random.seed(31415)
np.random.seed(9265)
word_vectors = np.concatenate((
(np.random.rand(num_words, dim_vectors) - .5) / dim_vectors,
np.zeros((num_words, dim_vectors))),
axis=0
)
params['sgd']['step'] = 0.2
params['sgd']['iterations'] = 40000
params['sgd']['tolerance'] = 1e-48
params['sgd']['anneal_every'] = 20000
params['sgd']['anneal_factor'] = 0.5
word_vectors0 = sgd(
lambda vec: word2vec_sgd_wrapper(skipgram, tokens, vec, dataset, params, neg_sampling_cost_and_gradient),
word_vectors, params, postprocessing=normalize_rows, use_saved=True, print_every=100, save_params_every=5000)
# sanity check: cost at convergence should be around or below 10
# sum the input and output word vectors
word_vectors = (word_vectors0[:num_words,:] + word_vectors0[num_words:,:])
print "\n=== For autograder ==="
check_words = ["the", "a", "an", "movie", "ordinary", "but", "and"]
check_idx = [tokens[word] for word in check_words]
check_vecs = word_vectors[check_idx, :]
print check_vecs
# Visualize the word vectors you trained
dim_vectors = word_vectors.shape[1]
dataset = StanfordSentiment()
train_features, train_labels, words = get_data(dataset, word_vectors, dtype='train')
weights = np.random.randn(dim_vectors, 5) # D x NUM_LABELS array
# We will do batch optimization
params = AttrDict({
'sgd' : {'batch_size': 50, 'step': 3.0, 'iterations': iterations, 'tolerance': 0,
'anneal_every': 10000, 'anneal_factor': 0.5},
'dataset' : {}
})
print "Starting SGD..."
weights = sgd(lambda weights: softmax_wrapper(train_features, train_labels, weights, regularization),
weights, params, postprocessing=None, use_saved=False, print_every=500, save_params_every=1000)
_, _, pred = softmax_regression(train_features, train_labels, weights)
print "Train precision (%%): %f" % precision(train_labels, pred)
save_data(words, train_labels, pred, 'data_train.txt')
print "Testing on dev dataset"
dev_features, dev_labels, dev_words = get_data(dataset, word_vectors, dtype='dev')
print dev_features.shape, weights.shape
_, _, pred = softmax_regression(dev_features, dev_labels, weights)
print "Dev precision (%%): %f" % precision(dev_labels, pred)
save_data(dev_words, dev_labels, pred, 'data_dev.txt')
test_features, test_labels, test_words = get_data(dataset, word_vectors, dtype='test')
# Context size
context_size = 5
print "Training word vectors"
# Reset the random seed to make sure that everyone gets the same results
random.seed(31415)
np.random.seed(9265)
word_vectors = np.concatenate((
(np.random.rand(num_words, dim_vectors) - 0.5) / dim_vectors,
np.zeros((num_words, dim_vectors))),
axis=0
)
word_vectors0 = sgd(
lambda vec: word2vec_sgd_wrapper(skipgram, tokens, vec, dataset, C, neg_sampling_cost_and_gradient),
word_vectors, 0.3, 40000, posprocessing=normalize_rows, use_saved=True, print_every=10, tolerance=1e-8)
# sanity check: cost at convergence should be around or below 10
# sum the input and output word vectors
word_vectors = (word_vectors0[:num_words,:] + word_vectors0[num_words:,:])
print "\n=== For autograder ==="
check_words = ["the", "a", "an", "movie", "ordinary", "but", "and"]
checkIdx = [tokens[word] for word in check_words]
checkVecs = word_vectors[checkIdx, :]
print checkVecs
# In[ ]:
