def test_d3_3_sample():
global c2i, i2c
model_fname = "deliverable_3.2.mod"
# does the file exist?
assert_true(os.path.exists(model_fname))
# try loading saved model
trained_model = torch.load(model_fname)
# try making a new town:
new_town = lm.sample(trained_model, c2i, i2c)
# verify that it's not too long:
assert_less(len(new_town), 200)
# make sure that it doesn't start with BOS_SYM
assert_false(new_town[0] == vocab.BOS_SYM)
# verify that its probability is lower than gibberish:
gibberish_string = "asdflkjiopqutepoiuqrfm"
p_gibberish = lm.compute_prob(trained_model, gibberish_string,
c2i) / len(gibberish_string)
p_good = lm.compute_prob(trained_model, new_town, c2i) / len(new_town)
# negative log probs, so bigger is smaller
assert_greater(p_gibberish, p_good)
def test_d3_2_training():
global corpus, c2i
model_fname = "deliverable_3.2.mod"
# does the file exist?
assert_true(os.path.exists(model_fname))
# try loading saved model
trained_model = torch.load(model_fname)
# was it made correctly (as per the assignment's instructions)?
eq_(trained_model.input_lookup.num_embeddings, len(c2i))
eq_(trained_model.input_lookup.embedding_dim, 25)
eq_(trained_model.lstm.num_layers, 1)
eq_(trained_model.lstm.hidden_size, 50)
# ask it for gibberish prob, make sure that's lower than actual town name prob
gibberish_string = "asdflkjiopqutepoiuqrfm"
good_fake_string = "Little Brockton-upon-Thyme"
p_gibberish = lm.compute_prob(trained_model, gibberish_string, c2i)
p_good = lm.compute_prob(trained_model, good_fake_string, c2i)
p_gibberish_norm = p_gibberish / len(gibberish_string)
p_good_norm = p_good / len(good_fake_string)
# p_gibberish should be larger than p_good, since we're in negative log space
assert_greater(p_gibberish_norm, p_good_norm)