def __init__(self):
super().__init__()
ch1 = ValueChoice([16, 32])
kernel = ValueChoice([3, 5])
self.conv1 = nn.Conv2d(1, ch1, kernel, padding=kernel // 2)
self.batch_norm = nn.BatchNorm2d(ch1)
self.conv2 = nn.Conv2d(ch1, 64, 3, padding=1)
self.dropout1 = LayerChoice(
[nn.Dropout(.25), nn.Dropout(.5),
nn.Dropout(.75)])
self.fc = nn.Linear(64, 10)
self.rpfc = nn.Repeat(nn.Linear(10, 10), (1, 4))
def _make_stage(self, stage_idx, inp, oup, se, stride, act):
# initialize them first because they are related to layer_count.
exp, ks, se_blocks = [], [], []
for _ in range(4):
exp.append(
nn.ValueChoice(list(self.expand_ratios),
label=f'exp_{self.layer_count}'))
ks.append(nn.ValueChoice([3, 5, 7],
label=f'ks_{self.layer_count}'))
if se:
# if SE is true, assign a layer choice to SE
se_blocks.append(lambda hidden_ch: nn.LayerChoice(
[nn.Identity(), SELayer(hidden_ch)],
label=f'se_{self.layer_count}'))
else:
se_blocks.append(None)
self.layer_count += 1
blocks = [
# stride = 2
InvertedResidual(inp,
oup,
exp[0],
ks[0],
stride,
squeeze_and_excite=se_blocks[0],
activation_layer=act),
# stride = 1, residual connection should be automatically enabled
InvertedResidual(oup,
oup,
exp[1],
ks[1],
squeeze_and_excite=se_blocks[1],
activation_layer=act),
InvertedResidual(oup,
oup,
exp[2],
ks[2],
squeeze_and_excite=se_blocks[2],
activation_layer=act),
InvertedResidual(oup,
oup,
exp[3],
ks[3],
squeeze_and_excite=se_blocks[3],
activation_layer=act)
]
# mutable depth
return nn.Repeat(blocks, depth=(1, 4), label=f'depth_{stage_idx}')
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):
super().__init__()
self.stem = nn.Conv2d(1, 5, 7, stride=4)
self.cells = nn.Repeat(
lambda index: nn.Cell(
{
'conv1':
lambda _, __, inp: nn.Conv2d(
(5 if index == 0 else 3 * 4)
if inp is not None and inp < 1 else 4, 4, 1),
'conv2':
lambda _, __, inp: nn.Conv2d(
(5 if index == 0 else 3 * 4)
if inp is not None and inp < 1 else 4,
4,
3,
padding=1),
},
3,
merge_op='loose_end'), (1, 3))
self.fc = nn.Linear(3 * 4, 10)
def __init__(self):
super().__init__()
self.block = nn.Repeat(AddOne(), (3, 5))
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
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(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):
super().__init__()
self.block = nn.Repeat(lambda index: nn.LayerChoice(
[AddOne(), nn.Identity()]), (2, 3),
label='rep')
def __init__(self):
super().__init__()
self.block = nn.Repeat(nn.LayerChoice(
[AddOne(), nn.Identity()], label='lc'), (3, 5),
label='rep')
def __init__(
self,
search_embed_dim: Tuple[int, ...] = (192, 216, 240),
search_mlp_ratio: Tuple[float, ...] = (3.5, 4.0),
search_num_heads: Tuple[int, ...] = (3, 4),
search_depth: Tuple[int, ...] = (12, 13, 14),
img_size: int = 224,
patch_size: int = 16,
in_chans: int = 3,
num_classes: int = 1000,
qkv_bias: bool = False,
drop_rate: float = 0.,
attn_drop_rate: float = 0.,
drop_path_rate: float = 0.,
pre_norm: bool = True,
global_pool: bool = False,
abs_pos: bool = True,
qk_scale: Optional[float] = None,
rpe: bool = True,
):
super().__init__()
embed_dim = nn.ValueChoice(list(search_embed_dim), label="embed_dim")
fixed_embed_dim = nn.ModelParameterChoice(
list(search_embed_dim), label="embed_dim")
depth = nn.ValueChoice(list(search_depth), label="depth")
self.patch_embed = nn.Conv2d(
in_chans,
cast(int, embed_dim),
kernel_size=patch_size,
stride=patch_size)
self.patches_num = int((img_size // patch_size) ** 2)
self.global_pool = global_pool
self.cls_token = nn.Parameter(torch.zeros(1, 1, cast(int, fixed_embed_dim)))
trunc_normal_(self.cls_token, std=.02)
dpr = [
x.item() for x in torch.linspace(
0,
drop_path_rate,
max(search_depth))] # stochastic depth decay rule
self.abs_pos = abs_pos
if self.abs_pos:
self.pos_embed = nn.Parameter(torch.zeros(
1, self.patches_num + 1, cast(int, fixed_embed_dim)))
trunc_normal_(self.pos_embed, std=.02)
self.blocks = nn.Repeat(lambda index: nn.LayerChoice([
TransformerEncoderLayer(embed_dim=embed_dim,
fixed_embed_dim=fixed_embed_dim,
num_heads=num_heads, mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, drop_rate=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[index],
rpe_length=img_size // patch_size,
qk_scale=qk_scale, rpe=rpe,
pre_norm=pre_norm,)
for mlp_ratio, num_heads in itertools.product(search_mlp_ratio, search_num_heads)
], label=f'layer{index}'), depth)
self.pre_norm = pre_norm
if self.pre_norm:
self.norm = nn.LayerNorm(cast(int, embed_dim))
self.head = nn.Linear(
cast(int, embed_dim),
num_classes) if num_classes > 0 else nn.Identity()
