def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
expand_ratio=6,
mid_channels=None):
super(MBInvertedConvLayer, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.expand_ratio = expand_ratio
self.mid_channels = mid_channels
if self.mid_channels is None:
feature_dim = round(self.in_channels * self.expand_ratio)
else:
feature_dim = self.mid_channels
if self.expand_ratio == 1:
self.inverted_bottleneck = nn.Sequential()
else:
self.inverted_bottleneck = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(self.in_channels,
feature_dim,
1,
1,
0,
bias=False)),
('bn', nn.BatchNorm2d(feature_dim)),
('act', nn.ReLU6(inplace=True)),
]))
pad = get_same_padding(self.kernel_size)
self.depth_conv = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(feature_dim,
feature_dim,
kernel_size,
stride,
pad,
groups=feature_dim,
bias=False)),
('bn', nn.BatchNorm2d(feature_dim)),
('act', nn.ReLU6(inplace=True)),
]))
self.point_linear = nn.Sequential(
OrderedDict([
('conv',
nn.Conv2d(feature_dim, out_channels, 1, 1, 0, bias=False)),
('bn', nn.BatchNorm2d(out_channels)),
]))
def __init__(self, alpha, depths, convops, kernel_sizes, num_layers,
skips, num_classes=1000, dropout=0.2):
super().__init__()
assert alpha > 0.0
assert len(depths) == len(convops) == len(kernel_sizes) == len(num_layers) == len(skips) == 7
self.alpha = alpha
self.num_classes = num_classes
depths = _get_depths([_FIRST_DEPTH] + depths, alpha)
base_filter_sizes = [16, 24, 40, 80, 96, 192, 320]
exp_ratios = [3, 3, 3, 6, 6, 6, 6]
strides = [1, 2, 2, 2, 1, 2, 1]
layers = [
# First layer: regular conv.
nn.Conv2d(3, depths[0], 3, padding=1, stride=2, bias=False),
nn.BatchNorm2d(depths[0], momentum=_BN_MOMENTUM),
nn.ReLU(inplace=True),
]
count = 0
# for conv, prev_depth, depth, ks, skip, stride, repeat, exp_ratio in \
# zip(convops, depths[:-1], depths[1:], kernel_sizes, skips, strides, num_layers, exp_ratios):
for filter_size, exp_ratio, stride in zip(base_filter_sizes, exp_ratios, strides):
# TODO: restrict that "choose" can only be used within mutator
ph = nn.Placeholder(label=f'mutable_{count}', **{
'kernel_size_options': [1, 3, 5],
'n_layer_options': [1, 2, 3, 4],
'op_type_options': ['__mutated__.base_mnasnet.RegularConv',
'__mutated__.base_mnasnet.DepthwiseConv',
'__mutated__.base_mnasnet.MobileConv'],
# 'se_ratio_options': [0, 0.25],
'skip_options': ['identity', 'no'],
'n_filter_options': [int(filter_size*x) for x in [0.75, 1.0, 1.25]],
'exp_ratio': exp_ratio,
'stride': stride,
'in_ch': depths[0] if count == 0 else None
})
layers.append(ph)
'''if conv == "mconv":
# MNASNet blocks: stacks of inverted residuals.
layers.append(_stack_inverted_residual(prev_depth, depth, ks, skip,
stride, exp_ratio, repeat, _BN_MOMENTUM))
else:
# Normal conv and depth-separated conv
layers += _stack_normal_conv(prev_depth, depth, ks, skip, conv == "dconv",
stride, repeat, _BN_MOMENTUM)'''
count += 1
if count >= 2:
break
layers += [
# Final mapping to classifier input.
nn.Conv2d(depths[7], 1280, 1, padding=0, stride=1, bias=False),
nn.BatchNorm2d(1280, momentum=_BN_MOMENTUM),
nn.ReLU(inplace=True),
]
self.layers = nn.Sequential(*layers)
self.classifier = nn.Sequential(nn.Dropout(p=dropout, inplace=True),
nn.Linear(1280, num_classes))
self._initialize_weights()
def __init__(self, kernel_size, in_ch, out_ch, skip, exp_ratio, stride):
super().__init__()
self.kernel_size = kernel_size
self.in_ch = in_ch
self.out_ch = out_ch
self.skip = skip
self.exp_ratio = exp_ratio
self.stride = stride
mid_ch = in_ch * exp_ratio
self.layers = nn.Sequential(
# Pointwise
nn.Conv2d(in_ch, mid_ch, 1, bias=False),
nn.BatchNorm2d(mid_ch, momentum=BN_MOMENTUM),
nn.ReLU(inplace=False),
# Depthwise
nn.Conv2d(mid_ch,
mid_ch,
kernel_size,
padding=(kernel_size - 1) // 2,
stride=stride,
groups=mid_ch,
bias=False),
nn.BatchNorm2d(mid_ch, momentum=BN_MOMENTUM),
nn.ReLU(inplace=False),
# Linear pointwise. Note that there's no activation.
nn.Conv2d(mid_ch, out_ch, 1, bias=False),
nn.BatchNorm2d(out_ch, momentum=BN_MOMENTUM))
def _stack_normal_conv(in_ch, out_ch, kernel_size, skip, dconv, stride, repeats, bn_momentum):
assert repeats >= 1
stack = []
for i in range(repeats):
s = stride if i == 0 else 1
if dconv:
modules = [
nn.Conv2d(in_ch, in_ch, kernel_size, padding=kernel_size // 2, stride=s, groups=in_ch, bias=False),
nn.BatchNorm2d(in_ch, momentum=bn_momentum),
nn.ReLU(inplace=True),
nn.Conv2d(in_ch, out_ch, 1, padding=0, stride=1, bias=False),
nn.BatchNorm2d(out_ch, momentum=bn_momentum)
]
else:
modules = [
nn.Conv2d(in_ch, out_ch, kernel_size, padding=kernel_size // 2, stride=s, bias=False),
nn.ReLU(inplace=True),
nn.BatchNorm2d(out_ch, momentum=bn_momentum)
]
if skip and in_ch == out_ch and s == 1:
# use different implementation for skip and noskip to align with pytorch
stack.append(_ResidualBlock(nn.Sequential(*modules)))
else:
stack += modules
in_ch = out_ch
return stack
def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
super().__init__()
self.net = nn.Sequential(
nn.ReLU(),
nn.Conv2d(C_in, C_out, kernel_size, stride, padding, bias=False),
nn.BatchNorm2d(C_out, affine=affine)
)
def __init__(self,
in_ch,
out_ch,
kernel_size,
stride,
expansion_factor,
skip,
bn_momentum=0.1):
super(_InvertedResidual, self).__init__()
assert stride in [1, 2]
assert kernel_size in [3, 5]
mid_ch = in_ch * expansion_factor
self.apply_residual = skip and in_ch == out_ch and stride == 1
self.layers = nn.Sequential(
# Pointwise
nn.Conv2d(in_ch, mid_ch, 1, bias=False),
nn.BatchNorm2d(mid_ch, momentum=bn_momentum),
nn.ReLU(inplace=False),
# Depthwise
nn.Conv2d(mid_ch,
mid_ch,
kernel_size,
padding=kernel_size // 2,
stride=stride,
groups=mid_ch,
bias=False),
nn.BatchNorm2d(mid_ch, momentum=bn_momentum),
nn.ReLU(inplace=False),
# Linear pointwise. Note that there's no activation.
nn.Conv2d(mid_ch, out_ch, 1, bias=False),
nn.BatchNorm2d(out_ch, momentum=bn_momentum))
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion,
stride),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def __init__(self,
input_size=224,
first_conv_channels=16,
last_conv_channels=1024,
n_classes=1000,
affine=False):
super().__init__()
assert input_size % 32 == 0
self.stage_blocks = [4, 4, 8, 4]
self.stage_channels = [64, 160, 320, 640]
self._input_size = input_size
self._feature_map_size = input_size
self._first_conv_channels = first_conv_channels
self._last_conv_channels = last_conv_channels
self._n_classes = n_classes
self._affine = affine
self._layerchoice_count = 0
# building first layer
self.first_conv = nn.Sequential(
nn.Conv2d(3, first_conv_channels, 3, 2, 1, bias=False),
nn.BatchNorm2d(first_conv_channels, affine=affine),
nn.ReLU(inplace=True),
)
self._feature_map_size //= 2
p_channels = first_conv_channels
features = []
for num_blocks, channels in zip(self.stage_blocks,
self.stage_channels):
features.extend(self._make_blocks(num_blocks, p_channels,
channels))
p_channels = channels
self.features = nn.Sequential(*features)
self.conv_last = nn.Sequential(
nn.Conv2d(p_channels, last_conv_channels, 1, 1, 0, bias=False),
nn.BatchNorm2d(last_conv_channels, affine=affine),
nn.ReLU(inplace=True),
)
self.globalpool = nn.AvgPool2d(self._feature_map_size)
self.dropout = nn.Dropout(0.1)
self.classifier = nn.Sequential(
nn.Linear(last_conv_channels, n_classes, bias=False), )
self._initialize_weights()
def _stack_inverted_residual(in_ch, out_ch, kernel_size, skip, stride, exp_factor, repeats, bn_momentum):
""" Creates a stack of inverted residuals. """
assert repeats >= 1
# First one has no skip, because feature map size changes.
first = _InvertedResidual(in_ch, out_ch, kernel_size, stride, exp_factor, skip, bn_momentum=bn_momentum)
remaining = []
for _ in range(1, repeats):
remaining.append(_InvertedResidual(out_ch, out_ch, kernel_size, 1, exp_factor, skip, bn_momentum=bn_momentum))
return nn.Sequential(first, *remaining)
def __init__(self, C_in, C_out, kernel_size, stride, padding, dilation, affine=True):
super().__init__()
self.net = nn.Sequential(
nn.ReLU(),
nn.Conv2d(C_in, C_in, kernel_size, stride, padding, dilation=dilation, groups=C_in,
bias=False),
nn.Conv2d(C_in, C_out, 1, stride=1, padding=0, bias=False),
nn.BatchNorm2d(C_out, affine=affine)
)
def __init__(self,
in_channels: int,
out_channels: int,
mid_channels: int,
*,
kernel_size: int,
stride: int,
sequence: str = "pdp",
affine: bool = True):
super().__init__()
assert stride in [1, 2]
assert kernel_size in [3, 5, 7]
self.channels = in_channels // 2 if stride == 1 else in_channels
self.in_channels = in_channels
self.out_channels = out_channels
self.mid_channels = mid_channels
self.kernel_size = kernel_size
self.stride = stride
self.pad = kernel_size // 2
self.oup_main = out_channels - self.channels
self.affine = affine
assert self.oup_main > 0
self.branch_main = nn.Sequential(
*self._decode_point_depth_conv(sequence))
if stride == 2:
self.branch_proj = nn.Sequential(
# dw
nn.Conv2d(self.channels,
self.channels,
kernel_size,
stride,
self.pad,
groups=self.channels,
bias=False),
nn.BatchNorm2d(self.channels, affine=affine),
# pw-linear
nn.Conv2d(self.channels, self.channels, 1, 1, 0, bias=False),
nn.BatchNorm2d(self.channels, affine=affine),
nn.ReLU(inplace=True))
else:
# empty block to be compatible with torchscript
self.branch_proj = nn.Sequential()
def __init__(self):
super(LargeModel, self).__init__()
dim = 15
n = 4 * 100
self.emb = nn.Embedding(n, dim)
self.lin1 = nn.Linear(dim, 1)
self.seq = nn.Sequential(
self.emb,
self.lin1,
)
def __init__(self,
in_features,
out_features,
bias=True,
use_bn=False,
act_func=None,
dropout_rate=0,
ops_order='weight_bn_act'):
super(LinearLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.bias = bias
self.use_bn = use_bn
self.act_func = act_func
self.dropout_rate = dropout_rate
self.ops_order = ops_order
""" modules """
modules = {}
# batch norm
if self.use_bn:
if self.bn_before_weight:
modules['bn'] = nn.BatchNorm1d(in_features)
else:
modules['bn'] = nn.BatchNorm1d(out_features)
else:
modules['bn'] = None
# activation
modules['act'] = build_activation(self.act_func,
self.ops_list[0] != 'act')
# dropout
if self.dropout_rate > 0:
modules['dropout'] = nn.Dropout(self.dropout_rate, inplace=True)
else:
modules['dropout'] = None
# linear
modules['weight'] = {
'linear': nn.Linear(self.in_features, self.out_features, self.bias)
}
# add modules
for op in self.ops_list:
if modules[op] is None:
continue
elif op == 'weight':
if modules['dropout'] is not None:
self.add_module('dropout', modules['dropout'])
for key in modules['weight']:
self.add_module(key, modules['weight'][key])
else:
self.add_module(op, modules[op])
self.sequence = nn.Sequential(self._modules)
def __init__(self,
channels: int,
reduction: int = 4,
activation_layer: Optional[Callable[..., nn.Module]] = None):
super().__init__()
if activation_layer is None:
activation_layer = nn.Sigmoid
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
nn.Linear(channels, make_divisible(channels // reduction, 8)),
nn.ReLU(inplace=True),
nn.Linear(make_divisible(channels // reduction, 8), channels),
activation_layer())
def __init__(self, input_size, C, n_classes):
""" assuming input size 7x7 or 8x8 """
assert input_size in [7, 8]
super().__init__()
self.net = nn.Sequential(
nn.ReLU(inplace=True),
nn.AvgPool2d(5, stride=input_size - 5, padding=0, count_include_pad=False), # 2x2 out
nn.Conv2d(C, 128, kernel_size=1, bias=False),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 768, kernel_size=2, bias=False), # 1x1 out
nn.BatchNorm2d(768),
nn.ReLU(inplace=True)
)
self.linear = nn.Linear(768, n_classes)
def __init__(self,
in_channels,
out_channels,
use_bn=True,
act_func='relu',
dropout_rate=0,
ops_order='weight_bn_act'):
super(Base2DLayer, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.use_bn = use_bn
self.act_func = act_func
self.dropout_rate = dropout_rate
self.ops_order = ops_order
""" modules """
modules = {}
# batch norm
if self.use_bn:
if self.bn_before_weight:
modules['bn'] = nn.BatchNorm2d(in_channels)
else:
modules['bn'] = nn.BatchNorm2d(out_channels)
else:
modules['bn'] = None
# activation
modules['act'] = build_activation(self.act_func,
self.ops_list[0] != 'act')
# dropout
if self.dropout_rate > 0:
modules['dropout'] = nn.Dropout2d(self.dropout_rate, inplace=True)
else:
modules['dropout'] = None
# weight
modules['weight'] = self.weight_op()
# add modules
for op in self.ops_list:
if modules[op] is None:
continue
elif op == 'weight':
if modules['dropout'] is not None:
self.add_module('dropout', modules['dropout'])
for key in modules['weight']:
self.add_module(key, modules['weight'][key])
else:
self.add_module(op, modules[op])
self.sequence = nn.Sequential(self._modules)
def __init__(self, num_labels: int = 1000,
base_widths: Tuple[int, ...] = (32, 16, 32, 40, 80, 96, 192, 320, 1280),
dropout_rate: float = 0.,
width_mult: float = 1.0,
bn_eps: float = 1e-3,
bn_momentum: float = 0.1):
super().__init__()
assert len(base_widths) == 9
# include the last stage info widths here
widths = [make_divisible(width * width_mult, 8) for width in base_widths]
downsamples = [True, False, True, True, True, False, True, False]
self.num_labels = num_labels
self.dropout_rate = dropout_rate
self.bn_eps = bn_eps
self.bn_momentum = bn_momentum
self.stem = ConvBNReLU(3, widths[0], stride=2, norm_layer=nn.BatchNorm2d)
blocks: List[nn.Module] = [
# first stage is fixed
DepthwiseSeparableConv(widths[0], widths[1], kernel_size=3, stride=1)
]
# https://github.com/ultmaster/AceNAS/blob/46c8895fd8a05ffbc61a6b44f1e813f64b4f66b7/searchspace/proxylessnas/__init__.py#L21
for stage in range(2, 8):
# Rather than returning a fixed module here,
# we return a builder that dynamically creates module for different `repeat_idx`.
builder = inverted_residual_choice_builder(
[3, 6], [3, 5, 7], downsamples[stage], widths[stage - 1], widths[stage], f's{stage}')
if stage < 7:
blocks.append(nn.Repeat(builder, (1, 4), label=f's{stage}_depth'))
else:
# No mutation for depth in the last stage.
# Directly call builder to initiate one block
blocks.append(builder(0))
self.blocks = nn.Sequential(*blocks)
# final layers
self.feature_mix_layer = ConvBNReLU(widths[7], widths[8], kernel_size=1, norm_layer=nn.BatchNorm2d)
self.global_avg_pooling = nn.AdaptiveAvgPool2d(1)
self.dropout_layer = nn.Dropout(dropout_rate)
self.classifier = nn.Linear(widths[-1], num_labels)
reset_parameters(self, bn_momentum=bn_momentum, bn_eps=bn_eps)
def __init__(self, C: int, num_labels: int, dataset: Literal['imagenet',
'cifar']):
super().__init__()
if dataset == 'imagenet':
# assuming input size 14x14
stride = 2
elif dataset == 'cifar':
stride = 3
self.features = nn.Sequential(
nn.ReLU(inplace=True),
nn.AvgPool2d(5, stride=stride, padding=0, count_include_pad=False),
nn.Conv2d(C, 128, 1, bias=False), nn.BatchNorm2d(128),
nn.ReLU(inplace=True), nn.Conv2d(128, 768, 2, bias=False),
nn.BatchNorm2d(768), nn.ReLU(inplace=True))
self.classifier = nn.Linear(768, num_labels)
def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
super().__init__()
self.net = nn.Sequential(
DilConv(C_in,
C_in,
kernel_size,
stride,
padding,
dilation=1,
affine=affine),
DilConv(C_in,
C_out,
kernel_size,
1,
padding,
dilation=1,
affine=affine))
def __init__(self,
input_size,
in_channels,
channels,
n_classes,
n_layers,
n_nodes=4,
stem_multiplier=3,
auxiliary=False):
super().__init__()
self.in_channels = in_channels
self.channels = channels
self.n_classes = n_classes
self.n_layers = n_layers
self.aux_pos = 2 * n_layers // 3 if auxiliary else -1
c_cur = stem_multiplier * self.channels
self.stem = nn.Sequential(
nn.Conv2d(in_channels, c_cur, 3, 1, 1, bias=False),
nn.BatchNorm2d(c_cur))
# for the first cell, stem is used for both s0 and s1
# [!] channels_pp and channels_p is output channel size, but c_cur is input channel size.
channels_pp, channels_p, c_cur = c_cur, c_cur, channels
self.cells = nn.ModuleList()
reduction_p, reduction = False, False
for i in range(n_layers):
reduction_p, reduction = reduction, False
# Reduce featuremap size and double channels in 1/3 and 2/3 layer.
if i in [n_layers // 3, 2 * n_layers // 3]:
c_cur *= 2
reduction = True
cell = Cell(n_nodes, channels_pp, channels_p, c_cur, reduction_p,
reduction)
self.cells.append(cell)
c_cur_out = c_cur * n_nodes
channels_pp, channels_p = channels_p, c_cur_out
#if i == self.aux_pos:
# self.aux_head = AuxiliaryHead(input_size // 4, channels_p, n_classes)
self.gap = nn.AdaptiveAvgPool2d(1)
self.linear = nn.Linear(channels_p, n_classes)
def __init__(self, config):
super(SNLIClassifier, self).__init__()
self.config = config
self.embed = nn.Embedding(config.n_embed, config.d_embed)
self.projection = Linear(config.d_embed, config.d_proj)
self.encoder = Encoder(config)
self.dropout = nn.Dropout(p=config.dp_ratio)
self.relu = nn.ReLU()
seq_in_size = 2 * config.d_hidden
if self.config.birnn:
seq_in_size *= 2
lin_config = [seq_in_size] * 2
self.out = nn.Sequential(Linear(*lin_config), self.relu,
self.dropout, Linear(*lin_config),
self.relu, self.dropout,
Linear(*lin_config), self.relu,
self.dropout,
Linear(seq_in_size, config.d_out))
def __init__(self, config):
super(SNLIClassifier, self).__init__()
self.embed = nn.Embedding(config["n_embed"], config["d_embed"])
self.projection = Linear(config["d_embed"], config["d_proj"])
self.encoder = Encoder(config)
self.dropout = nn.Dropout(p=config["dp_ratio"])
self.relu = nn.ReLU()
seq_in_size = 2 * config["d_hidden"]
if config["birnn"]:
seq_in_size *= 2
lin_config = [seq_in_size] * 2
self.out = nn.Sequential(Linear(*lin_config), self.relu,
self.dropout, Linear(*lin_config),
self.relu, self.dropout,
Linear(*lin_config), self.relu,
self.dropout,
Linear(seq_in_size, config["d_out"]))
self.fix_emb = config["fix_emb"]
self.project = config["projection"]
def __init__(self,
C_in,
C_out,
kernel_length,
stride,
padding,
affine=True):
super(FacConv, self).__init__()
self.net = nn.Sequential(
nn.ReLU(),
nn.Conv2d(C_in,
C_in, (kernel_length, 1),
stride,
padding,
bias=False),
nn.Conv2d(C_in,
C_out, (1, kernel_length),
stride,
padding,
bias=False), nn.BatchNorm2d(C_out, affine=affine))
def __init__(self, nc, ndf):
super(DCGANDiscriminator, self).__init__()
self.main = nn.Sequential(
# input is (nc) x 64 x 64
nn.Conv2d(nc, ndf, 4, 2, 1, bias=False),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ndf) x 32 x 32
nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(ndf * 2),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ndf*2) x 16 x 16
nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(ndf * 4),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ndf*4) x 8 x 8
nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False),
nn.BatchNorm2d(ndf * 8),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ndf*8) x 4 x 4
nn.Conv2d(ndf * 8, 1, 4, 1, 0, bias=False),
nn.Sigmoid())
def __init__(self, nz, ngf, nc):
super(DCGANGenerator, self).__init__()
self.main = nn.Sequential(
# input is Z, going into a convolution
nn.ConvTranspose2d(nz, ngf * 8, 4, 1, 0, bias=False),
nn.BatchNorm2d(ngf * 8),
nn.ReLU(True),
# state size. (ngf*8) x 4 x 4
nn.ConvTranspose2d(ngf * 8, ngf * 4, 4, 2, 1, bias=False),
nn.BatchNorm2d(ngf * 4),
nn.ReLU(True),
# state size. (ngf*4) x 8 x 8
nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1, bias=False),
nn.BatchNorm2d(ngf * 2),
nn.ReLU(True),
# state size. (ngf*2) x 16 x 16
nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 1, bias=False),
nn.BatchNorm2d(ngf),
nn.ReLU(True),
# state size. (ngf) x 32 x 32
nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False),
nn.Tanh()
# state size. (nc) x 64 x 64
)
def __init__(self,
op_candidates: List[str],
merge_op: Literal['all', 'loose_end'] = 'all',
num_nodes_per_cell: int = 4,
width: Union[Tuple[int, ...], int] = 16,
num_cells: Union[Tuple[int, ...], int] = 20,
dataset: Literal['cifar', 'imagenet'] = 'imagenet',
auxiliary_loss: bool = False):
super().__init__()
self.dataset = dataset
self.num_labels = 10 if dataset == 'cifar' else 1000
self.auxiliary_loss = auxiliary_loss
# preprocess the specified width and depth
if isinstance(width, Iterable):
C = nn.ValueChoice(list(width), label='width')
else:
C = width
self.num_cells: nn.MaybeChoice[int] = cast(int, num_cells)
if isinstance(num_cells, Iterable):
self.num_cells = nn.ValueChoice(list(num_cells), label='depth')
num_cells_per_stage = [
(i + 1) * self.num_cells // 3 - i * self.num_cells // 3
for i in range(3)
]
# auxiliary head is different for network targetted at different datasets
if dataset == 'imagenet':
self.stem0 = nn.Sequential(
nn.Conv2d(3,
cast(int, C // 2),
kernel_size=3,
stride=2,
padding=1,
bias=False),
nn.BatchNorm2d(cast(int, C // 2)),
nn.ReLU(inplace=True),
nn.Conv2d(cast(int, C // 2),
cast(int, C),
3,
stride=2,
padding=1,
bias=False),
nn.BatchNorm2d(C),
)
self.stem1 = nn.Sequential(
nn.ReLU(inplace=True),
nn.Conv2d(cast(int, C),
cast(int, C),
3,
stride=2,
padding=1,
bias=False),
nn.BatchNorm2d(C),
)
C_pprev = C_prev = C_curr = C
last_cell_reduce = True
elif dataset == 'cifar':
self.stem = nn.Sequential(
nn.Conv2d(3, cast(int, 3 * C), 3, padding=1, bias=False),
nn.BatchNorm2d(cast(int, 3 * C)))
C_pprev = C_prev = 3 * C
C_curr = C
last_cell_reduce = False
else:
raise ValueError(f'Unsupported dataset: {dataset}')
self.stages = nn.ModuleList()
for stage_idx in range(3):
if stage_idx > 0:
C_curr *= 2
# For a stage, we get C_in, C_curr, and C_out.
# C_in is only used in the first cell.
# C_curr is number of channels for each operator in current stage.
# C_out is usually `C * num_nodes_per_cell` because of concat operator.
cell_builder = CellBuilder(op_candidates, C_pprev, C_prev, C_curr,
num_nodes_per_cell, merge_op,
stage_idx > 0, last_cell_reduce)
stage: Union[NDSStage, nn.Sequential] = NDSStage(
cell_builder, num_cells_per_stage[stage_idx])
if isinstance(stage, NDSStage):
stage.estimated_out_channels_prev = cast(int, C_prev)
stage.estimated_out_channels = cast(
int, C_curr * num_nodes_per_cell)
stage.downsampling = stage_idx > 0
self.stages.append(stage)
# NOTE: output_node_indices will be computed on-the-fly in trial code.
# When constructing model space, it's just all the nodes in the cell,
# which happens to be the case of one-shot supernet.
# C_pprev is output channel number of last second cell among all the cells already built.
if len(stage) > 1:
# Contains more than one cell
C_pprev = len(cast(nn.Cell,
stage[-2]).output_node_indices) * C_curr
else:
# Look up in the out channels of last stage.
C_pprev = C_prev
# This was originally,
# C_prev = num_nodes_per_cell * C_curr.
# but due to loose end, it becomes,
C_prev = len(cast(nn.Cell, stage[-1]).output_node_indices) * C_curr
# Useful in aligning the pprev and prev cell.
last_cell_reduce = cell_builder.last_cell_reduce
if stage_idx == 2:
C_to_auxiliary = C_prev
if auxiliary_loss:
assert isinstance(
self.stages[2], nn.Sequential
), 'Auxiliary loss can only be enabled in retrain mode.'
self.stages[2] = SequentialBreakdown(
cast(nn.Sequential, self.stages[2]))
self.auxiliary_head = AuxiliaryHead(
C_to_auxiliary, self.num_labels,
dataset=self.dataset) # type: ignore
self.global_pooling = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = nn.Linear(cast(int, C_prev), self.num_labels)
def __init__(self,
op_candidates: List[str],
merge_op: Literal['all', 'loose_end'] = 'all',
num_nodes_per_cell: int = 4,
width: Union[Tuple[int], int] = 16,
num_cells: Union[Tuple[int], int] = 20,
dataset: Literal['cifar', 'imagenet'] = 'imagenet',
auxiliary_loss: bool = False):
super().__init__()
self.dataset = dataset
self.num_labels = 10 if dataset == 'cifar' else 1000
self.auxiliary_loss = auxiliary_loss
# preprocess the specified width and depth
if isinstance(width, Iterable):
C = nn.ValueChoice(list(width), label='width')
else:
C = width
if isinstance(num_cells, Iterable):
num_cells = nn.ValueChoice(list(num_cells), label='depth')
num_cells_per_stage = [
i * num_cells // 3 - (i - 1) * num_cells // 3 for i in range(3)
]
# auxiliary head is different for network targetted at different datasets
if dataset == 'imagenet':
self.stem0 = nn.Sequential(
nn.Conv2d(3,
C // 2,
kernel_size=3,
stride=2,
padding=1,
bias=False),
nn.BatchNorm2d(C // 2),
nn.ReLU(inplace=True),
nn.Conv2d(C // 2, C, 3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(C),
)
self.stem1 = nn.Sequential(
nn.ReLU(inplace=True),
nn.Conv2d(C, C, 3, stride=2, padding=1, bias=False),
nn.BatchNorm2d(C),
)
C_pprev = C_prev = C_curr = C
last_cell_reduce = True
elif dataset == 'cifar':
self.stem = nn.Sequential(
nn.Conv2d(3, 3 * C, 3, padding=1, bias=False),
nn.BatchNorm2d(3 * C))
C_pprev = C_prev = 3 * C
C_curr = C
last_cell_reduce = False
self.stages = nn.ModuleList()
for stage_idx in range(3):
if stage_idx > 0:
C_curr *= 2
# For a stage, we get C_in, C_curr, and C_out.
# C_in is only used in the first cell.
# C_curr is number of channels for each operator in current stage.
# C_out is usually `C * num_nodes_per_cell` because of concat operator.
cell_builder = CellBuilder(op_candidates, C_pprev, C_prev, C_curr,
num_nodes_per_cell, merge_op,
stage_idx > 0, last_cell_reduce)
stage = nn.Repeat(cell_builder, num_cells_per_stage[stage_idx])
self.stages.append(stage)
# C_pprev is output channel number of last second cell among all the cells already built.
if len(stage) > 1:
# Contains more than one cell
C_pprev = len(stage[-2].output_node_indices) * C_curr
else:
# Look up in the out channels of last stage.
C_pprev = C_prev
# This was originally,
# C_prev = num_nodes_per_cell * C_curr.
# but due to loose end, it becomes,
C_prev = len(stage[-1].output_node_indices) * C_curr
# Useful in aligning the pprev and prev cell.
last_cell_reduce = cell_builder.last_cell_reduce
if stage_idx == 2:
C_to_auxiliary = C_prev
if auxiliary_loss:
assert isinstance(
self.stages[2], nn.Sequential
), 'Auxiliary loss can only be enabled in retrain mode.'
self.stages[2] = SequentialBreakdown(self.stages[2])
self.auxiliary_head = AuxiliaryHead(C_to_auxiliary,
self.num_labels,
dataset=self.dataset)
self.global_pooling = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = nn.Linear(C_prev, self.num_labels)
lambda C, stride, affine: nn.AvgPool2d(
5, stride=stride, padding=2, count_include_pad=False),
'max_pool_2x2':
lambda C, stride, affine: nn.MaxPool2d(2, stride=stride, padding=0),
'max_pool_3x3':
lambda C, stride, affine: nn.MaxPool2d(3, stride=stride, padding=1),
'max_pool_5x5':
lambda C, stride, affine: nn.MaxPool2d(5, stride=stride, padding=2),
'max_pool_7x7':
lambda C, stride, affine: nn.MaxPool2d(7, stride=stride, padding=3),
'skip_connect':
lambda C, stride, affine: nn.Identity()
if stride == 1 else FactorizedReduce(C, C, affine=affine),
'conv_1x1':
lambda C, stride, affine: nn.Sequential(
nn.ReLU(inplace=False),
nn.Conv2d(C, C, 1, stride=stride, padding=0, bias=False),
nn.BatchNorm2d(C, affine=affine)),
'conv_3x3':
lambda C, stride, affine: nn.Sequential(
nn.ReLU(inplace=False),
nn.Conv2d(C, C, 3, stride=stride, padding=1, bias=False),
nn.BatchNorm2d(C, affine=affine)),
'sep_conv_3x3':
lambda C, stride, affine: SepConv(C, C, 3, stride, 1, affine=affine),
'sep_conv_5x5':
lambda C, stride, affine: SepConv(C, C, 5, stride, 2, affine=affine),
'sep_conv_7x7':
lambda C, stride, affine: SepConv(C, C, 7, stride, 3, affine=affine),
'dil_conv_3x3':
lambda C, stride, affine: DilConv(C, C, 3, stride, 2, 2, affine=affine),
'dil_conv_5x5':
def __init__(self,
num_labels: int = 1000,
base_widths: Tuple[int, ...] = (16, 16, 32, 64, 128, 256, 512,
1024),
width_multipliers: Tuple[float, ...] = (0.5, 0.625, 0.75, 1.0,
1.25, 1.5, 2.0),
expand_ratios: Tuple[int, ...] = (1, 2, 3, 4, 5, 6),
dropout_rate: float = 0.2,
bn_eps: float = 1e-3,
bn_momentum: float = 0.1):
super().__init__()
self.widths = [
nn.ValueChoice([
make_divisible(base_width * mult, 8)
for mult in width_multipliers
],
label=f'width_{i}')
for i, base_width in enumerate(base_widths)
]
self.expand_ratios = expand_ratios
blocks = [
# Stem
ConvBNReLU(3,
self.widths[0],
nn.ValueChoice([3, 5], label='ks_0'),
stride=2,
activation_layer=h_swish),
SeparableConv(self.widths[0],
self.widths[0],
activation_layer=nn.ReLU),
]
# counting for kernel sizes and expand ratios
self.layer_count = 2
blocks += [
# Body
self._make_stage(1, self.widths[0], self.widths[1], False, 2,
nn.ReLU),
self._make_stage(2, self.widths[1], self.widths[2], True, 2,
nn.ReLU),
self._make_stage(1, self.widths[2], self.widths[3], False, 2,
h_swish),
self._make_stage(1, self.widths[3], self.widths[4], True, 1,
h_swish),
self._make_stage(1, self.widths[4], self.widths[5], True, 2,
h_swish),
]
# Head
blocks += [
ConvBNReLU(self.widths[5],
self.widths[6],
1,
1,
activation_layer=h_swish),
nn.AdaptiveAvgPool2d(1),
ConvBNReLU(self.widths[6],
self.widths[7],
1,
1,
norm_layer=nn.Identity,
activation_layer=h_swish),
]
self.blocks = nn.Sequential(*blocks)
self.classifier = nn.Sequential(
nn.Dropout(dropout_rate),
nn.Linear(self.widths[7], num_labels),
)
reset_parameters(self, bn_momentum=bn_momentum, bn_eps=bn_eps)
def __init__(self,
num_labels: int = 1000,
channel_search: bool = False,
affine: bool = False):
super().__init__()
self.num_labels = num_labels
self.channel_search = channel_search
self.affine = affine
# the block number in each stage. 4 stages in total. 20 blocks in total.
self.stage_repeats = [4, 4, 8, 4]
# output channels for all stages, including the very first layer and the very last layer
self.stage_out_channels = [-1, 16, 64, 160, 320, 640, 1024]
# building first layer
out_channels = self.stage_out_channels[1]
self.first_conv = nn.Sequential(
nn.Conv2d(3, out_channels, 3, 2, 1, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
)
self.features = []
global_block_idx = 0
for stage_idx, num_repeat in enumerate(self.stage_repeats):
for block_idx in range(num_repeat):
# count global index to give names to choices
global_block_idx += 1
# get ready for input and output
in_channels = out_channels
out_channels = self.stage_out_channels[stage_idx + 2]
stride = 2 if block_idx == 0 else 1
# mid channels can be searched
base_mid_channels = out_channels // 2
if self.channel_search:
k_choice_list = [
int(base_mid_channels * (.2 * k)) for k in range(1, 9)
]
mid_channels = nn.ValueChoice(
k_choice_list, label=f'channel_{global_block_idx}')
else:
mid_channels = int(base_mid_channels)
choice_block = nn.LayerChoice(
[
ShuffleNetBlock(in_channels,
out_channels,
mid_channels=mid_channels,
kernel_size=3,
stride=stride,
affine=affine),
ShuffleNetBlock(in_channels,
out_channels,
mid_channels=mid_channels,
kernel_size=5,
stride=stride,
affine=affine),
ShuffleNetBlock(in_channels,
out_channels,
mid_channels=mid_channels,
kernel_size=7,
stride=stride,
affine=affine),
ShuffleXceptionBlock(in_channels,
out_channels,
mid_channels=mid_channels,
stride=stride,
affine=affine)
],
label=f'layer_{global_block_idx}')
self.features.append(choice_block)
self.features = nn.Sequential(*self.features)
# final layers
last_conv_channels = self.stage_out_channels[-1]
self.conv_last = nn.Sequential(
nn.Conv2d(out_channels, last_conv_channels, 1, 1, 0, bias=False),
nn.BatchNorm2d(last_conv_channels, affine=affine),
nn.ReLU(inplace=True),
)
self.globalpool = nn.AdaptiveAvgPool2d((1, 1))
self.dropout = nn.Dropout(0.1)
self.classifier = nn.Sequential(
nn.Linear(last_conv_channels, num_labels, bias=False), )
self._initialize_weights()
