def eval_w2v_wordsim999():
fasttext = we.WordEmbedding('D://Codes/NSE/data/raw/embeddings/fastText-crawl-300d-2M.vec')
simlex = get_simlex999('D://Codes/NSE/data/output/vocab_unigram.txt')
results = []
for pair in simlex:
w1 = pair[0]
w2 = pair[1]
if w1 not in fasttext.words or w2 not in fasttext.words:
print('Cant find', w1, w2)
continue
w1v = fasttext.vecs[fasttext.index[w1]]
w2v = fasttext.vecs[fasttext.index[w2]]
results.append([w1, w2, pair[2], np.sum(np.multiply(w1v, w2v))])
print(stats.pearsonr([i[2] for i in results], [i[3] for i in results])[0], stats.spearmanr([i[2] for i in results], [i[3] for i in results]).correlation)
def eval_w2v_wordsim999():
sswe = we.WordEmbedding('D://Codes/NSE/data/raw/embeddings/sswe-r.txt')
simlex = get_simlex999('D://Codes/NSE/data/output/vocab_unigram.txt')
results = []
for pair in simlex:
w1 = pair[0]
w2 = pair[1]
if w1 not in sswe.words or w2 not in sswe.words:
print('Cant find', w1, w2)
continue
w1v = sswe.vecs[sswe.index[w1]]
w2v = sswe.vecs[sswe.index[w2]]
results.append([w1, w2, pair[2], np.sum(np.multiply(w1v, w2v))])
print(stats.pearsonr([i[2] for i in results], [i[3] for i in results])[0], stats.spearmanr([i[2] for i in results], [i[3] for i in results]).correlation)
for i in sorted(results, key=lambda x: x[3]):
print(i)
def eval_w2v_principal_component():
sswe = we.WordEmbedding('D://Codes/NSE/data/raw/embeddings/sswe-r.txt')
wordpair = [line.strip().split() for line in open('D://Codes/NSE/data/used/seeds/wordpairs-greater-than-0.5').readlines()]
sub_words_p = [line[0] for line in wordpair]
sub_words_n = [line[1] for line in wordpair]
matrix = []
for index in range(len(sub_words_p)):
w1 = sub_words_p[index]
w2 = sub_words_n[index]
if w1 not in sswe.words or w2 not in sswe.words:
print('Cant find', w1, w2)
continue
w1v = sswe.vecs[sswe.index[w1]]
w2v = sswe.vecs[sswe.index[w2]]
center = (w1v + w2v) / 2
matrix.append(w1v - center)
matrix.append(w2v - center)
pca = PCA(n_components=10)
pca.fit(matrix)
print(pca.explained_variance_ratio_)
def run():
new_words = []
matrix = []
coll = get_sentiwords()
fasttext = we.WordEmbedding(
'D://Codes/NSE/data/raw/embeddings/fastText-crawl-300d-2M.vec')
for i in coll:
word = i[0]
if word not in fasttext.words:
# print(word, 'can not be found in fasttext')
continue
else:
if word not in new_words:
new_words.append(word)
matrix.append(fasttext.vecs[fasttext.index[word]])
f1 = open('D://Codes/NSE/data/used/embeddings/word-picked', 'w')
f1.writelines("\n".join(new_words))
f2 = open('D://Codes/NSE/data/used/embeddings/word-picked.vec', 'w')
f2.write("\n".join([
w + " " + " ".join([str(x) for x in v])
for w, v in zip(new_words, matrix)
]))
import we
import numpy as np
picked = we.WordEmbedding('D://Codes/NSE/data/used/embeddings/word-picked.vec')
import argparse
import random
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--f', default='embeddings/glove', type=str)
parser.add_argument('--o', default='my_emb', type=str)
parser.add_argument('--dim', default=300, type=int)
args = parser.parse_args()
if args.o == 'my_emb':
output_path = f'my_emb_{random.randint(1,10000)}.txt'
else:
output_path = args.o
print(f"Converting: {args.f}")
E = we.WordEmbedding(args.f)
lines = [f'{len(E.words)} {args.dim}']
for word in tqdm(E.words):
vector = ' '.join([str(v) for v in E.v(word)])
word = ' '.join([w for w in word.split()])
line = word + ' ' + vector
lines.append(line)
with open(output_path, 'w') as handle:
handle.write('\n'.join(lines))
print(f"File saved in {output_path}")
idb is calculated using original debiased therefore you need to provide it here."
)
parser.add_argument(
'--dictf',
default="../data/glove_ns.pkl",
type=str,
help="pkl file for neighbour dict (see `make_neighbours.py` for this)")
parser.add_argument(
'--bias_list',
default="../data/debias.pkl",
type=str,
help="pkl file for list of words to debias (V_d, debias set in paper)."
)
args = parser.parse_args()
E_orig = we.WordEmbedding(args.orig_emb)
X = we.WordEmbedding(args.f)
g = utils.get_g(E_orig)
biased_words = get_words(args.bias_list)
vals, wts = gipe(biased_words, X, E_orig, g, args.dictf, args.idb_thresh,
args.n)
gipe_score = score(vals, wts)
print(gipe_score)
fname = "gipe-results.txt"
with open(fname, 'a+') as handle:
handle.write(
"=========================================================\n")
handle.write(
parser.add_argument(
"gendered_words_filename",
help="File containing words not to neutralize (one per line)")
parser.add_argument("equalize_filename", help="???.bin")
parser.add_argument("debiased_filename", help="???.bin")
args = parser.parse_args()
print(args)
with open(args.definitional_filename, "r") as f:
defs = json.load(f)
with open(args.equalize_filename, "r") as f:
equalize_pairs = json.load(f)
with open(args.gendered_words_filename, "r") as f:
gender_specific_words = json.load(f)
E = we.WordEmbedding(args.embedding_filename)
print("Debiasing...")
debias(E, gender_specific_words, defs, equalize_pairs)
print("Saving to file...")
if args.embedding_filename[-4:] == args.debiased_filename[-4:] == ".bin":
E.save_w2v(args.debiased_filename)
else:
E.save(args.debiased_filename)
print("\n\nDone!\n")
def get_embedding_and_g(filename):
E = we.WordEmbedding(filename)
g = utils.get_g()
g = torch.Tensor(g)
return E, g
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--emb_file', default="../embeddings/glove", type=str)
parser.add_argument('--dict_dir', default="output", type=str)
parser.add_argument('--out_dir', default="finals", type=str)
parser.add_argument('--out_fname', default="RAN-GloVe", type=str)
args = parser.parse_args()
emb_file = args.emb_file
dict_dir = args.dict_dir
out_fname = args.out_fname
out_dir = args.out_dir
E = we.WordEmbedding(emb_file)
fnames = glob.glob(f'{dict_dir}/*.dict.pickle')
D = {}
for fname in fnames:
print(f"Reading {fname}...")
d = read(fname)
for k in d:
D[k] = d[k]
new_embs = copy.deepcopy(E.vecs)
for w in tqdm(D):
new_embs[E.index[w]] = D[w]
if not os.path.exists(out_dir):
# Hyperparameters and other
EMB = 'data/glove' #path to embedding in vector-vocab format. See we.py to know how this format works.
words_list_file = conf.get(
'WORDS_LIST_FILE'
) #pickle file list of words to debias (small list from bulakbasi et. al.)
bs = conf.getint('BS')
nz = conf.getint('NZ')
lrd = conf.getflooat('LRD')
lrg = conf.getfloat('LRG')
beta1 = conf.getfloat('BETA1')
epochs = conf.getfloat('EPOCHS')
PATH_D = conf.get('PATH_D')
PATH_G = conf.get('PATH_G')
E = we.WordEmbedding(EMB) #Load embedding
g = gender_subspace_simple(E) #Get gender direction
with open(words_list_file, 'rb') as handle:
words = pickle.load(handle) #Load word list to debias
words = clean_vocab(words, E) #Clean word list
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# init dataset and dataloader
train_dataset = gender_set(words, E, g, device)
train_loader = DataLoader(train_dataset, batch_size=bs, shuffle=True)
netG = Generator(nz).to(device)
netD = SexistDiscriminator().to(device)
