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, C_in, C_out, affine=True): super(FactorizedReduce, self).__init__() self.relu = nn.ReLU() self.conv1 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False) self.conv2 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False) self.bn = nn.BatchNorm2d(C_out, affine=affine)
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 self.conv = nn.Conv2d(in_ch, out_ch, kernel_size, padding=kernel_size // 2, stride=stride, bias=False) self.relu = nn.ReLU(inplace=False) self.bn = nn.BatchNorm2d(out_ch, momentum=BN_MOMENTUM)
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, pool_type, C, kernel_size, stride, padding, affine=True): super().__init__() if pool_type.lower() == 'max': self.pool = nn.MaxPool2d(kernel_size, stride, padding) elif pool_type.lower() == 'avg': self.pool = nn.AvgPool2d(kernel_size, stride, padding, count_include_pad=False) else: raise ValueError() self.bn = nn.BatchNorm2d(C, affine=affine)
def __init__(self, C_in, C_out, kernel_length, stride, padding, affine=True): super().__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, 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, C_in, C_out, kernel_size, stride, padding, dilation, affine=True): super().__init__( nn.ReLU(inplace=False), nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=C_in, bias=False), nn.Conv2d(C_in, C_out, kernel_size=1, padding=0, bias=False), nn.BatchNorm2d(C_out, affine=affine), )
def __init__(self, C_in, C_out, affine=True): super().__init__() if isinstance(C_out, int): assert C_out % 2 == 0 else: # is a value choice assert all(c % 2 == 0 for c in C_out.all_options()) self.relu = nn.ReLU(inplace=False) self.conv_1 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False) self.conv_2 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False) self.bn = nn.BatchNorm2d(C_out, affine=affine) self.pad = nn.ConstantPad2d((0, 1, 0, 1), 0)
def __init__(self): super().__init__() self.m = nn.BatchNorm2d(2)
'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': lambda C, stride, affine: DilConv(C, C, 5, stride, 4, 2, affine=affine),
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)
def __init__(self): super(Fuse, self).__init__() self.conv = nn.Conv2d(3, 2, kernel_size=1, stride=2, padding=3, bias=False) self.bn = nn.BatchNorm2d(2)
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=False), ] 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=False), ] self.layers = nn.Sequential(*layers) self.classifier = nn.Sequential(nn.Dropout(p=dropout, inplace=False), nn.Linear(1280, num_classes)) self._initialize_weights()
def __init__(self): super().__init__( nn.Conv2d(3, 3, 1, 1, bias=False), nn.BatchNorm2d(3), nn.ReLU(inplace=False) )
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()
def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True): super(StdConv, self).__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, some_ch): super().__init__() self.some_ch = some_ch self.batch_norm = nn.BatchNorm2d(some_ch)
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): super().__init__() self.m = nn.BatchNorm2d(128, affine=False, momentum=0.3)