def test_linsoftmax_test(self):
"""
Test function of linear layer in eval mode (softmax)
"""
lin = layers.LinSoftmax(20, 10).eval()
o = lin(torch.randn(1, 20, 12, 24))
self.assertGreaterEqual(o[0].min(), 0)
def test_linsoftmax_aug(self):
"""
Test basic function of linear layer with 1-augmentation.
"""
lin = layers.LinSoftmax(20, 10, True)
o = lin(torch.randn(1, 20, 12, 24))
self.assertEqual(o[0].shape, (1, 10, 12, 24))
def test_linsoftmax(self):
"""
Test basic function of linear layer.
"""
lin = layers.LinSoftmax(20, 10)
o = lin(torch.randn(1, 20, 12, 24))
self.assertEqual(o[0].shape, (1, 10, 12, 24))
def test_linsoftmax_train(self):
"""
Test function of linear layer in training mode (log_softmax)
"""
lin = layers.LinSoftmax(20, 10).train()
o = lin(torch.randn(1, 20, 12, 24))
self.assertLess(o[0].max(), 0)
def build_output(
self, input: Tuple[int, int, int, int], block: str
) -> Union[Tuple[None, None, None], Tuple[Tuple[int, int, int, int], str,
Callable]]:
"""
Builds an output layer.
"""
pattern = re.compile(
r'(O)(?P<name>{\w+})?(?P<dim>2|1|0)(?P<type>l|s|c)(?P<aug>a)?(?P<out>\d+)'
)
m = pattern.match(block)
if not m:
return None, None, None
if int(m.group('dim')) != 1:
raise ValueError('non-2d output not supported, yet')
nl = m.group('type')
if nl not in ['s', 'c']:
raise ValueError('only softmax and ctc supported in output')
if nl == 'c':
self.criterion = nn.CTCLoss(reduction='none')
aug = True if m.group('aug') else False
lin = layers.LinSoftmax(input[1], int(m.group('out')), aug)
logger.debug('{}\t\tlinear\taugmented {} out {}'.format(
self.idx + 1, aug, m.group('out')))
return lin.get_shape(input), self.get_layer_name(
m.group(1), m.group('name')), lin
def test_linsoftmax_aug_test(self):
"""
Test function of linear layer in eval mode (softmax) with 1-augmentation
"""
lin = layers.LinSoftmax(20, 10, True).eval()
o = lin(torch.randn(1, 20, 12, 24))
self.assertGreaterEqual(o.min(), 0)
def test_linsoftmax_aug_train(self):
"""
Test function of linear layer in training mode (log_softmax) with 1-augmentation
"""
lin = layers.LinSoftmax(20, 10, True).train()
o = lin(torch.randn(1, 20, 12, 24))
self.assertLess(o.max(), 0)
def test_linsoftmax_resize_remove(self):
"""
Tests resizing of a fully connected layer.
"""
lin = layers.LinSoftmax(20, 10)
w_cp = lin.lin.weight.clone()
b_cp = lin.lin.bias.clone()
lin.resize(5, (1, 5, 6, 7, 9))
self.assertTrue(w_cp[(0, 2, 3, 4, 8), :].eq(lin.lin.weight).all())
self.assertTrue(b_cp[(0, 2, 3, 4, 8), ].eq(lin.lin.bias).all())
def test_linsoftmax_resize_add(self):
"""
Tests resizing of a fully connected layer.
"""
lin = layers.LinSoftmax(20, 10)
w_cp = lin.lin.weight.clone()
b_cp = lin.lin.bias.clone()
lin.resize(25)
self.assertTrue(w_cp.eq(lin.lin.weight[:10, :]).all())
self.assertTrue(b_cp.eq(lin.lin.bias[:10]).all())
self.assertTrue(lin.lin.weight.shape[0] == 25)
self.assertTrue(lin.lin.bias.shape[0] == 25)
def test_linsoftmax_resize_both(self):
"""
Tests resizing of a fully connected layer.
"""
lin = layers.LinSoftmax(20, 10)
w_cp = lin.lin.weight.clone()
b_cp = lin.lin.bias.clone()
lin.resize(25, (1, 5, 6, 7, 9))
self.assertTrue(w_cp[(0, 2, 3, 4,
8), :].eq(lin.lin.weight[:5, :]).all())
self.assertTrue(b_cp[(0, 2, 3, 4, 8), ].eq(lin.lin.bias[:5]).all())
self.assertTrue(lin.lin.weight.shape[0] == 25)
self.assertTrue(lin.lin.bias.shape[0] == 25)
def build_output(
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 output layer.
"""
pattern = re.compile(
r'(O)(?P<name>{\w+})?(?P<dim>2|1|0)(?P<type>l|s|c)(?P<aug>a)?(?P<out>\d+)'
)
m = pattern.match(blocks[idx])
if not m:
return None, None, None
dim = int(m.group('dim'))
nl = m.group('type')
outdim = int(m.group('out'))
if dim == 0:
raise ValueError('categorical output not supported, yet.')
if nl == 'c' and dim == 2:
raise ValueError('CTC not supported for heatmap output')
if nl in ['l', 's'] and int(m.group('out')) >= 1:
self.criterion = nn.BCELoss()
elif nl == 'c':
self.criterion = nn.CTCLoss(reduction='sum', zero_infinity=True)
else:
raise ValueError('unsupported output specification')
# heatmap output
if dim == 2:
act = 's' if nl == 'l' else 'm'
fn = layers.ActConv2D(input[1], outdim, (1, 1), (1, 1), act)
self.idx += 1
logger.debug(
'{}\t\tconv\tkernel 1 x 1 filters {} stride 1 activation {}'.
format(self.idx, outdim, nl))
return fn.get_shape(input), [
VGSLBlock(blocks[idx], m.group('type'), m.group('name'),
self.idx)
], fn
else:
aug = True if m.group('aug') else False
lin = layers.LinSoftmax(input[1], int(m.group('out')), aug)
self.idx += 1
logger.debug('{}\t\tlinear\taugmented {} out {}'.format(
self.idx, aug, m.group('out')))
return lin.get_shape(input), [
VGSLBlock(blocks[idx], m.group(1), m.group('name'), self.idx)
], lin
