def __init__(self,
config,
last_channel,
scale=1.0,
num_classes=1000,
with_pool=True):
super().__init__()
self.config = config
self.scale = scale
self.last_channel = last_channel
self.num_classes = num_classes
self.with_pool = with_pool
self.firstconv_in_channels = config[0].in_channels
self.lastconv_in_channels = config[-1].in_channels
self.lastconv_out_channels = self.lastconv_in_channels * 6
norm_layer = partial(nn.BatchNorm2D, epsilon=0.001, momentum=0.99)
self.conv = ConvNormActivation(in_channels=3,
out_channels=self.firstconv_in_channels,
kernel_size=3,
stride=2,
padding=1,
groups=1,
activation_layer=nn.Hardswish,
norm_layer=norm_layer)
self.blocks = nn.Sequential(*[
InvertedResidual(in_channels=cfg.in_channels,
expanded_channels=cfg.expanded_channels,
out_channels=cfg.out_channels,
filter_size=cfg.kernel,
stride=cfg.stride,
use_se=cfg.use_se,
activation_layer=cfg.activation_layer,
norm_layer=norm_layer) for cfg in self.config
])
self.lastconv = ConvNormActivation(
in_channels=self.lastconv_in_channels,
out_channels=self.lastconv_out_channels,
kernel_size=1,
stride=1,
padding=0,
groups=1,
norm_layer=norm_layer,
activation_layer=nn.Hardswish)
if with_pool:
self.avgpool = nn.AdaptiveAvgPool2D(1)
if num_classes > 0:
self.classifier = nn.Sequential(
nn.Linear(self.lastconv_out_channels, self.last_channel),
nn.Hardswish(), nn.Dropout(p=0.2),
nn.Linear(self.last_channel, num_classes))
def _create_act(act):
if act == "hardswish":
return nn.Hardswish()
elif act == "relu":
return nn.ReLU()
elif act is None:
return Identity()
else:
raise RuntimeError(
"The activation function is not supported: {}".format(act))
def __init__(self,
class_num=1000,
scale=1.0,
dropout_prob=0.2,
class_expand=1280):
super().__init__()
self.scale = scale
self.class_num = class_num
self.class_expand = class_expand
stage_repeats = [3, 7, 3]
stage_out_channels = [
-1, 24,
make_divisible(116 * scale),
make_divisible(232 * scale),
make_divisible(464 * scale), 1024
]
self.conv1 = ConvBNLayer(in_channels=3,
out_channels=stage_out_channels[1],
kernel_size=3,
stride=2)
self.max_pool = MaxPool2D(kernel_size=3, stride=2, padding=1)
block_list = []
for stage_id, num_repeat in enumerate(stage_repeats):
for i in range(num_repeat):
if i == 0:
block = ESBlock2(
in_channels=stage_out_channels[stage_id + 1],
out_channels=stage_out_channels[stage_id + 2])
else:
block = ESBlock1(
in_channels=stage_out_channels[stage_id + 2],
out_channels=stage_out_channels[stage_id + 2])
block_list.append(block)
self.blocks = nn.Sequential(*block_list)
self.conv2 = ConvBNLayer(in_channels=stage_out_channels[-2],
out_channels=stage_out_channels[-1],
kernel_size=1)
self.avg_pool = AdaptiveAvgPool2D(1)
self.last_conv = Conv2D(in_channels=stage_out_channels[-1],
out_channels=self.class_expand,
kernel_size=1,
stride=1,
padding=0,
bias_attr=False)
self.hardswish = nn.Hardswish()
self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
self.fc = Linear(self.class_expand, self.class_num)
def __init__(self,
num_channels,
filter_size,
num_filters,
stride,
num_groups=1):
super().__init__()
self.conv = Conv2D(in_channels=num_channels,
out_channels=num_filters,
kernel_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=num_groups,
weight_attr=ParamAttr(initializer=KaimingNormal()),
bias_attr=False)
self.bn = BatchNorm(num_filters,
param_attr=ParamAttr(regularizer=L2Decay(0.0)),
bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
self.hardswish = nn.Hardswish()
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
groups=1,
if_act=True):
super().__init__()
self.conv = Conv2D(in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=(kernel_size - 1) // 2,
groups=groups,
weight_attr=ParamAttr(initializer=KaimingNormal()),
bias_attr=False)
self.bn = BatchNorm(out_channels,
param_attr=ParamAttr(regularizer=L2Decay(0.0)),
bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
self.if_act = if_act
self.hardswish = nn.Hardswish()
def get_activation_layer(activation):
"""
Create activation layer from string/function.
Parameters:
----------
activation : function, or str, or nn.Module
Activation function or name of activation function.
Returns:
-------
nn.Module
Activation layer.
"""
assert activation is not None
if isfunction(activation):
return activation()
elif isinstance(activation, str):
if activation == "relu":
return nn.ReLU()
elif activation == "relu6":
return nn.ReLU6()
elif activation == "swish":
return nn.Swish()
elif activation == "hswish":
return nn.Hardswish()
elif activation == "sigmoid":
return nn.Sigmoid()
elif activation == "hsigmoid":
return nn.Hardsigmoid()
elif activation == "identity":
return Identity()
else:
raise NotImplementedError()
else:
assert isinstance(activation, nn.Layer)
return activation
def __init__(
self,
inverted_residual_setting: List[InvertedResidualConfig],
last_channel: int,
num_classes: int = 1000,
block: Optional[Callable[..., nn.Layer]] = None,
norm_layer: Optional[Callable[..., nn.Layer]] = None,
dropout: float = 0.2,
**kwargs: Any,
) -> None:
"""
MobileNet V3 main class
Args:
inverted_residual_setting (List[InvertedResidualConfig]): Network structure
last_channel (int): The number of channels on the penultimate layer
num_classes (int): Number of classes
block (Optional[Callable[..., nn.Layer]]): Module specifying inverted residual building block for mobilenet
norm_layer (Optional[Callable[..., nn.Layer]]): Module specifying the normalization layer to use
dropout (float): The droupout probability
"""
super().__init__()
if not inverted_residual_setting:
raise ValueError(
"The inverted_residual_setting should not be empty")
elif not (isinstance(inverted_residual_setting, Sequence) and all([
isinstance(s, InvertedResidualConfig)
for s in inverted_residual_setting
])):
raise TypeError(
"The inverted_residual_setting should be List[InvertedResidualConfig]"
)
if block is None:
block = InvertedResidual
if norm_layer is None:
norm_layer = partial(nn.BatchNorm2D, epsilon=0.001, momentum=0.01)
layers: List[nn.Layer] = []
# building first layer
firstconv_output_channels = inverted_residual_setting[0].input_channels
layers.append(
ConvNormActivation(
3,
firstconv_output_channels,
kernel_size=3,
stride=2,
norm_layer=norm_layer,
activation_layer=nn.Hardswish,
))
# building inverted residual blocks
for cnf in inverted_residual_setting:
layers.append(block(cnf, norm_layer))
# building last several layers
lastconv_input_channels = inverted_residual_setting[-1].out_channels
lastconv_output_channels = 6 * lastconv_input_channels
layers.append(
ConvNormActivation(
lastconv_input_channels,
lastconv_output_channels,
kernel_size=1,
norm_layer=norm_layer,
activation_layer=nn.Hardswish,
))
self.features = nn.Sequential(*layers)
self.avgpool = nn.AdaptiveAvgPool2D(1)
self.classifier = nn.Sequential(
nn.Linear(lastconv_output_channels, last_channel),
nn.Hardswish(),
nn.Dropout(p=dropout),
nn.Linear(last_channel, num_classes),
)
def __init__(self,
config,
stages_pattern,
scale=1.0,
class_num=1000,
inplanes=STEM_CONV_NUMBER,
class_squeeze=LAST_SECOND_CONV_LARGE,
class_expand=LAST_CONV,
dropout_prob=0.2,
return_patterns=None,
return_stages=None):
super().__init__()
self.cfg = config
self.scale = scale
self.inplanes = inplanes
self.class_squeeze = class_squeeze
self.class_expand = class_expand
self.class_num = class_num
self.conv = ConvBNLayer(
in_c=3,
out_c=_make_divisible(self.inplanes * self.scale),
filter_size=3,
stride=2,
padding=1,
num_groups=1,
if_act=True,
act="hardswish")
self.blocks = nn.Sequential(* [
ResidualUnit(
in_c=_make_divisible(self.inplanes * self.scale if i == 0 else
self.cfg[i - 1][2] * self.scale),
mid_c=_make_divisible(self.scale * exp),
out_c=_make_divisible(self.scale * c),
filter_size=k,
stride=s,
use_se=se,
act=act) for i, (k, exp, c, se, act, s) in enumerate(self.cfg)
])
self.last_second_conv = ConvBNLayer(
in_c=_make_divisible(self.cfg[-1][2] * self.scale),
out_c=_make_divisible(self.scale * self.class_squeeze),
filter_size=1,
stride=1,
padding=0,
num_groups=1,
if_act=True,
act="hardswish")
self.avg_pool = AdaptiveAvgPool2D(1)
self.last_conv = Conv2D(
in_channels=_make_divisible(self.scale * self.class_squeeze),
out_channels=self.class_expand,
kernel_size=1,
stride=1,
padding=0,
bias_attr=False)
self.hardswish = nn.Hardswish()
if dropout_prob is not None:
self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
else:
self.dropout = None
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
self.fc = Linear(self.class_expand, class_num)
super().init_res(
stages_pattern,
return_patterns=return_patterns,
return_stages=return_stages)
def func_test_layer_str(self):
module = nn.ELU(0.2)
self.assertEqual(str(module), 'ELU(alpha=0.2)')
module = nn.CELU(0.2)
self.assertEqual(str(module), 'CELU(alpha=0.2)')
module = nn.GELU(True)
self.assertEqual(str(module), 'GELU(approximate=True)')
module = nn.Hardshrink()
self.assertEqual(str(module), 'Hardshrink(threshold=0.5)')
module = nn.Hardswish(name="Hardswish")
self.assertEqual(str(module), 'Hardswish(name=Hardswish)')
module = nn.Tanh(name="Tanh")
self.assertEqual(str(module), 'Tanh(name=Tanh)')
module = nn.Hardtanh(name="Hardtanh")
self.assertEqual(str(module),
'Hardtanh(min=-1.0, max=1.0, name=Hardtanh)')
module = nn.PReLU(1, 0.25, name="PReLU", data_format="NCHW")
self.assertEqual(
str(module),
'PReLU(num_parameters=1, data_format=NCHW, init=0.25, dtype=float32, name=PReLU)'
)
module = nn.ReLU()
self.assertEqual(str(module), 'ReLU()')
module = nn.ReLU6()
self.assertEqual(str(module), 'ReLU6()')
module = nn.SELU()
self.assertEqual(
str(module),
'SELU(scale=1.0507009873554805, alpha=1.6732632423543772)')
module = nn.LeakyReLU()
self.assertEqual(str(module), 'LeakyReLU(negative_slope=0.01)')
module = nn.Sigmoid()
self.assertEqual(str(module), 'Sigmoid()')
module = nn.Hardsigmoid()
self.assertEqual(str(module), 'Hardsigmoid()')
module = nn.Softplus()
self.assertEqual(str(module), 'Softplus(beta=1, threshold=20)')
module = nn.Softshrink()
self.assertEqual(str(module), 'Softshrink(threshold=0.5)')
module = nn.Softsign()
self.assertEqual(str(module), 'Softsign()')
module = nn.Swish()
self.assertEqual(str(module), 'Swish()')
module = nn.Tanhshrink()
self.assertEqual(str(module), 'Tanhshrink()')
module = nn.ThresholdedReLU()
self.assertEqual(str(module), 'ThresholdedReLU(threshold=1.0)')
module = nn.LogSigmoid()
self.assertEqual(str(module), 'LogSigmoid()')
module = nn.Softmax()
self.assertEqual(str(module), 'Softmax(axis=-1)')
module = nn.LogSoftmax()
self.assertEqual(str(module), 'LogSoftmax(axis=-1)')
module = nn.Maxout(groups=2)
self.assertEqual(str(module), 'Maxout(groups=2, axis=1)')
module = nn.Linear(2, 4, name='linear')
self.assertEqual(
str(module),
'Linear(in_features=2, out_features=4, dtype=float32, name=linear)'
)
module = nn.Upsample(size=[12, 12])
self.assertEqual(
str(module),
'Upsample(size=[12, 12], mode=nearest, align_corners=False, align_mode=0, data_format=NCHW)'
)
module = nn.UpsamplingNearest2D(size=[12, 12])
self.assertEqual(
str(module),
'UpsamplingNearest2D(size=[12, 12], data_format=NCHW)')
module = nn.UpsamplingBilinear2D(size=[12, 12])
self.assertEqual(
str(module),
'UpsamplingBilinear2D(size=[12, 12], data_format=NCHW)')
module = nn.Bilinear(in1_features=5, in2_features=4, out_features=1000)
self.assertEqual(
str(module),
'Bilinear(in1_features=5, in2_features=4, out_features=1000, dtype=float32)'
)
module = nn.Dropout(p=0.5)
self.assertEqual(str(module),
'Dropout(p=0.5, axis=None, mode=upscale_in_train)')
module = nn.Dropout2D(p=0.5)
self.assertEqual(str(module), 'Dropout2D(p=0.5, data_format=NCHW)')
module = nn.Dropout3D(p=0.5)
self.assertEqual(str(module), 'Dropout3D(p=0.5, data_format=NCDHW)')
module = nn.AlphaDropout(p=0.5)
self.assertEqual(str(module), 'AlphaDropout(p=0.5)')
module = nn.Pad1D(padding=[1, 2], mode='constant')
self.assertEqual(
str(module),
'Pad1D(padding=[1, 2], mode=constant, value=0.0, data_format=NCL)')
module = nn.Pad2D(padding=[1, 0, 1, 2], mode='constant')
self.assertEqual(
str(module),
'Pad2D(padding=[1, 0, 1, 2], mode=constant, value=0.0, data_format=NCHW)'
)
module = nn.ZeroPad2D(padding=[1, 0, 1, 2])
self.assertEqual(str(module),
'ZeroPad2D(padding=[1, 0, 1, 2], data_format=NCHW)')
module = nn.Pad3D(padding=[1, 0, 1, 2, 0, 0], mode='constant')
self.assertEqual(
str(module),
'Pad3D(padding=[1, 0, 1, 2, 0, 0], mode=constant, value=0.0, data_format=NCDHW)'
)
module = nn.CosineSimilarity(axis=0)
self.assertEqual(str(module), 'CosineSimilarity(axis=0, eps=1e-08)')
module = nn.Embedding(10, 3, sparse=True)
self.assertEqual(str(module), 'Embedding(10, 3, sparse=True)')
module = nn.Conv1D(3, 2, 3)
self.assertEqual(str(module),
'Conv1D(3, 2, kernel_size=[3], data_format=NCL)')
module = nn.Conv1DTranspose(2, 1, 2)
self.assertEqual(
str(module),
'Conv1DTranspose(2, 1, kernel_size=[2], data_format=NCL)')
module = nn.Conv2D(4, 6, (3, 3))
self.assertEqual(str(module),
'Conv2D(4, 6, kernel_size=[3, 3], data_format=NCHW)')
module = nn.Conv2DTranspose(4, 6, (3, 3))
self.assertEqual(
str(module),
'Conv2DTranspose(4, 6, kernel_size=[3, 3], data_format=NCHW)')
module = nn.Conv3D(4, 6, (3, 3, 3))
self.assertEqual(
str(module),
'Conv3D(4, 6, kernel_size=[3, 3, 3], data_format=NCDHW)')
module = nn.Conv3DTranspose(4, 6, (3, 3, 3))
self.assertEqual(
str(module),
'Conv3DTranspose(4, 6, kernel_size=[3, 3, 3], data_format=NCDHW)')
module = nn.PairwiseDistance()
self.assertEqual(str(module), 'PairwiseDistance(p=2.0)')
module = nn.InstanceNorm1D(2)
self.assertEqual(str(module),
'InstanceNorm1D(num_features=2, epsilon=1e-05)')
module = nn.InstanceNorm2D(2)
self.assertEqual(str(module),
'InstanceNorm2D(num_features=2, epsilon=1e-05)')
module = nn.InstanceNorm3D(2)
self.assertEqual(str(module),
'InstanceNorm3D(num_features=2, epsilon=1e-05)')
module = nn.GroupNorm(num_channels=6, num_groups=6)
self.assertEqual(
str(module),
'GroupNorm(num_groups=6, num_channels=6, epsilon=1e-05)')
module = nn.LayerNorm([2, 2, 3])
self.assertEqual(
str(module),
'LayerNorm(normalized_shape=[2, 2, 3], epsilon=1e-05)')
module = nn.BatchNorm1D(1)
self.assertEqual(
str(module),
'BatchNorm1D(num_features=1, momentum=0.9, epsilon=1e-05, data_format=NCL)'
)
module = nn.BatchNorm2D(1)
self.assertEqual(
str(module),
'BatchNorm2D(num_features=1, momentum=0.9, epsilon=1e-05)')
module = nn.BatchNorm3D(1)
self.assertEqual(
str(module),
'BatchNorm3D(num_features=1, momentum=0.9, epsilon=1e-05, data_format=NCDHW)'
)
module = nn.SyncBatchNorm(2)
self.assertEqual(
str(module),
'SyncBatchNorm(num_features=2, momentum=0.9, epsilon=1e-05)')
module = nn.LocalResponseNorm(size=5)
self.assertEqual(
str(module),
'LocalResponseNorm(size=5, alpha=0.0001, beta=0.75, k=1.0)')
module = nn.AvgPool1D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'AvgPool1D(kernel_size=2, stride=2, padding=0)')
module = nn.AvgPool2D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'AvgPool2D(kernel_size=2, stride=2, padding=0)')
module = nn.AvgPool3D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'AvgPool3D(kernel_size=2, stride=2, padding=0)')
module = nn.MaxPool1D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'MaxPool1D(kernel_size=2, stride=2, padding=0)')
module = nn.MaxPool2D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'MaxPool2D(kernel_size=2, stride=2, padding=0)')
module = nn.MaxPool3D(kernel_size=2, stride=2, padding=0)
self.assertEqual(str(module),
'MaxPool3D(kernel_size=2, stride=2, padding=0)')
module = nn.AdaptiveAvgPool1D(output_size=16)
self.assertEqual(str(module), 'AdaptiveAvgPool1D(output_size=16)')
module = nn.AdaptiveAvgPool2D(output_size=3)
self.assertEqual(str(module), 'AdaptiveAvgPool2D(output_size=3)')
module = nn.AdaptiveAvgPool3D(output_size=3)
self.assertEqual(str(module), 'AdaptiveAvgPool3D(output_size=3)')
module = nn.AdaptiveMaxPool1D(output_size=16, return_mask=True)
self.assertEqual(
str(module), 'AdaptiveMaxPool1D(output_size=16, return_mask=True)')
module = nn.AdaptiveMaxPool2D(output_size=3, return_mask=True)
self.assertEqual(str(module),
'AdaptiveMaxPool2D(output_size=3, return_mask=True)')
module = nn.AdaptiveMaxPool3D(output_size=3, return_mask=True)
self.assertEqual(str(module),
'AdaptiveMaxPool3D(output_size=3, return_mask=True)')
module = nn.SimpleRNNCell(16, 32)
self.assertEqual(str(module), 'SimpleRNNCell(16, 32)')
module = nn.LSTMCell(16, 32)
self.assertEqual(str(module), 'LSTMCell(16, 32)')
module = nn.GRUCell(16, 32)
self.assertEqual(str(module), 'GRUCell(16, 32)')
module = nn.PixelShuffle(3)
self.assertEqual(str(module), 'PixelShuffle(upscale_factor=3)')
module = nn.SimpleRNN(16, 32, 2)
self.assertEqual(
str(module),
'SimpleRNN(16, 32, num_layers=2\n (0): RNN(\n (cell): SimpleRNNCell(16, 32)\n )\n (1): RNN(\n (cell): SimpleRNNCell(32, 32)\n )\n)'
)
module = nn.LSTM(16, 32, 2)
self.assertEqual(
str(module),
'LSTM(16, 32, num_layers=2\n (0): RNN(\n (cell): LSTMCell(16, 32)\n )\n (1): RNN(\n (cell): LSTMCell(32, 32)\n )\n)'
)
module = nn.GRU(16, 32, 2)
self.assertEqual(
str(module),
'GRU(16, 32, num_layers=2\n (0): RNN(\n (cell): GRUCell(16, 32)\n )\n (1): RNN(\n (cell): GRUCell(32, 32)\n )\n)'
)
module1 = nn.Sequential(
('conv1', nn.Conv2D(1, 20, 5)), ('relu1', nn.ReLU()),
('conv2', nn.Conv2D(20, 64, 5)), ('relu2', nn.ReLU()))
self.assertEqual(
str(module1),
'Sequential(\n '\
'(conv1): Conv2D(1, 20, kernel_size=[5, 5], data_format=NCHW)\n '\
'(relu1): ReLU()\n '\
'(conv2): Conv2D(20, 64, kernel_size=[5, 5], data_format=NCHW)\n '\
'(relu2): ReLU()\n)'
)
module2 = nn.Sequential(
nn.Conv3DTranspose(4, 6, (3, 3, 3)),
nn.AvgPool3D(kernel_size=2, stride=2, padding=0),
nn.Tanh(name="Tanh"), module1, nn.Conv3D(4, 6, (3, 3, 3)),
nn.MaxPool3D(kernel_size=2, stride=2, padding=0), nn.GELU(True))
self.assertEqual(
str(module2),
'Sequential(\n '\
'(0): Conv3DTranspose(4, 6, kernel_size=[3, 3, 3], data_format=NCDHW)\n '\
'(1): AvgPool3D(kernel_size=2, stride=2, padding=0)\n '\
'(2): Tanh(name=Tanh)\n '\
'(3): Sequential(\n (conv1): Conv2D(1, 20, kernel_size=[5, 5], data_format=NCHW)\n (relu1): ReLU()\n'\
' (conv2): Conv2D(20, 64, kernel_size=[5, 5], data_format=NCHW)\n (relu2): ReLU()\n )\n '\
'(4): Conv3D(4, 6, kernel_size=[3, 3, 3], data_format=NCDHW)\n '\
'(5): MaxPool3D(kernel_size=2, stride=2, padding=0)\n '\
'(6): GELU(approximate=True)\n)'
)
def __init__(self, scale=1.0, class_expand=1280):
super(PPLCNet, self).__init__()
self.scale = scale
self.class_expand = class_expand
self.conv1 = ConvBNLayer(
num_channels=3,
filter_size=3,
num_filters=make_divisible(16 * scale),
stride=2)
self.blocks2 = nn.Sequential(*[
DepthwiseSeparable(
num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks2"])
])
self.blocks3 = nn.Sequential(*[
DepthwiseSeparable(
num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks3"])
])
self.blocks4 = nn.Sequential(*[
DepthwiseSeparable(
num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks4"])
])
self.blocks5 = nn.Sequential(*[
DepthwiseSeparable(
num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks5"])
])
self.blocks6 = nn.Sequential(*[
DepthwiseSeparable(
num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks6"])
])
self.avg_pool = AdaptiveAvgPool2D(1)
self.last_conv = Conv2D(
in_channels=make_divisible(NET_CONFIG["blocks6"][-1][2] * scale),
out_channels=self.class_expand,
kernel_size=1,
stride=1,
padding=0,
bias_attr=False)
self.hardswish = nn.Hardswish()
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
def __init__(self, block_args, global_params):
super().__init__()
self._block_args = block_args
self._bn_mom = global_params.batch_norm_momentum
self._bn_eps = global_params.batch_norm_epsilon
self.has_se = (self._block_args.se_ratio
is not None) and (0 < self._block_args.se_ratio <= 1)
self.id_skip = block_args.id_skip # skip connection and drop connect
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Expansion phase
inp = self._block_args.input_filters # number of input channels
oup = self._block_args.input_filters * self._block_args.expand_ratio # number of output channels
if self._block_args.expand_ratio != 1:
self._expand_conv = Conv2d(in_channels=inp,
out_channels=oup,
kernel_size=1,
bias_attr=False)
self._bn0 = nn.BatchNorm2D(num_features=oup,
momentum=self._bn_mom,
epsilon=self._bn_eps)
# Depthwise convolution phase
k = self._block_args.kernel_size
s = self._block_args.stride
self._depthwise_conv = Conv2d(
in_channels=oup,
out_channels=oup,
groups=oup, # groups makes it depthwise
kernel_size=k,
stride=s,
bias_attr=False)
self._bn1 = nn.BatchNorm2D(num_features=oup,
momentum=self._bn_mom,
epsilon=self._bn_eps)
# Squeeze and Excitation layer, if desired
if self.has_se:
num_squeezed_channels = max(
1,
int(self._block_args.input_filters *
self._block_args.se_ratio))
self._se_reduce = Conv2d(in_channels=oup,
out_channels=num_squeezed_channels,
kernel_size=1)
self._se_expand = Conv2d(in_channels=num_squeezed_channels,
out_channels=oup,
kernel_size=1)
# Output phase
final_oup = self._block_args.output_filters
self._project_conv = Conv2d(in_channels=oup,
out_channels=final_oup,
kernel_size=1,
bias_attr=False)
self._bn2 = nn.BatchNorm2D(num_features=final_oup,
momentum=self._bn_mom,
epsilon=self._bn_eps)
self._swish = nn.Hardswish()
def set_swish(self, memory_efficient=True):
"""Sets swish function as memory efficient (for training) or standard (for export)"""
self._swish = nn.Hardswish() if memory_efficient else nn.Swish()
for block in self._blocks:
block.set_swish(memory_efficient)
def __init__(self, blocks_args=None, global_params=None):
super().__init__()
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Batch norm parameters
bn_mom = self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
# Stem
in_channels = 3 # rgb
out_channels = round_filters(
32, self._global_params) # number of output channels
self._conv_stem = Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=2,
bias_attr=False)
self._bn0 = nn.BatchNorm2D(num_features=out_channels,
momentum=bn_mom,
epsilon=bn_eps)
# Build blocks
self._blocks = nn.LayerList([])
for block_args in self._blocks_args:
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params))
# The first block needs to take care of stride and filter size increase.
self._blocks.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args, self._global_params))
# Head
in_channels = block_args.output_filters # output of final block
out_channels = round_filters(1280, self._global_params)
self._conv_head = Conv2d(in_channels,
out_channels,
kernel_size=1,
bias_attr=False)
self._bn1 = nn.BatchNorm2D(num_features=out_channels,
momentum=bn_mom,
epsilon=bn_eps)
# Final linear layer
self._avg_pooling = nn.AdaptiveAvgPool2D(1)
self._dropout = nn.Dropout(self._global_params.dropout_rate)
self._fc = nn.Linear(out_channels, self._global_params.num_classes)
self._swish = nn.Hardswish()
def __init__(self,
inverted_residual_setting,
last_channel,
num_classes=1000,
block=None,
norm_layer=None,
**kwargs):
"""
MobileNet V3 main class
Args:
inverted_residual_setting (List[InvertedResidualConfig]): Network structure
last_channel (int): The number of channels on the penultimate layer
num_classes (int): Number of classes
block (Optional[Callable[..., nn.Layer]]): Module specifying inverted residual building block for mobilenet
norm_layer (Optional[Callable[..., nn.Layer]]): Module specifying the normalization layer to use
"""
super().__init__()
if not inverted_residual_setting:
raise ValueError(
"The inverted_residual_setting should not be empty")
elif not (isinstance(inverted_residual_setting, Sequence) and all([
isinstance(s, InvertedResidualConfig)
for s in inverted_residual_setting
])):
raise TypeError(
"The inverted_residual_setting should be List[InvertedResidualConfig]"
)
if block is None:
block = InvertedResidual
if norm_layer is None:
norm_layer = partial(nn.BatchNorm2d, eps=0.001, momentum=0.01)
layers = []
# building first layer
firstconv_output_channels = inverted_residual_setting[0].input_channels
layers.append(
ConvBNActivation(3,
firstconv_output_channels,
kernel_size=3,
stride=2,
norm_layer=norm_layer,
activation_layer=nn.Hardswish))
# building inverted residual blocks
for cnf in inverted_residual_setting:
layers.append(block(cnf, norm_layer))
# building last several layers
lastconv_input_channels = inverted_residual_setting[-1].out_channels
lastconv_output_channels = 6 * lastconv_input_channels
layers.append(
ConvBNActivation(lastconv_input_channels,
lastconv_output_channels,
kernel_size=1,
norm_layer=norm_layer,
activation_layer=nn.Hardswish))
self.features = nn.Sequential(*layers)
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.classifier = nn.Sequential(
nn.Linear(lastconv_output_channels, last_channel),
nn.Hardswish(),
nn.Dropout(p=0.2, ),
nn.Linear(last_channel, num_classes),
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
nn.init.zeros_(m.bias)
def __init__(self,
stages_pattern,
scale=1.0,
class_num=1000,
dropout_prob=0.2,
class_expand=1280,
return_patterns=None,
return_stages=None):
super().__init__()
self.scale = scale
self.class_expand = class_expand
self.conv1 = ConvBNLayer(num_channels=3,
filter_size=3,
num_filters=make_divisible(16 * scale),
stride=2)
self.blocks2 = nn.Sequential(*[
DepthwiseSeparable(num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks2"])
])
self.blocks3 = nn.Sequential(*[
DepthwiseSeparable(num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks3"])
])
self.blocks4 = nn.Sequential(*[
DepthwiseSeparable(num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks4"])
])
self.blocks5 = nn.Sequential(*[
DepthwiseSeparable(num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks5"])
])
self.blocks6 = nn.Sequential(*[
DepthwiseSeparable(num_channels=make_divisible(in_c * scale),
num_filters=make_divisible(out_c * scale),
dw_size=k,
stride=s,
use_se=se)
for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks6"])
])
self.avg_pool = AdaptiveAvgPool2D(1)
self.last_conv = Conv2D(in_channels=make_divisible(
NET_CONFIG["blocks6"][-1][2] * scale),
out_channels=self.class_expand,
kernel_size=1,
stride=1,
padding=0,
bias_attr=False)
self.hardswish = nn.Hardswish()
self.dropout = Dropout(p=dropout_prob, mode="downscale_in_infer")
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
self.fc = Linear(self.class_expand, class_num)
super().init_res(stages_pattern,
return_patterns=return_patterns,
return_stages=return_stages)
