def test_resnest_stem():
# Test default stem_channels
model = ResNet(50)
assert model.stem_channels == 64
assert model.conv1.out_channels == 64
assert model.norm1.num_features == 64
# Test default stem_channels, with base_channels=32
model = ResNet(50, base_channels=32)
assert model.stem_channels == 32
assert model.conv1.out_channels == 32
assert model.norm1.num_features == 32
assert model.layer1[0].conv1.in_channels == 32
# Test stem_channels=64
model = ResNet(50, stem_channels=64)
assert model.stem_channels == 64
assert model.conv1.out_channels == 64
assert model.norm1.num_features == 64
assert model.layer1[0].conv1.in_channels == 64
# Test stem_channels=64, with base_channels=32
model = ResNet(50, stem_channels=64, base_channels=32)
assert model.stem_channels == 64
assert model.conv1.out_channels == 64
assert model.norm1.num_features == 64
assert model.layer1[0].conv1.in_channels == 64
# Test stem_channels=128
model = ResNet(depth=50, stem_channels=128)
model.init_weights()
model.train()
assert model.conv1.out_channels == 128
assert model.layer1[0].conv1.in_channels == 128
# Test V1d stem_channels
model = ResNetV1d(depth=50, stem_channels=128)
model.init_weights()
model.train()
assert model.stem[0].out_channels == 64
assert model.stem[1].num_features == 64
assert model.stem[3].out_channels == 64
assert model.stem[4].num_features == 64
assert model.stem[6].out_channels == 128
assert model.stem[7].num_features == 128
assert model.layer1[0].conv1.in_channels == 128
def __init__(self,
depth=50,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 2),
out_indices=(0, 1, 2, 3),
base_channels=64,
**kwargs):
super(ResDepthModel, self).__init__()
self.backbone = ResNetV1d(depth=depth,
strides=strides,
dilations=dilations,
out_indices=out_indices,
base_channels=base_channels)
# output channel =
feat_dim = self.backbone.feat_dim
self.head = DepthPredictHead2Up(feat_dim)
def test_resnest_stem():
# Test default stem_channels
model = ResNet(50)
assert model.stem_channels == 64
assert model.conv1.out_channels == 64
assert model.norm1.num_features == 64
# Test default stem_channels, with base_channels=3
model = ResNet(50, base_channels=3)
assert model.stem_channels == 3
assert model.conv1.out_channels == 3
assert model.norm1.num_features == 3
assert model.layer1[0].conv1.in_channels == 3
# Test stem_channels=3
model = ResNet(50, stem_channels=3)
assert model.stem_channels == 3
assert model.conv1.out_channels == 3
assert model.norm1.num_features == 3
assert model.layer1[0].conv1.in_channels == 3
# Test stem_channels=3, with base_channels=2
model = ResNet(50, stem_channels=3, base_channels=2)
assert model.stem_channels == 3
assert model.conv1.out_channels == 3
assert model.norm1.num_features == 3
assert model.layer1[0].conv1.in_channels == 3
# Test V1d stem_channels
model = ResNetV1d(depth=50, stem_channels=6)
model.train()
assert model.stem[0].out_channels == 3
assert model.stem[1].num_features == 3
assert model.stem[3].out_channels == 3
assert model.stem[4].num_features == 3
assert model.stem[6].out_channels == 6
assert model.stem[7].num_features == 6
assert model.layer1[0].conv1.in_channels == 6
def test_resnet_backbone():
"""Test resnet backbone."""
with pytest.raises(KeyError):
# ResNet depth should be in [18, 34, 50, 101, 152]
ResNet(20)
with pytest.raises(AssertionError):
# In ResNet: 1 <= num_stages <= 4
ResNet(50, num_stages=0)
with pytest.raises(AssertionError):
# len(stage_with_dcn) == num_stages
dcn = dict(type='DCN', deformable_groups=1, fallback_on_stride=False)
ResNet(50, dcn=dcn, stage_with_dcn=(True, ))
with pytest.raises(AssertionError):
# len(stage_with_plugin) == num_stages
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
stages=(False, True, True),
position='after_conv3')
]
ResNet(50, plugins=plugins)
with pytest.raises(AssertionError):
# In ResNet: 1 <= num_stages <= 4
ResNet(50, num_stages=5)
with pytest.raises(AssertionError):
# len(strides) == len(dilations) == num_stages
ResNet(50, strides=(1, ), dilations=(1, 1), num_stages=3)
with pytest.raises(TypeError):
# pretrained must be a string path
model = ResNet(50)
model.init_weights(pretrained=0)
with pytest.raises(AssertionError):
# Style must be in ['pytorch', 'caffe']
ResNet(50, style='tensorflow')
# Test ResNet50 norm_eval=True
model = ResNet(50, norm_eval=True)
model.init_weights()
model.train()
assert check_norm_state(model.modules(), False)
# Test ResNet50 with torchvision pretrained weight
model = ResNet(depth=50, norm_eval=True)
model.init_weights('torchvision://resnet50')
model.train()
assert check_norm_state(model.modules(), False)
# Test ResNet50 with first stage frozen
frozen_stages = 1
model = ResNet(50, frozen_stages=frozen_stages)
model.init_weights()
model.train()
assert model.norm1.training is False
for layer in [model.conv1, model.norm1]:
for param in layer.parameters():
assert param.requires_grad is False
for i in range(1, frozen_stages + 1):
layer = getattr(model, f'layer{i}')
for mod in layer.modules():
if isinstance(mod, _BatchNorm):
assert mod.training is False
for param in layer.parameters():
assert param.requires_grad is False
# Test ResNet50V1d with first stage frozen
model = ResNetV1d(depth=50, frozen_stages=frozen_stages)
assert len(model.stem) == 9
model.init_weights()
model.train()
check_norm_state(model.stem, False)
for param in model.stem.parameters():
assert param.requires_grad is False
for i in range(1, frozen_stages + 1):
layer = getattr(model, f'layer{i}')
for mod in layer.modules():
if isinstance(mod, _BatchNorm):
assert mod.training is False
for param in layer.parameters():
assert param.requires_grad is False
# Test ResNet18 forward
model = ResNet(18)
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 64, 56, 56])
assert feat[1].shape == torch.Size([1, 128, 28, 28])
assert feat[2].shape == torch.Size([1, 256, 14, 14])
assert feat[3].shape == torch.Size([1, 512, 7, 7])
# Test ResNet18 with checkpoint forward
model = ResNet(18, with_cp=True)
for m in model.modules():
if is_block(m):
assert m.with_cp
# Test ResNet50 with BatchNorm forward
model = ResNet(50)
for m in model.modules():
if is_norm(m):
assert isinstance(m, _BatchNorm)
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 with layers 1, 2, 3 out forward
model = ResNet(50, out_indices=(0, 1, 2))
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 3
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
# Test ResNet50 with checkpoint forward
model = ResNet(50, with_cp=True)
for m in model.modules():
if is_block(m):
assert m.with_cp
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 with GroupNorm forward
model = ResNet(
50, norm_cfg=dict(type='GN', num_groups=32, requires_grad=True))
for m in model.modules():
if is_norm(m):
assert isinstance(m, GroupNorm)
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 with 1 GeneralizedAttention after conv2, 1 NonLocal2D
# after conv2, 1 ContextBlock after conv3 in layers 2, 3, 4
plugins = [
dict(
cfg=dict(
type='GeneralizedAttention',
spatial_range=-1,
num_heads=8,
attention_type='0010',
kv_stride=2),
stages=(False, True, True, True),
position='after_conv2'),
dict(cfg=dict(type='NonLocal2D'), position='after_conv2'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
stages=(False, True, True, False),
position='after_conv3')
]
model = ResNet(50, plugins=plugins)
for m in model.layer1.modules():
if is_block(m):
assert not hasattr(m, 'context_block')
assert not hasattr(m, 'gen_attention_block')
assert m.nonlocal_block.in_channels == 64
for m in model.layer2.modules():
if is_block(m):
assert m.nonlocal_block.in_channels == 128
assert m.gen_attention_block.in_channels == 128
assert m.context_block.in_channels == 512
for m in model.layer3.modules():
if is_block(m):
assert m.nonlocal_block.in_channels == 256
assert m.gen_attention_block.in_channels == 256
assert m.context_block.in_channels == 1024
for m in model.layer4.modules():
if is_block(m):
assert m.nonlocal_block.in_channels == 512
assert m.gen_attention_block.in_channels == 512
assert not hasattr(m, 'context_block')
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 with 1 ContextBlock after conv2, 1 ContextBlock after
# conv3 in layers 2, 3, 4
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16, postfix=1),
stages=(False, True, True, False),
position='after_conv3'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16, postfix=2),
stages=(False, True, True, False),
position='after_conv3')
]
model = ResNet(50, plugins=plugins)
for m in model.layer1.modules():
if is_block(m):
assert not hasattr(m, 'context_block')
assert not hasattr(m, 'context_block1')
assert not hasattr(m, 'context_block2')
for m in model.layer2.modules():
if is_block(m):
assert not hasattr(m, 'context_block')
assert m.context_block1.in_channels == 512
assert m.context_block2.in_channels == 512
for m in model.layer3.modules():
if is_block(m):
assert not hasattr(m, 'context_block')
assert m.context_block1.in_channels == 1024
assert m.context_block2.in_channels == 1024
for m in model.layer4.modules():
if is_block(m):
assert not hasattr(m, 'context_block')
assert not hasattr(m, 'context_block1')
assert not hasattr(m, 'context_block2')
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 zero initialization of residual
model = ResNet(50, zero_init_residual=True)
model.init_weights()
for m in model.modules():
if isinstance(m, Bottleneck):
assert all_zeros(m.norm3)
elif isinstance(m, BasicBlock):
assert all_zeros(m.norm2)
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNetV1d forward
model = ResNetV1d(depth=50)
model.init_weights()
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50 stem_channels
model = ResNet(depth=50, stem_channels=128)
model.init_weights()
model.train()
assert model.conv1.out_channels == 128
assert model.layer1[0].conv1.in_channels == 128
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])
# Test ResNet50V1d stem_channels
model = ResNetV1d(depth=50, stem_channels=128)
model.init_weights()
model.train()
assert model.stem[0].out_channels == 64
assert model.stem[3].out_channels == 64
assert model.stem[6].out_channels == 128
assert model.layer1[0].conv1.in_channels == 128
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert len(feat) == 4
assert feat[0].shape == torch.Size([1, 256, 56, 56])
assert feat[1].shape == torch.Size([1, 512, 28, 28])
assert feat[2].shape == torch.Size([1, 1024, 14, 14])
assert feat[3].shape == torch.Size([1, 2048, 7, 7])