def _sample(self, lines, biases=None, styles=None):
num_samples = len(lines)
max_tsteps = 40 * max([len(i) for i in lines])
biases = biases if biases is not None else [0.5] * num_samples
x_prime = np.zeros([num_samples, 1600, 3])
x_prime_len = np.zeros([num_samples])
chars = np.zeros([num_samples, 120])
chars_len = np.zeros([num_samples])
if styles is not None:
for i, (cs, style) in enumerate(zip(lines, styles)):
if isinstance(style, dict):
x_p = style['stroke']
c_p = style['text']
else:
x_p = np.load('styles/style-{}-strokes.npy'.format(style))
c_p = np.load('styles/style-{}-chars.npy'.format(
style)).tostring().decode('utf-8')
c_p = str(c_p) + " " + cs
c_p = drawing.encode_ascii(c_p)
c_p = np.array(c_p)
x_prime[i, :len(x_p), :] = x_p
x_prime_len[i] = len(x_p)
chars[i, :len(c_p)] = c_p
chars_len[i] = len(c_p)
else:
for i in range(num_samples):
encoded = drawing.encode_ascii(lines[i])
chars[i, :len(encoded)] = encoded
chars_len[i] = len(encoded)
[samples] = self.nn.session.run(
[self.nn.sampled_sequence],
feed_dict={
self.nn.prime: styles is not None,
self.nn.x_prime: x_prime,
self.nn.x_prime_len: x_prime_len,
self.nn.num_samples: num_samples,
self.nn.sample_tsteps: max_tsteps,
self.nn.c: chars,
self.nn.c_len: chars_len,
self.nn.bias: biases
})
samples = [
sample[~np.all(sample == 0.0, axis=1)] for sample in samples
]
return samples
def process_chars(chars):
chars = chars.strip()
if any(c not in drawing.alphabet
for c in chars) or len(chars) > MAX_CHAR_LEN:
return None
chars = drawing.encode_ascii(chars)[:MAX_CHAR_LEN]
return chars
def collect_data(fnames):
# fnames = []
# for dirpath, dirnames, filenames in os.walk('data/raw_npz/'):
# if dirnames:
# continue
# for filename in filenames:
# if filename.startswith('.'):
# continue
# if not filename.endswith('.npz'):
# continue
# print(os.path.join(dirpath,filename))
# fnames.append(os.path.join(dirpath, filename))
x_out = []
c_out = []
for i, fname in enumerate(fnames):
print('loading \'' + fname + '\'...')
dataset = np.load(fname, allow_pickle=True)
assert (len(dataset['samples']) == len(dataset['texts']))
for strokes, text in zip(tqdm(dataset['samples'], file=sys.stdout),
dataset['texts']):
x_out.append(strokes)
c_out.append(
drawing.encode_ascii(text.strip())[:drawing.MAX_CHAR_LEN])
return x_out, c_out
def _sample(self, line, x_p, c_p, bias=None):
max_tsteps = 40 * len(line)
biases = [bias] if bias is not None else [0.5]
x_prime = np.zeros([1, 2000, 3])
x_prime_len = np.zeros([1])
chars = np.zeros([1, 120])
chars_len = np.zeros([1])
c_p = c_p + " " + line
c_p = drawing.encode_ascii(c_p)
c_p = np.array(c_p)
x_prime[0, :len(x_p), :] = x_p
x_prime_len[0] = len(x_p)
chars[0, :len(c_p)] = c_p
chars_len[0] = len(c_p)
[samples] = self.nn.session.run(
[self.nn.sampled_sequence],
feed_dict={
self.nn.prime: True,
self.nn.x_prime: x_prime,
self.nn.x_prime_len: x_prime_len,
self.nn.num_samples: 1,
self.nn.sample_tsteps: max_tsteps,
self.nn.c: chars,
self.nn.c_len: chars_len,
self.nn.bias: biases
}
)
return samples[0]
def get_ascii_sequences(filename):
sequences = open(filename, 'r').read()
sequences = sequences.replace(r'%%%%%%%%%%%', '\n')
sequences = [i.strip() for i in sequences.split('\n')]
lines = sequences[sequences.index('CSR:') + 2:]
lines = [line.strip() for line in lines if line.strip()]
lines = [drawing.encode_ascii(line)[:drawing.MAX_CHAR_LEN] for line in lines]
return lines
def _predict_strokes(self, lines_v):
num_samples = len(lines_v)
max_tsteps = 40*max([len(i) for i in lines_v])
biases = [self.bias]*num_samples if self.bias is not None else [0.5]*num_samples
x_prime = np.zeros([num_samples, 1200, 3])
x_prime_len = np.zeros([num_samples])
chars = np.zeros([num_samples, 120])
chars_len = np.zeros([num_samples])
if self.style is not None:
for i, cs in enumerate(lines_v):
x_p = np.load(os.path.join(self.styles_dir, 'style-{}-strokes.npy'.format(self.style)))
c_p = np.load(os.path.join(self.styles_dir, 'style-{}-chars.npy'.format(self.style))).tostring().decode('utf-8')
c_p = str(c_p) + " " + cs
c_p = drawing.encode_ascii(c_p)
c_p = np.array(c_p)
x_prime[i, :len(x_p), :] = x_p
x_prime_len[i] = len(x_p)
chars[i, :len(c_p)] = c_p
chars_len[i] = len(c_p)
else:
for i in range(num_samples):
encoded = drawing.encode_ascii(lines_v[i])
chars[i, :len(encoded)] = encoded
chars_len[i] = len(encoded)
[samples] = self.nn.session.run(
[self.nn.sampled_sequence],
feed_dict={
self.nn.prime: self.style is not None,
self.nn.x_prime: x_prime,
self.nn.x_prime_len: x_prime_len,
self.nn.num_samples: num_samples,
self.nn.sample_tsteps: max_tsteps,
self.nn.c: chars,
self.nn.c_len: chars_len,
self.nn.bias: biases
}
)
strokes_v = [sample[~np.all(sample == 0.0, axis=1)] for sample in samples]
return strokes_v
def get_ascii_sequences(filename):
sequences = open(filename, 'r').read()
sequences = sequences.replace(r'%%%%%%%%%%%', '\n')
sequences = [i.strip() for i in sequences.split('\n')]
lines = sequences[sequences.index('CSR:') + 2:]
lines = [line.strip() for line in lines if line.strip()]
lines = [drawing.encode_ascii(line)[:drawing.MAX_CHAR_LEN] for line in lines]
return lines
def _sample(self, lines, biases=None, styles=None):
num_samples = len(lines)
max_tsteps = 40*max([len(i) for i in lines])
biases = biases if biases is not None else [0.5]*num_samples
x_prime = np.zeros([num_samples, 1200, 3])
x_prime_len = np.zeros([num_samples])
chars = np.zeros([num_samples, 120])
chars_len = np.zeros([num_samples])
if styles is not None:
for i, (cs, style) in enumerate(zip(lines, styles)):
x_p = np.load('styles/style-{}-strokes.npy'.format(style))
c_p = np.load('styles/style-{}-chars.npy'.format(style))
c_p = str(c_p) + " " + cs
c_p = drawing.encode_ascii(c_p)
c_p = np.array(c_p)
x_prime[i, :len(x_p), :] = x_p
x_prime_len[i] = len(x_p)
chars[i, :len(c_p)] = c_p
chars_len[i] = len(c_p)
else:
for i in range(num_samples):
encoded = drawing.encode_ascii(lines[i])
chars[i, :len(encoded)] = encoded
chars_len[i] = len(encoded)
[samples] = self.nn.session.run(
[self.nn.sampled_sequence],
feed_dict={
self.nn.prime: styles is not None,
self.nn.x_prime: x_prime,
self.nn.x_prime_len: x_prime_len,
self.nn.num_samples: num_samples,
self.nn.sample_tsteps: max_tsteps,
self.nn.c: chars,
self.nn.c_len: chars_len,
self.nn.bias: biases
}
)
samples = [sample[~np.all(sample == 0.0, axis=1)] for sample in samples]
return samples
def collect_data():
fnames = []
for dirpath, dirnames, filenames in os.walk('data/raw_deepwriting/'):
if dirnames:
continue
for filename in filenames:
if filename.startswith('.'):
continue
if not filename.endswith('.npz'):
continue
fnames.append(os.path.join(dirpath, filename))
x_out = []
c_out = []
for i, fname in enumerate(fnames):
print(i, fname)
dataset = HandWritingDatasetConditional(fname)
assert (len(dataset.samples) == len(dataset.texts))
for i2, (strokes, text, line_onehot, eoc_labels,
bow_labels) in enumerate(
zip(dataset.samples, dataset.texts,
dataset.one_hot_char_labels, dataset.eoc_labels,
dataset.bow_labels)):
if i2 % 200 == 0:
print(i2, '\t', '/', len(dataset.samples))
x_out.append(dataset.undo_preprocess(strokes))
line_int = dataset.one_hot_to_int_labels(line_onehot)
line_chars = int_labels_to_text(line_int, eoc_labels, bow_labels,
dataset)
#line_chars += " | " + text
c_out.append(
drawing.encode_ascii(
line_chars.strip())[:drawing.MAX_CHAR_LEN])
return x_out, c_out
