The ro_sgns from alexgrinch

ro_sgns's People

Contributors

Stargazers

Watchers

ro_sgns's Issues

SGD-SGNS

Dear Alex

Is the method SGD-SGNS included in the repo?

If not, may I know which implementation you used for experiment?

Thanks,
Matt

Data issue

In enwik_experiments.ipynb, the data data/enwik/enwik9.txt seems to be needed. Could you please provide this data in the repo?

Thanks,
Matt

Corr Test

Dear Alex

I am trying to run the correlation test.

Based on your datasets_corr function in functions.py, I am doing something as follows

opt_experiment(model_enwik,
               mode='PS', 
               d=100,
               eta = 5e-5, # step size
               MAX_ITER=7,
               from_iter=0,
               start_from='SVD',
               init=(True, C_svd, W_svd))

from scipy.spatial.distance import cosine
from scipy.stats import spearmanr

datasets_path = ".\datasets"
indices = np.load(open(datasets_path+'/indices.npz', 'rb'))
name = 'wordsim_sim'

pairs_num = indices['0'+name].size
idx = np.arange(pairs_num)
        
ind1 = indices['0'+name][idx]
ind2 = indices['1'+name][idx]
scores = indices['2'+name][idx]

W, C = model_enwik.W, model_enwik.C
G = W
G = G / np.linalg.norm(G, axis=0)
cosines = (G[:,ind1]*G[:,ind2]).sum(axis=0)
spearmanr(cosines, scores)[0]```

The code works find, but the result is -0.11981128713175491, which I expect to be 0.729 in the paper. I am pretty sure I misunderstand something and code something wrong.

Could you please help to take a look at what is going wrong?


Thanks,
Matt

Recommend Projects