
pip3 install -r requirements.txtThe glove.6B.200d.txt under wordvec_files/ contains only the first 300 words from the original file. You can download the original files like glove.6B.200d.txt from here
After download the glove files, put the glove.6B.200d.txt file under directory wordvec_files/, if you do not use the glove.6B.200d.txt, you need to change the args of wordvec_file in main.py, then to create pretrained embedding and vocab cache using the glove file, you should run python main.py in the terminal, it takes about 30 seconds on my MacBook Pro 2018(13.3-inch) and it will print a few examples as below:
The you can play with it with a much faster speed, for instance:
python main.py --pred_word love --topk 10 means find the most similar 10 words of love according to the cosine similarity between their word vectors.
For GPU user, you can specify --use_gpu True in terminal only if torch.cuda.is_available() == True, otherwise the model will compute on cpu as default.
Test on my MacBook Pro 2018(13.3-inch)
| Metrics | My implementation | Gensim |
|---|---|---|
| Time to load data | 40+ seconds first time and less than 1 second after that | 120 seconds |
| Time to find 100 similar words | 0.65 seconds | 0.22 seconds |
| Time to find 1000 similar words | 0.76 seconds | 0.15 seconds |
| Time to find 10000 similar words | 0.91 seconds | 0.04 seconds |
My implementation costs more time when the topk becomes bigger because I use a heap to keep the topk most similar words rather than sort the cosine similarity between the target word and the whole vocabs and then return the first topk words as Gensim does, which is much faster than my impletation when topk is very big.
But due to the time to load wordvec_file for Gensim is too long——up to about 2 minutes, which is very annoying when you want to test a few different words many times. My implementation takes less than 1 second to load wordvec_file which is much faster, so it is very easy to play with it repeatly without waiting for a long time before getting the answer everytime.