def indices_to_data(indices, seqlength, num_chars):
num_examples = len(indices) - seqlength
x = np.zeros([num_examples, seqlength], dtype='float32')
y = np.zeros([num_examples], dtype='int32')
for example_num in range(0, len(indices) - seqlength):
start = example_num
end = start + seqlength
x[example_num, :] = indices[start:end]
y[example_num] = indices[end]
return to_one_hot(x, num_chars), to_one_hot(y, num_chars)
def gen_sequence(model,
seqlength,
char_to_index,
index_to_char,
seed='The quick brown fox jumps',
num_chars=100):
assert len(seed) >= seqlength
samples = []
indices = [char_to_index[c] for c in seed]
data = np.zeros((1, seqlength))
for i, index in enumerate(indices[:seqlength]):
data[0, i] = index
data = to_one_hot(data, len(char_to_index))
for i in range(num_chars):
# Pick the character that got assigned the highest probability
preds = model(data)
# ix = np.argmax(preds.ravel())
# Alternatively, to sample from the distribution instead:
ix = np.random.choice(np.arange(len(char_to_index)), p=preds.ravel())
samples.append(ix)
data[0, 0:seqlength - 1] = data[0, 1:] # bump down
data[0, seqlength - 1] = ix # insert latest
random_snippet = seed + ''.join(index_to_char[index] for index in samples)
return random_snippet
def load_labels(data_dir, filename):
path = os.path.join(data_dir, filename)
if not os.path.exists(path):
download(data_dir, filename)
with gzip.open(path, 'rb') as f:
data = np.frombuffer(f.read(), np.uint8, offset=8)
return to_one_hot(data, 10)
def load_labels(data_dir, filename):
path = os.path.join(data_dir, 'cifar-10-batches-py', filename)
with open(path, 'rb') as f:
labels = np.asarray(pk.load(f, encoding='latin1')['labels'])
return to_one_hot(labels, 10)