nesl / nlp_adversarial_examples Goto Github PK

hey, can you please tell where have you applied the max changes constraint in genetic attack ?
I do see a parameter max_changes=0.4 in attack function but cant see it being used anywhere.
Thanks

In attack method, the iteration index "i" is used when generating "childs", which I believe may cause some iteration bugs...
for i in range(self.max_iters): # print(i) pop_preds = self.batch_model.predict(self.sess, np.array(pop)) pop_scores = pop_preds[:, target] print('\t\t', i, ' -- ', np.max(pop_scores)) pop_ranks = np.argsort(pop_scores)[::-1] top_attack = pop_ranks[0] logits = np.exp(pop_scores / self.temp) select_probs = logits / np.sum(logits) if np.argmax(pop_preds[top_attack, :]) == target: return pop[top_attack] elite = [pop[top_attack]] # elite #print(select_probs.shape) parent1_idx = np.random.choice(self.pop_size, size=self.pop_size-1, p=select_probs) parent2_idx = np.random.choice(self.pop_size, size=self.pop_size-1, p=select_probs) childs = [self.crossover(pop[parent1_idx[i]], pop[parent2_idx[i]]) for i in range(self.pop_size-1)] childs = [self.perturb(x, x_orig, neigbhours_list, neighbours_dist, w_select_probs, target) for x in childs] pop = elite + childs

The size of embedding_matrix in 30050001 due to which there is a memory error while doing -2np.dot(embedding_matrix.T , embedding_matrix).
How do I fix this?

Hi!
When I try to run download_counterfitted_vectors.sh, it shows a error (403 forbidden) and I cannot download the counter-filterd vectors.
Can you check this out?

Hi,

I keep getting a segmentation fault when I run compute_dist_mat.py:

$ python compute_dist_mat.py 
Using TensorFlow backend.
Segmentation fault

All the previous steps in the README seem to complete without error. I've tried this on two different computers with python 2.7 and tensorflow 1.4 and 1.9, respectively. Any ideas on how to get around this? Or if someone could share the output of this script, that would also be appreciated. Thanks!

Hi there,

I am highly interested in your fantastic work. However, I am a little bit confused about the baseline attack you used in your code and your paper. In BaselineDemo.ipynb, what is the NewBaseline attack that you imported? I couldn't find it in Attacks.py but PerturbSentBaseline(). Also, you mentioned in your paper that you use a baseline method to greedily find adversarial texts, could you please be more specific on the baseline attack you use? Thank you so much.

Where can I find the code for adversarial attacks on SNLI model?