def test_point_head():
inputs = [torch.randn(1, 32, 45, 45)]
point_head = PointHead(
in_channels=[32], in_index=[0], channels=16, num_classes=19)
assert len(point_head.fcs) == 3
fcn_head = FCNHead(in_channels=32, channels=16, num_classes=19)
if torch.cuda.is_available():
head, inputs = to_cuda(point_head, inputs)
head, inputs = to_cuda(fcn_head, inputs)
prev_output = fcn_head(inputs)
test_cfg = ConfigDict(
subdivision_steps=2, subdivision_num_points=8196, scale_factor=2)
output = point_head.forward_test(inputs, prev_output, None, test_cfg)
assert output.shape == (1, point_head.num_classes, 180, 180)
# test multiple losses case
inputs = [torch.randn(1, 32, 45, 45)]
point_head_multiple_losses = PointHead(
in_channels=[32],
in_index=[0],
channels=16,
num_classes=19,
loss_decode=[
dict(type='CrossEntropyLoss', loss_name='loss_1'),
dict(type='CrossEntropyLoss', loss_name='loss_2')
])
assert len(point_head_multiple_losses.fcs) == 3
fcn_head_multiple_losses = FCNHead(
in_channels=32,
channels=16,
num_classes=19,
loss_decode=[
dict(type='CrossEntropyLoss', loss_name='loss_1'),
dict(type='CrossEntropyLoss', loss_name='loss_2')
])
if torch.cuda.is_available():
head, inputs = to_cuda(point_head_multiple_losses, inputs)
head, inputs = to_cuda(fcn_head_multiple_losses, inputs)
prev_output = fcn_head_multiple_losses(inputs)
test_cfg = ConfigDict(
subdivision_steps=2, subdivision_num_points=8196, scale_factor=2)
output = point_head_multiple_losses.forward_test(inputs, prev_output, None,
test_cfg)
assert output.shape == (1, point_head.num_classes, 180, 180)
fake_label = torch.ones([1, 180, 180], dtype=torch.long)
if torch.cuda.is_available():
fake_label = fake_label.cuda()
loss = point_head_multiple_losses.losses(output, fake_label)
assert 'pointloss_1' in loss
assert 'pointloss_2' in loss
def _to_container(cfg):
"""
mmdet will assert the type of dict/list.
So convert omegaconf objects to dict/list.
"""
if isinstance(cfg, DictConfig):
cfg = OmegaConf.to_container(cfg, resolve=True)
from mmcv.utils import ConfigDict
return ConfigDict(cfg)
def test_point_head():
inputs = [torch.randn(1, 32, 45, 45)]
point_head = PointHead(
in_channels=[32], in_index=[0], channels=16, num_classes=19)
assert len(point_head.fcs) == 3
fcn_head = FCNHead(in_channels=32, channels=16, num_classes=19)
if torch.cuda.is_available():
head, inputs = to_cuda(point_head, inputs)
head, inputs = to_cuda(fcn_head, inputs)
prev_output = fcn_head(inputs)
test_cfg = ConfigDict(
subdivision_steps=2, subdivision_num_points=8196, scale_factor=2)
output = point_head.forward_test(inputs, prev_output, None, test_cfg)
assert output.shape == (1, point_head.num_classes, 180, 180)
def test_transformer():
config = ConfigDict(
dict(
encoder=dict(
type='DetrTransformerEncoder',
num_layers=6,
transformerlayers=dict(
type='BaseTransformerLayer',
attn_cfgs=[
dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.1)
],
feedforward_channels=2048,
ffn_dropout=0.1,
operation_order=('self_attn', 'norm', 'ffn', 'norm'))),
decoder=dict(
type='DetrTransformerDecoder',
return_intermediate=True,
num_layers=6,
transformerlayers=dict(
type='DetrTransformerDecoderLayer',
attn_cfgs=dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.1),
feedforward_channels=2048,
ffn_dropout=0.1,
operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
'ffn', 'norm')),
)))
transformer = Transformer(**config)
transformer.init_weights()
def __init__(self,
attn_cfgs=None,
ffn_cfgs=dict(
type='FFN',
embed_dims=256,
feedforward_channels=1024,
num_fcs=2,
ffn_drop=0.,
act_cfg=dict(type='ReLU', inplace=True),
),
operation_order=None,
norm_cfg=dict(type='LN'),
init_cfg=None,
batch_first=False,
**kwargs):
deprecated_args = dict(feedforward_channels='feedforward_channels',
ffn_dropout='ffn_drop',
ffn_num_fcs='num_fcs')
for ori_name, new_name in deprecated_args.items():
if ori_name in kwargs:
warnings.warn(
f'The arguments `{ori_name}` in BaseTransformerLayer '
f'has been deprecated, now you should set `{new_name}` '
f'and other FFN related arguments '
f'to a dict named `ffn_cfgs`. ', DeprecationWarning)
ffn_cfgs[new_name] = kwargs[ori_name]
super(BaseTransformerLayer, self).__init__(init_cfg)
self.batch_first = batch_first
assert set(operation_order) & set(
['self_attn', 'norm', 'ffn', 'cross_attn']) == \
set(operation_order), f'The operation_order of' \
f' {self.__class__.__name__} should ' \
f'contains all four operation type ' \
f"{['self_attn', 'norm', 'ffn', 'cross_attn']}"
num_attn = operation_order.count('self_attn') + operation_order.count(
'cross_attn')
if isinstance(attn_cfgs, dict):
attn_cfgs = [copy.deepcopy(attn_cfgs) for _ in range(num_attn)]
else:
assert num_attn == len(attn_cfgs), f'The length ' \
f'of attn_cfg {num_attn} is ' \
f'not consistent with the number of attention' \
f'in operation_order {operation_order}.'
self.num_attn = num_attn
self.operation_order = operation_order
self.norm_cfg = norm_cfg
self.pre_norm = operation_order[0] == 'norm'
self.attentions = ModuleList()
index = 0
for operation_name in operation_order:
if operation_name in ['self_attn', 'cross_attn']:
if 'batch_first' in attn_cfgs[index]:
assert self.batch_first == attn_cfgs[index]['batch_first']
else:
attn_cfgs[index]['batch_first'] = self.batch_first
attention = build_attention(attn_cfgs[index])
# Some custom attentions used as `self_attn`
# or `cross_attn` can have different behavior.
attention.operation_name = operation_name
self.attentions.append(attention)
index += 1
self.embed_dims = self.attentions[0].embed_dims
self.ffns = ModuleList()
num_ffns = operation_order.count('ffn')
if isinstance(ffn_cfgs, dict):
ffn_cfgs = ConfigDict(ffn_cfgs)
if isinstance(ffn_cfgs, dict):
ffn_cfgs = [copy.deepcopy(ffn_cfgs) for _ in range(num_ffns)]
assert len(ffn_cfgs) == num_ffns
for ffn_index in range(num_ffns):
if 'embed_dims' not in ffn_cfgs[ffn_index]:
ffn_cfgs['embed_dims'] = self.embed_dims
else:
assert ffn_cfgs[ffn_index]['embed_dims'] == self.embed_dims
self.ffns.append(
build_feedforward_network(ffn_cfgs[ffn_index],
dict(type='FFN')))
self.norms = ModuleList()
num_norms = operation_order.count('norm')
for _ in range(num_norms):
self.norms.append(build_norm_layer(norm_cfg, self.embed_dims)[1])
def test_detr_transformer_dencoder_encoder_layer():
config = ConfigDict(
dict(
return_intermediate=True,
num_layers=6,
transformerlayers=dict(
type='DetrTransformerDecoderLayer',
attn_cfgs=dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.1),
feedforward_channels=2048,
ffn_dropout=0.1,
operation_order=(
'norm',
'self_attn',
'norm',
'cross_attn',
'norm',
'ffn',
))))
assert DetrTransformerDecoder(**config).layers[0].pre_norm
assert len(DetrTransformerDecoder(**config).layers) == 6
DetrTransformerDecoder(**config)
with pytest.raises(AssertionError):
config = ConfigDict(
dict(
return_intermediate=True,
num_layers=6,
transformerlayers=[
dict(
type='DetrTransformerDecoderLayer',
attn_cfgs=dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.1),
feedforward_channels=2048,
ffn_dropout=0.1,
operation_order=('self_attn', 'norm', 'cross_attn',
'norm', 'ffn', 'norm'))
] * 5))
DetrTransformerDecoder(**config)
config = ConfigDict(
dict(
num_layers=6,
transformerlayers=dict(
type='DetrTransformerDecoderLayer',
attn_cfgs=dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.1),
feedforward_channels=2048,
ffn_dropout=0.1,
operation_order=('norm', 'self_attn', 'norm', 'cross_attn',
'norm', 'ffn', 'norm'))))
with pytest.raises(AssertionError):
# len(operation_order) == 6
DetrTransformerEncoder(**config)
