def plot_log_p(filename, codec, model):
from questions.likelihood import log_likelihood
with open(os.path.join('datasets', filename + '.pkl'), 'rb') as f:
lls = []
data = pkl.load(f)
for i in trange(len(data)):
text = data[i]
text = codec.encode(text).to(device)
## TODO: complete the code in the function `log_likelihood`
lls.append(log_likelihood(model, text))
lls = np.asarray(lls)
with open(os.path.join('submit', filename + '_raw.pkl'), 'wb') as f:
pkl.dump(lls, f, protocol=pkl.HIGHEST_PROTOCOL)
plt.figure()
plt.hist(lls)
plt.xlabel('Log-likelihood')
plt.xlim([-600, 0])
plt.ylabel('Counts')
plt.title(filename)
plt.savefig(os.path.join('submit', filename + '.png'), bbox_inches='tight')
plt.show()
plt.close()
print("# Figure written to %s.png." % filename)
def classification(model, text):
"""
Classify whether the string `text` is randomly generated or not.
:param model: The GPT-2 model
:param texts: A tensor of shape (1, T), where T is the length of the text
:return: True if `text` is a random string. Otherwise return False
"""
with torch.no_grad():
## TODO: Return True if `text` is a random string. Or else return False.
ll = log_likelihood(model, text)
return ll < -300