def get_validate_phrases(args):
pairs = []
phrases = []
for filename in ["valid.txt"]:
with codecs.open(os.path.join(args.data_dir, filename),
encoding="utf-8") as f:
f.readline()
for line in f:
parts = line.strip().split("\t")
pair = {
"text_1": parts[3],
"text_2": parts[4],
"decision": float(parts[0])
}
pairs.append(pair)
pairs = pairs
true = [x["decision"] for x in pairs]
with open(os.path.join(args.save_dir, 'chars_vocab.pkl'), 'rb') as f:
chars, _ = cPickle.load(f)
for pair in pairs:
phrases.append(noise_generator(pair["text_1"], args.noise_level,
chars))
phrases.append(noise_generator(pair["text_2"], args.noise_level,
chars))
return phrases, true
def get_validate_entailment(args):
pairs = []
phrases = []
import pandas as pd
valid_path = os.path.join(args.data_dir, "valid.txt")
if valid_path.__contains__("quora"):
full_df = pd.read_csv(valid_path, sep='\t')[:300]
decision = "duplicate"
else:
decision = "gold_label"
full_df = pd.read_csv(valid_path)
for index, row in full_df.iterrows():
pair = {
"text_1": row['sentence1'],
"text_2": row["sentence2"],
"decision": int(row[decision])
} #ist(filter(lambda x: x.isdigit(), row["gold_label"]))[0]
pairs.append(pair)
pairs = pairs
true = [x["decision"] for x in pairs]
with open(os.path.join(args.save_dir, 'chars_vocab.pkl'), 'rb') as f:
chars, _ = cPickle.load(f)
for pair in pairs:
phrases.append(noise_generator(pair["text_1"], args.noise_level,
chars))
phrases.append(noise_generator(pair["text_2"], args.noise_level,
chars))
return phrases, true
def svm_robust_score(args, data, labels):
idx_for_split = int(0.2 * len(data))
phrases = []
pred = []
for index, row in data.iterrows():
phrases.append(
noise_generator(row["sentence1"], args.noise_level, chars))
phrases.append(
noise_generator(row["sentence2"], args.noise_level, chars))
from sample import sample_multi
results = np.squeeze(
np.vsplit(sample_multi(args.save_dir, phrases, args.model_type),
len(phrases)))
for i in range(0, len(results), 2):
v1 = results[i]
v2 = results[i + 1]
if (v1 == np.zeros_like(v1)).all() or (v2 == np.zeros_like(v2)).all():
print(i)
pred.append(1 - cosine(v1, v2))
if math.isnan(pred[-1]):
pred[-1] = 0.5
pr = pd.DataFrame(pred)
train = pr.iloc[idx_for_split:]
test = pr.iloc[:idx_for_split]
train_label = labels[idx_for_split:]
test_label = labels[:idx_for_split]
roc_auc = linear_svm(train, test, train_label, test_label)
with open("results_entail_" + args.model_type + ".txt", "at") as f_out:
# f_out.write("robust,%.2f,%.3f\n" % (args.noise_level, mean_squared_error(true, pred)))
f_out.write(args.mode + ",%.2f,%.3f\n" % (args.noise_level, roc_auc))
def get_robust_score(args, pairs, true):
if "robust" in args.mode:
pred = []
phrases = []
for index, row in pairs.iterrows():
phrases.append(
noise_generator(row["sentence1"], args.noise_level, chars))
phrases.append(
noise_generator(row["sentence2"], args.noise_level, chars))
from sample import sample_multi
results = np.vsplit(
sample_multi(args.save_dir, phrases, args.model_type),
len(phrases))
for i in range(0, len(results), 2):
v1 = results[i]
v2 = results[i + 1]
if (v1 == np.zeros_like(v1)).all() or (v2
== np.zeros_like(v2)).all():
print(i)
pred.append(1 - cosine(v1, v2))
if math.isnan(pred[-1]):
pred[-1] = 0.5
with open("results_entail" + args.model_type + ".txt", "at") as f_out:
# f_out.write("robust,%.2f,%.3f\n" % (args.noise_level, mean_squared_error(true, pred)))
f_out.write(args.mode + ",%.2f,%.3f\n" %
(args.noise_level, roc_auc_score(true, pred)))
def svm_robust_score(args,data, labels):
idx_for_split = int(0.2 * len(data))
phrases = []
pred = []
for index, row in data.iterrows():
phrases.append(noise_generator(row["sentence1"], args.noise_level, chars))
phrases.append(noise_generator(row["sentence2"], args.noise_level, chars))
from sample import sample_multi
results = np.squeeze(np.vsplit(sample_multi(args.save_dir, phrases, args.model_type), len(phrases)))
#pairs_vectors = zip(results[0::2], results[1::2])
# df = pd.DataFrame(columns=['cosine', 'canberra', 'cityblock', 'euclidean', 'minkowski', 'braycurtis',"skew_q1","skew_q2"\
# "kur_q1","kur_q2", "skew_diff", "kur_diff"])
# df['cosine'] = [cosine(x, y) for (x, y) in pairs_vectors]
# print(len(df))
# df['canberra'] = [canberra(x, y) for (x, y) in zip(results[0::2], results[1::2])]
# df['cityblock'] = [cityblock(x, y) for (x, y) in zip(results[0::2], results[1::2])]
# df['euclidean'] = [euclidean(x, y) for (x, y) in zip(results[0::2], results[1::2])]
# df['minkowski'] = [minkowski(x, y, 3) for (x, y) in zip(results[0::2], results[1::2])]
# df['braycurtis'] = [braycurtis(x, y) for (x, y) in zip(results[0::2], results[1::2])]
# question1_vec = results[0::2]
# question2_vec = results[1::2]
# data['skew_q1'] = [skew(x) for x in question1_vec]
# data['skew_q2'] = [skew(x) for x in question2_vec]
# data['kur_q1'] = [kurtosis(x) for x in question1_vec]
# data['kur_q2'] = [kurtosis(x) for x in question2_vec]
#
# data['skew_diff'] = np.abs(data['skew_q1'] - data['skew_q2'])
# data['kur_diff'] = np.abs(data['kur_q1'] - data['kur_q2'])
for i in range(0, len(results), 2):
v1 = results[i]
v2 = results[i + 1]
if (v1 == np.zeros_like(v1)).all() or (v2 == np.zeros_like(v2)).all():
print(i)
pred.append(1 - cosine(v1, v2))
if math.isnan(pred[-1]):
pred[-1] = 0.5
# pr = pd.DataFrame(pred)
# train = df.iloc[idx_for_split:]
# test = df.iloc[:idx_for_split]
# train_label = labels[idx_for_split:]
# test_label = labels[:idx_for_split]
roc_auc= roc_auc_score(labels, pred)
# clf = catboost.CatBoostClassifier(depth=6, iterations=5000, learning_rate=0.1, thread_count=16)
# clf.fit(train, train_label)
# y_proba = clf.predict_proba(test)[:, 1]
# roc_auc = roc_auc_score(test_label, y_proba)
with open("results_quora" + args.model_type + ".txt", "at") as f_out:
# f_out.write("robust,%.2f,%.3f\n" % (args.noise_level, mean_squared_error(true, pred)))
f_out.write(args.mode + ",%.2f,%.3f\n" % (args.noise_level, roc_auc))
def generator_loss(generator, discriminator, criterion, num_images, z_dim):
noise = noise_generator(num_images, z_dim)
gen_out = generator(noise)
# discriminator output on fake images generated by generator
gen_fake_out = discriminator(gen_out)
# calculating loss on fake images as generator wants discrimiator to treat theses images as real(1)
gen_fake_loss = criterion(gen_fake_out, torch.ones_like(gen_fake_out))
return gen_fake_loss
def get_w2v_results(args, pairs, true):
pred = []
w2v = Word2Vec.load(args.word2vec_model)
def get_mean_vec(phrase):
tokens = word_tokenize(phrase)
vectors = [np.zeros((w2v.vector_size,))]
for token in tokens:
if token in w2v:
vector = w2v[token]
vectors.append(vector)
return np.mean(vectors, axis=0)
for index, pair in pairs.iterrows():
v1 = get_mean_vec(noise_generator(pair["sentence1"], args.noise_level, chars))
v2 = get_mean_vec(noise_generator(pair["sentence2"], args.noise_level, chars))
pred.append(1 - cosine(v1, v2))
with open("results" + args.mode + ".txt", "at") as f_out:
# f_out.write("word2vec,%.2f,%.3f\n" % (args.noise_level, mean_squared_error(true, pred)))
f_out.write("word2vec,%.2f,%.3f\n" % (args.noise_level, roc_auc_score(true, pred)))
def discriminator_loss(generator, discriminator, criterion, real, num_images,
z_dim):
noise = noise_generator(num_images, z_dim)
gen_out = generator(noise)
# discriminator output on fake images generated by generator
disc_fake_out = discriminator(gen_out.detach())
# calculating loss on fake images as discrimiator wants theses images to be fake(0)
disc_fake_loss = criterion(disc_fake_out, torch.zeros_like(disc_fake_out))
#discriminator output on real images
disc_real_out = discriminator(real)
# calculating loss on real images as discrimiator wants theses images to be real(1)
disc_real_loss = criterion(disc_real_out, torch.ones_like(disc_real_out))
#averaging both loss
disc_loss = (disc_real_loss + disc_fake_loss) / 2
return disc_loss
if "word2vec" in args.mode:
pred = []
w2v = Word2Vec.load_word2vec_format(args.word2vec_model, binary=True)
def get_mean_vec(phrase):
tokens = word_tokenize(phrase)
vectors = [np.zeros((w2v.vector_size,))]
for token in tokens:
if token in w2v:
vector = w2v[token]
vectors.append(vector)
return np.mean(vectors, axis=0)
for pair in tqdm(pairs):
v1 = get_mean_vec(noise_generator(pair["text_1"], args.noise_level, chars))
v2 = get_mean_vec(noise_generator(pair["text_2"], args.noise_level, chars))
pred.append(1 - cosine(v1, v2))
with open("results_w2v_MRPC.txt", "at") as f_out:
f_out.write("word2vec,%.2f,%.3f\n" % (args.noise_level, roc_auc_score(true, pred)))
if "robust" in args.mode:
pred = []
phrases = []
for pair in pairs:
phrases.append(noise_generator(pair["text_1"], args.noise_level, chars))
phrases.append(noise_generator(pair["text_2"], args.noise_level, chars))
from sample import sample_multi
results = np.vsplit(sample_multi(args.save_dir, phrases, args.model_type), len(phrases))
for i in range(0, len(results), 2):
v1 = results[i]
def train(n_epochs=200, batch_size=128, lr=0.00001, z_dim=64, hidden_dim=128):
cur_step = 0
display_step = 500
criterion = nn.BCEWithLogitsLoss()
mean_discriminator_loss = 0
mean_generator_loss = 0
# generator model with optimizer
generator = Generator(z_dim=z_dim, img_dim=784,
hidden_dim=hidden_dim).to(device)
generator_opt = torch.optim.Adam(generator.parameters(), lr=lr)
# discriminator model with optimizer
discriminator = Discriminator(img_dim=784,
hidden_dim=hidden_dim).to(device)
discriminator_opt = torch.optim.Adam(discriminator.parameters(), lr=lr)
dataloader = get_dataloader(batch_size)
for epoch in range(n_epochs):
for real, _ in dataloader:
cur_batch_size = len(real)
# Flatten the batch of real images from the dataset
real = real.view(cur_batch_size, -1).to(device)
# Zero out the gradients before backpropagation
discriminator_opt.zero_grad()
# Calculate discriminator loss
disc_loss = discriminator_loss(generator, discriminator, criterion,
real, cur_batch_size, z_dim)
# Update gradients
disc_loss.backward(retain_graph=True)
# Update optimizer
discriminator_opt.step()
# Zero out the gradients before backpropagation
generator_opt.zero_grad()
# Calculate generator loss
gen_loss = generator_loss(generator, discriminator, criterion,
cur_batch_size, z_dim)
# Update gradients
gen_loss.backward(retain_graph=True)
# Update optimizer
generator_opt.step()
# Keep track of the average discriminator loss
mean_discriminator_loss += disc_loss.item() / display_step
# Keep track of the average generator loss
mean_generator_loss += gen_loss.item() / display_step
if cur_step % display_step == 0 and cur_step > 0:
print(
f"Epoch {epoch}, step {cur_step}: Generator loss: {mean_generator_loss}, discriminator loss: {mean_discriminator_loss}"
)
fake_noise = noise_generator(cur_batch_size, z_dim)
fake = generator(fake_noise)
generated_image_grid = show_tensor_images(
fake, tensorboard_writer=True)
tensorboard_writer(generated_image_grid,
epoch,
cur_step,
gen_image=True)
real_image_grid = show_tensor_images(real,
tensorboard_writer=True)
tensorboard_writer(real_image_grid,
epoch,
cur_step,
gen_image=False)
mean_generator_loss = 0
mean_discriminator_loss = 0
cur_step += 1