import numpy as np
import os, sys
import review_proc as rp
import matplotlib.pyplot as plt
init_prob = []
final_prob = []
for filename in os.listdir('./probs/'):
review_number = filename[0:-4]
rv = rp.review('./aclImdb/test/posneg/'+review_number+'.txt')
final_prob_scalar = np.load('./probs/'+review_number+'.npy')
delta_matrix = np.load('./diffs/actual_diff_'+review_number+'.npy')
if rv.sentiment == 'pos':
m = np.amin(delta_matrix,axis=0)
delta = np.amin(m)
i = np.argmin(m)
elif rv.sentiment == 'neg':
m = np.amax(delta_matrix,axis=0)
delta = np.amax(delta_matrix)
i = np.argmax(m)
#if final_prob_scalar < 0.5 and rv.sentiment == 'pos':
# print(i)
if final_prob_scalar > 0.5 and rv.sentiment == 'neg':
print(m)
final_prob.append(final_prob_scalar)
init_prob.append(final_prob_scalar-delta)
#plt.hist(final_prob,bins=50)
#plt.show()
word_embedding = np.load(word_embedding_filename)
word_to_embedding_index = np.load(word_to_embedding_index_filename).item()
except FileNotFoundError:
print('Word embedding not found, running word2vec')
word2vec.w2v(corpus_filename='./corpus/imdb_train_corpus.txt')
embedding_norm = np.linalg.norm(word_embedding, axis=1)
embedding_norm.shape = (10000, 1)
normalized_word_embedding = word_embedding / embedding_norm
m = word_to_embedding_index
# Reverse dictionary to look up words from indices
embedding_index_to_word = dict(zip(m.values(), m.keys()))
root = './aclImdb/test/posneg/'
for filename in os.listdir('./ggs_results/diffs/'):
rv = rp.review(root + filename[0:-4] + '.txt')
diff = np.load('./ggs_results/diffs/' + filename)
prob = np.load('./ggs_results/probs/' + filename)
print('Filename: ', filename, 'Initial Probability: ', prob[0][0])
if rv.sentiment == 'pos':
m = np.amin(diff)
else:
m = np.amax(diff)
prob_positive = prob[0, 0] + m
# Following are conditions where one-word replacement worked.
if prob_positive < 0.5 and rv.sentiment == 'pos':
continue
elif prob_positive > 0.5 and rv.sentiment == 'neg':
continue
g = tf.Graph()
tf.reset_default_graph()
embedding_norm = np.linalg.norm(word_embedding, axis=1)
embedding_norm.shape = (10000, 1)
normalized_word_embedding = word_embedding / embedding_norm
m = word_to_embedding_index
# Reverse dictionary to look up words from indices
embedding_index_to_word = dict(zip(m.values(), m.keys()))
c = [0, 0, 0]
t = [0, 0, 0]
for it, divs in enumerate([2, 4, 8]):
res_dir = './minmax/' + str(divs) + '/ii/'
for file_name in os.listdir(res_dir):
ii = np.load(res_dir + file_name)
jj = np.load(res_dir[:-3] + 'jj/' + file_name)
g = tf.Graph()
rv = rp.review('./aclImdb/test/posneg/' + file_name[:-4] + '.txt')
w = [embedding_index_to_word[i] for i in ii]
for n, j in enumerate(jj):
rv.tokens[j] = w[n]
with g.as_default():
global_step_tensor = \
tf.Variable(0, trainable = False, name = 'global_step')
# Create RNN graph
r = rnn.classifier(batch_size=1,
learning_rate=0.0,
hidden_size=16,
max_time=rv.length,
embeddings=word_embedding,
global_step=global_step_tensor)
with tf.Session() as sess:
tf.train.Saver().restore(
try:
word_embedding = np.load(word_embedding_filename)
word_to_embedding_index = np.load(word_to_embedding_index_filename).item()
except FileNotFoundError:
print('Word embedding not found, running word2vec')
word2vec.w2v(corpus_filename = './corpus/imdb_train_corpus.txt')
embedding_norm = np.linalg.norm(word_embedding,axis=1)
embedding_norm.shape = (10000,1)
normalized_word_embedding = word_embedding / embedding_norm
m = word_to_embedding_index
# Reverse dictionary to look up words from indices
embedding_index_to_word = dict(zip(m.values(), m.keys()))
g = tf.Graph()
rv = rp.review('./aclImdb/test/posneg/5164_10.txt')
#ii = [242,242,837,348]
#jj = [11,215,467,680]
#ii = [369,369,837]
#jj = [72,292,495]
ii = [882,348]
jj = [165,495]
w = [embedding_index_to_word[i] for i in ii]
print(w)
for n,j in enumerate(jj):
rv.tokens.insert(j,w[n])
rv.length = rv.length + len(ii)
with g.as_default():
global_step_tensor = tf.Variable(0, trainable = False, name = 'global_step')
# Create RNN graph
t = []
f = 0
root_dir = './aclImdb/test/posneg/'
restore_name = '4786_9.txt'
flag = True
for file_name in os.listdir(root_dir):
if file_name != restore_name and flag:
print(file_name)
continue
elif flag:
flag = False
continue
word_count = 0
g = tf.Graph()
print('Running attack on: ' + file_name)
rvo = rp.review(root_dir + file_name)
rvo.translate(rvo.length, word_to_embedding_index, embedding_index_to_word)
rvo.vec(word_embedding)
# Actual Neural Network
with g.as_default():
global_step_tensor = tf.Variable(0,
trainable=False,
name='global_step')
r = rnn.classifier(batch_size=1,
learning_rate=0.0,
hidden_size=8,
max_time=rvo.length,
embeddings=word_embedding,
global_step=global_step_tensor)
with tf.Session() as sess:
restore_name = './ckpts/gridckpt_8_10/imdb-rnn-e15.ckpt'
try:
word_embedding = np.load(word_embedding_filename)
word_to_embedding_index = np.load(word_to_embedding_index_filename).item()
except FileNotFoundError:
print('Word embedding not found, running word2vec')
word2vec.w2v(corpus_filename='./corpus/imdb_train_corpus.txt')
embedding_norm = np.linalg.norm(word_embedding, axis=1)
embedding_norm.shape = (10000, 1)
normalized_word_embedding = word_embedding / embedding_norm
m = word_to_embedding_index
# Reverse dictionary to look up words from indices
embedding_index_to_word = dict(zip(m.values(), m.keys()))
g = tf.Graph()
rv = rp.review('./aclImdb/test/posneg/9999_10.txt')
with g.as_default():
global_step_tensor = tf.Variable(0, trainable=False, name='global_step')
# Create RNN graph
r = rnn.classifier(batch_size=1,
learning_rate=0.0,
hidden_size=16,
max_time=1024,
embeddings=word_embedding,
global_step=global_step_tensor)
with tf.Session() as sess:
tf.train.Saver().restore(sess,
'./ckpts/gridckpt_16_10/imdb-rnn-e15.ckpt')
print(rv.tokens)
rv.translate(r.max_time, word_to_embedding_index,
embedding_index_to_word)
test_file_list = os.listdir('./aclImdb/test/posneg/')
with g.as_default():
global_step_tensor = tf.Variable(0, trainable=False, name='global_step')
# Create RNN graph
r = rnn.classifier(batch_size=num_top * per_batch,
learning_rate=0.0,
hidden_size=16,
max_time=1024,
embeddings=word_embedding,
global_step=global_step_tensor)
with tf.Session() as sess:
tf.train.Saver().restore(sess,
'./ckpts/gridckpt_16_10/imdb-rnn-e15.ckpt')
for test_file in test_file_list:
print('Processing ' + test_file)
rv = rp.review('./aclImdb/test/posneg/' + test_file)
rv.translate(r.max_time, word_to_embedding_index,
embedding_index_to_word)
rv.vec(word_embedding)
decision, probability, batch_grad = \
r.infer_batched_prob(sess,rv,rv.index_vector[0][0],50,list(range(10)))
rnn_sentiment = 'pos' if not decision[0] else 'neg'
print('Neural Net Decision: ', rnn_sentiment, ' Actual: ',
rv.sentiment)
if rnn_sentiment != rv.sentiment:
continue
grad = batch_grad[0][0, 0:rv.length, :]
W = word_embedding
G = grad
D = W @ (G.T)
c = np.sum(np.multiply(rv.vector_list, G), axis=1)
global_step_tensor = tf.Variable(0, trainable=False, name='global_step')
# Create RNN graph
r = rnn.classifier(batch_size=1,
learning_rate=0.0,
hidden_size=16,
max_time=1024,
embeddings=word_embedding,
global_step=global_step_tensor)
succ = 0
n = 0
with tf.Session() as sess:
tf.train.Saver().restore(sess,
'./ckpts/gridckpt_16_10/imdb-rnn-e15.ckpt')
for test_file in test_file_list:
review_name = test_file[12:-4] + '.txt'
rv = rp.review('./aclImdb/test/posneg/' + review_name)
rv.translate(r.max_time, word_to_embedding_index,
embedding_index_to_word)
rv.vec(word_embedding)
decision, probability, batch_grad = r.infer_dpg(sess, rv)
rnn_sentiment = 'pos' if not decision[0] else 'neg'
if rnn_sentiment != rv.sentiment:
continue
grad = batch_grad[0][0, 0:rv.length, :]
W = word_embedding
G = grad
D = W @ (G.T)
c = np.sum(np.multiply(rv.vector_list, G), axis=1)
d = D - c
x = np.load('./diffs/' + test_file)
x.shape = (100000, 1)
import numpy as np
import matplotlib.pyplot as plt
import os, sys, re
import review_proc as rp
V = 10000
num_top = 10
root = './ggs_results/'
for filename in os.listdir(root + 'diffs/'):
rv = rp.review('aclImdb/test/posneg/' + filename[0:-4] + '.txt')
grad = np.load(root + 'grads/' + filename)
diff = np.load(root + 'diffs/' + filename)
n = np.linalg.norm(grad, axis=1)
i = np.argsort(n)[-num_top:]
for index in i:
top_word = rv.tokens[index]
print(top_word)
print('_' * 20)
#n = n[i]; n.shape = (10,1)
#n = np.tile(n.T,(V,1))
