def test_bidi_rnn_layer_y_summarize(self):
"""
Test bidirectional summarizing RNN layer in y-dimension.
"""
rnn = layers.TransposedSummarizingRNN(10, 2, 'b', True, True)
o = rnn(torch.randn(1, 10, 32, 64))
self.assertEqual(o.shape, (1, 4, 1, 64))
def test_forward_rnn_layer_x_summarize(self):
"""
Test unidirectional summarizing RNN layer in x-dimension.
"""
rnn = layers.TransposedSummarizingRNN(10, 2, 'f', False, True)
o = rnn(torch.randn(1, 10, 32, 64))
self.assertEqual(o.shape, (1, 2, 32, 1))
def test_bidi_rnn_layer_x(self):
"""
Test bidirectional RNN layer in x-dimension.
"""
rnn = layers.TransposedSummarizingRNN(10, 2, 'b', False, False)
o = rnn(torch.randn(1, 10, 32, 64))
self.assertEqual(o[0].shape, (1, 4, 32, 64))
def build_rnn(
self, input: Tuple[int, int, int, int], block: str
) -> Union[Tuple[None, None, None], Tuple[Tuple[int, int, int, int], str,
Callable]]:
"""
Builds an LSTM/GRU layer returning number of outputs and layer.
"""
pattern = re.compile(
r'(?P<type>L|G)(?P<dir>f|r|b)(?P<dim>x|y)(?P<sum>s)?(?P<legacy>c|o)?(?P<name>{\w+})?(?P<out>\d+)'
)
m = pattern.match(block)
if not m:
return None, None, None
type = m.group('type')
direction = m.group('dir')
dim = m.group('dim') == 'y'
summarize = m.group('sum') == 's'
legacy = None
if m.group('legacy') == 'c':
legacy = 'clstm'
elif m.group('legacy') == 'o':
legacy = 'ocropy'
hidden = int(m.group(7))
fn = layers.TransposedSummarizingRNN(input[1], hidden, direction, dim,
summarize, legacy)
logger.debug(
'{}\t\trnn\tdirection {} transposed {} summarize {} out {} legacy {}'
.format(self.idx + 1, direction, dim, summarize, hidden, legacy))
return fn.get_shape(input), self.get_layer_name(type,
m.group('name')), fn
def build_rnn(
self, input: Tuple[int, int, int, int], blocks: List[str], idx: int
) -> Union[Tuple[None, None, None], Tuple[Tuple[int, int, int, int], str,
Callable]]:
"""
Builds an LSTM/GRU layer returning number of outputs and layer.
"""
pattern = re.compile(
r'(?P<type>L|G)(?P<dir>f|r|b)(?P<dim>x|y)(?P<sum>s)?(?P<legacy>c|o)?(?P<name>{\w+})?(?P<out>\d+)'
)
m = pattern.match(blocks[idx])
if not m:
return None, None, None
type = m.group('type')
direction = m.group('dir')
dim = m.group('dim') == 'y'
summarize = m.group('sum') == 's'
legacy = None
if m.group('legacy') == 'c':
legacy = 'clstm'
elif m.group('legacy') == 'o':
legacy = 'ocropy'
hidden = int(m.group(7))
fn = layers.TransposedSummarizingRNN(input[1], hidden, direction, dim,
summarize, legacy)
self.idx += 1
logger.debug(
f'{self.idx}\t\trnn\tdirection {direction} transposed {dim} '
f'summarize {summarize} out {hidden} legacy {legacy}')
return fn.get_shape(input), [
VGSLBlock(blocks[idx], type, m.group('name'), self.idx)
], fn
