def init_weights(self):
"""Initialize weights of FPN."""
for m in self.modules():
if isinstance(m, nn.Conv2d):
kaiming_init(m)
elif is_norm(m):
constant_init(m, 1)
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
normal_init(m, mean=0, std=0.01)
if is_norm(m):
constant_init(m, 1)
# Use prior in model initialization to improve stability
bias_cls = bias_init_with_prob(0.01)
torch.nn.init.constant_(self.cls_score.bias, bias_cls)
def init_weights(self):
caffe2_xavier_init(self.lateral_conv)
caffe2_xavier_init(self.fpn_conv)
for m in [self.lateral_norm, self.fpn_norm]:
constant_init(m, 1)
for m in self.dilated_encoder_blocks.modules():
if isinstance(m, nn.Conv2d):
normal_init(m, mean=0, std=0.01)
if is_norm(m):
constant_init(m, 1)
def _init_weights(module, init_linear='normal', std=0.01, bias=0):
assert init_linear in ['normal', 'kaiming']
for m in module.modules():
if isinstance(m, nn.Linear):
if init_linear == 'normal':
normal_init(m, std=std, bias=bias)
else:
kaiming_init(m, mode='fan_in', nonlinearity='relu')
elif is_norm(m):
if m.weight is not None:
nn.init.constant_(m.weight, 1)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
normal_init(m, mean=0, std=0.01)
if is_norm(m):
constant_init(m, 1)
# Use prior in model initialization to improve stability
for conv_pred, stride in zip(self.convs_pred, self.featmap_strides):
bias = conv_pred.bias.reshape(self.num_anchors, -1)
# init objectness with prior of 8 objects per feature map
# refer to https://github.com/ultralytics/yolov3
nn.init.constant_(bias.data[:, 4],
bias_init_with_prob(8 / (608 / stride)**2))
nn.init.constant_(bias.data[:, 5:], bias_init_with_prob(0.01))
def test_inverted_residual():
with pytest.raises(AssertionError):
# stride must be in [1, 2]
InvertedResidual(16, 16, 32, stride=3)
with pytest.raises(AssertionError):
# se_cfg must be None or dict
InvertedResidual(16, 16, 32, se_cfg=list())
with pytest.raises(AssertionError):
# in_channeld and mid_channels must be the same if
# with_expand_conv is False
InvertedResidual(16, 16, 32, with_expand_conv=False)
# Test InvertedResidual forward, stride=1
block = InvertedResidual(16, 16, 32, stride=1)
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
assert getattr(block, 'se', None) is None
assert block.with_res_shortcut
assert x_out.shape == torch.Size((1, 16, 56, 56))
# Test InvertedResidual forward, stride=2
block = InvertedResidual(16, 16, 32, stride=2)
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
assert not block.with_res_shortcut
assert x_out.shape == torch.Size((1, 16, 28, 28))
# Test InvertedResidual forward with se layer
se_cfg = dict(channels=32)
block = InvertedResidual(16, 16, 32, stride=1, se_cfg=se_cfg)
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
assert isinstance(block.se, SELayer)
assert x_out.shape == torch.Size((1, 16, 56, 56))
# Test InvertedResidual forward, with_expand_conv=False
block = InvertedResidual(32, 16, 32, with_expand_conv=False)
x = torch.randn(1, 32, 56, 56)
x_out = block(x)
assert getattr(block, 'expand_conv', None) is None
assert x_out.shape == torch.Size((1, 16, 56, 56))
# Test InvertedResidual forward with GroupNorm
block = InvertedResidual(16,
16,
32,
norm_cfg=dict(type='GN', num_groups=2))
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
for m in block.modules():
if is_norm(m):
assert isinstance(m, GroupNorm)
assert x_out.shape == torch.Size((1, 16, 56, 56))
# Test InvertedResidual forward with HSigmoid
block = InvertedResidual(16, 16, 32, act_cfg=dict(type='HSigmoid'))
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size((1, 16, 56, 56))
# Test InvertedResidual forward with checkpoint
block = InvertedResidual(16, 16, 32, with_cp=True)
x = torch.randn(1, 16, 56, 56)
x_out = block(x)
assert block.with_cp
assert x_out.shape == torch.Size((1, 16, 56, 56))
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
caffe2_xavier_init(m)
elif is_norm(m):
constant_init(m, 1.0)