def __init__(self, builder: ConvBuilder, deps):
super(LeNet5, self).__init__()
self.bd = builder
stem = builder.Sequential()
stem.add_module(
'conv1',
builder.Conv2d(in_channels=1,
out_channels=LENET5_DEPS[0],
kernel_size=5,
bias=True))
stem.add_module('maxpool1', builder.Maxpool2d(kernel_size=2))
stem.add_module(
'conv2',
builder.Conv2d(in_channels=LENET5_DEPS[0],
out_channels=LENET5_DEPS[1],
kernel_size=5,
bias=True))
stem.add_module('maxpool2', builder.Maxpool2d(kernel_size=2))
self.stem = stem
self.flatten = builder.Flatten()
self.linear1 = builder.Linear(in_features=LENET5_DEPS[1] * 16,
out_features=LENET5_DEPS[2])
self.relu1 = builder.ReLU()
self.linear2 = builder.Linear(in_features=LENET5_DEPS[2],
out_features=10)
def __init__(self, builder:ConvBuilder):
super(LeNet300, self).__init__()
self.flatten = builder.Flatten()
self.linear1 = builder.Linear(in_features=28*28, out_features=300, bias=True)
self.relu1 = builder.ReLU()
self.linear2 = builder.Linear(in_features=300, out_features=100, bias=True)
self.relu2 = builder.ReLU()
self.linear3 = builder.Linear(in_features=100, out_features=10, bias=True)
def __init__(self, num_classes, builder: ConvBuilder, deps):
super(VCNet, self).__init__()
self.stem = _create_vgg_stem(builder=builder, deps=deps)
self.flatten = builder.Flatten()
self.linear1 = builder.IntermediateLinear(in_features=deps[12],
out_features=512)
self.relu = builder.ReLU()
self.linear2 = builder.Linear(in_features=512,
out_features=num_classes)
def __init__(self,
builder: ConvBuilder,
num_blocks,
num_classes=1000,
deps=None):
super(SBottleneckResNet, self).__init__()
# self.mean_tensor = torch.from_numpy(np.array([0.485, 0.456, 0.406])).reshape(1, 3, 1, 1).cuda().type(torch.cuda.FloatTensor)
# self.std_tensor = torch.from_numpy(np.array([0.229, 0.224, 0.225])).reshape(1, 3, 1, 1).cuda().type(torch.cuda.FloatTensor)
# self.mean_tensor = torch.from_numpy(np.array([0.406, 0.456, 0.485])).reshape(1, 3, 1, 1).cuda().type(
# torch.cuda.FloatTensor)
# self.std_tensor = torch.from_numpy(np.array([0.225, 0.224, 0.229])).reshape(1, 3, 1, 1).cuda().type(
# torch.cuda.FloatTensor)
# self.mean_tensor = torch.from_numpy(np.array([0.5, 0.5, 0.5])).reshape(1, 3, 1, 1).cuda().type(
# torch.cuda.FloatTensor)
# self.std_tensor = torch.from_numpy(np.array([0.5, 0.5, 0.5])).reshape(1, 3, 1, 1).cuda().type(
# torch.cuda.FloatTensor)
if deps is None:
if num_blocks == [3, 4, 6, 3]:
deps = RESNET50_ORIGIN_DEPS_FLATTENED
elif num_blocks == [3, 4, 23, 3]:
deps = resnet_bottleneck_origin_deps_flattened(101)
else:
raise ValueError('???')
self.conv1 = builder.Conv2dBNReLU(3,
deps[0],
kernel_size=7,
stride=2,
padding=3)
self.maxpool = builder.Maxpool2d(kernel_size=3, stride=2, padding=1)
# every stage has num_block * 3 + 1
nls = [n * 3 + 1 for n in num_blocks] # num layers in each stage
self.stage1 = ResNetBottleneckStage(builder=builder,
in_planes=deps[0],
stage_deps=deps[1:nls[0] + 1])
self.stage2 = ResNetBottleneckStage(builder=builder,
in_planes=deps[nls[0]],
stage_deps=deps[nls[0] + 1:nls[0] +
1 + nls[1]],
stride=2)
self.stage3 = ResNetBottleneckStage(
builder=builder,
in_planes=deps[nls[0] + nls[1]],
stage_deps=deps[nls[0] + nls[1] + 1:nls[0] + 1 + nls[1] + nls[2]],
stride=2)
self.stage4 = ResNetBottleneckStage(
builder=builder,
in_planes=deps[nls[0] + nls[1] + nls[2]],
stage_deps=deps[nls[0] + nls[1] + nls[2] + 1:nls[0] + 1 + nls[1] +
nls[2] + nls[3]],
stride=2)
self.gap = builder.GAP(kernel_size=7)
self.fc = builder.Linear(deps[-1], num_classes)
def __init__(self, builder:ConvBuilder, num_classes):
super(MobileV1CifarNet, self).__init__()
self.conv1 = builder.Conv2dBNReLU(in_channels=3, out_channels=16, kernel_size=3, stride=1, padding=1)
blocks = []
in_planes = cifar_cfg[0]
for x in cifar_cfg:
out_planes = x if isinstance(x, int) else x[0]
stride = 1 if isinstance(x, int) else x[1]
blocks.append(MobileV1Block(builder=builder, in_planes=in_planes, out_planes=out_planes, stride=stride))
in_planes = out_planes
self.stem = builder.Sequential(*blocks)
self.gap = builder.GAP(kernel_size=8)
self.linear = builder.Linear(cifar_cfg[-1], num_classes)
def __init__(self, builder:ConvBuilder, deps):
super(LeNet5BN, self).__init__()
self.bd = builder
stem = builder.Sequential()
stem.add_module('conv1', builder.Conv2dBNReLU(in_channels=1, out_channels=deps[0], kernel_size=5))
stem.add_module('maxpool1', builder.Maxpool2d(kernel_size=2))
stem.add_module('conv2', builder.Conv2dBNReLU(in_channels=deps[0], out_channels=deps[1], kernel_size=5))
stem.add_module('maxpool2', builder.Maxpool2d(kernel_size=2))
self.stem = stem
self.flatten = builder.Flatten()
self.linear1 = builder.IntermediateLinear(in_features=deps[1] * 16, out_features=500)
self.relu1 = builder.ReLU()
self.linear2 = builder.Linear(in_features=500, out_features=10)
def __init__(self,
conv_idx,
builder: ConvBuilder,
preced_layer_idx,
in_features,
out_features,
bias=True):
super(AOFPFCReluLayer, self).__init__()
self.conv_idx = conv_idx
self.base_path = builder.Linear(in_features=in_features,
out_features=out_features,
bias=bias)
self.relu = builder.ReLU()
self.register_buffer('t_value', torch.zeros(1))
self.preced_layer_idx = preced_layer_idx
def __init__(self, block_counts, num_classes, builder:ConvBuilder, deps, use_dropout):
super(WRNCifarNet, self).__init__()
self.bd = builder
converted_deps = wrn_convert_flattened_deps(deps)
print('the converted deps is ', converted_deps)
self.conv1 = builder.Conv2d(in_channels=3, out_channels=converted_deps[0], kernel_size=3, stride=1, padding=1, bias=False)
self.stage1 = self._build_wrn_stage(num_blocks=block_counts[0], stage_input_channels=converted_deps[0],
stage_deps=converted_deps[1], downsample=False, use_dropout=use_dropout)
self.stage2 = self._build_wrn_stage(num_blocks=block_counts[1], stage_input_channels=converted_deps[1][-1][1],
stage_deps=converted_deps[2], downsample=True, use_dropout=use_dropout)
self.stage3 = self._build_wrn_stage(num_blocks=block_counts[2], stage_input_channels=converted_deps[2][-1][1],
stage_deps=converted_deps[3], downsample=True, use_dropout=use_dropout)
self.last_bn = builder.BatchNorm2d(num_features=converted_deps[3][-1][1])
self.linear = builder.Linear(in_features=converted_deps[3][-1][1], out_features=num_classes)
def __init__(self, builder:ConvBuilder, deps=SIMPLE_ALEXNET_DEPS):
super(AlexBN, self).__init__()
# self.bd = builder
stem = builder.Sequential()
stem.add_module('conv1', builder.Conv2dBNReLU(in_channels=3, out_channels=deps[0], kernel_size=11, stride=4, padding=2))
stem.add_module('maxpool1', builder.Maxpool2d(kernel_size=3, stride=2))
stem.add_module('conv2', builder.Conv2dBNReLU(in_channels=deps[0], out_channels=deps[1], kernel_size=5, padding=2))
stem.add_module('maxpool2', builder.Maxpool2d(kernel_size=3, stride=2))
stem.add_module('conv3',
builder.Conv2dBNReLU(in_channels=deps[1], out_channels=deps[2], kernel_size=3, padding=1))
stem.add_module('conv4',
builder.Conv2dBNReLU(in_channels=deps[2], out_channels=deps[3], kernel_size=3, padding=1))
stem.add_module('conv5',
builder.Conv2dBNReLU(in_channels=deps[3], out_channels=deps[4], kernel_size=3, padding=1))
stem.add_module('maxpool3', builder.Maxpool2d(kernel_size=3, stride=2))
self.stem = stem
self.flatten = builder.Flatten()
self.linear1 = builder.Linear(in_features=deps[4] * 6 * 6, out_features=4096)
self.relu1 = builder.ReLU()
self.drop1 = builder.Dropout(0.5)
self.linear2 = builder.Linear(in_features=4096, out_features=4096)
self.relu2 = builder.ReLU()
self.drop2 = builder.Dropout(0.5)
self.linear3 = builder.Linear(in_features=4096, out_features=1000)
def __init__(self, builder:ConvBuilder, num_classes, deps=None):
super(MobileV1ImagenetNet, self).__init__()
if deps is None:
deps = MI1_ORIGIN_DEPS
assert len(deps) == 27
self.conv1 = builder.Conv2dBNReLU(in_channels=3, out_channels=deps[0], kernel_size=3, stride=2, padding=1)
blocks = []
for block_idx in range(13):
depthwise_channels = int(deps[block_idx * 2 + 1])
pointwise_channels = int(deps[block_idx * 2 + 2])
stride = 2 if block_idx in [1, 3, 5, 11] else 1
blocks.append(MobileV1Block(builder=builder, in_planes=depthwise_channels, out_planes=pointwise_channels, stride=stride))
self.stem = builder.Sequential(*blocks)
self.gap = builder.GAP(kernel_size=7)
self.linear = builder.Linear(imagenet_cfg[-1], num_classes)
def __init__(self,
builder: ConvBuilder,
num_blocks,
num_classes=1000,
deps=None):
super(SBottleneckResNet, self).__init__()
if deps is None:
if num_blocks == [3, 4, 6, 3]:
deps = RESNET50_ORIGIN_DEPS_FLATTENED
elif num_blocks == [3, 4, 23, 3]:
deps = resnet_bottleneck_origin_deps_flattened(101)
else:
raise ValueError('???')
self.conv1 = builder.Conv2dBNReLU(3,
deps[0],
kernel_size=7,
stride=2,
padding=3)
self.maxpool = builder.Maxpool2d(kernel_size=3, stride=2, padding=1)
# every stage has num_block * 3 + 1
nls = [n * 3 + 1 for n in num_blocks] # num layers in each stage
self.stage1 = ResNetBottleneckStage(builder=builder,
in_planes=deps[0],
stage_deps=deps[1:nls[0] + 1])
self.stage2 = ResNetBottleneckStage(builder=builder,
in_planes=deps[nls[0]],
stage_deps=deps[nls[0] + 1:nls[0] +
1 + nls[1]],
stride=2)
self.stage3 = ResNetBottleneckStage(
builder=builder,
in_planes=deps[nls[0] + nls[1]],
stage_deps=deps[nls[0] + nls[1] + 1:nls[0] + 1 + nls[1] + nls[2]],
stride=2)
self.stage4 = ResNetBottleneckStage(
builder=builder,
in_planes=deps[nls[0] + nls[1] + nls[2]],
stage_deps=deps[nls[0] + nls[1] + nls[2] + 1:nls[0] + 1 + nls[1] +
nls[2] + nls[3]],
stride=2)
self.gap = builder.GAP(kernel_size=7)
self.fc = builder.Linear(deps[-1], num_classes)
def __init__(self, num_classes, builder: ConvBuilder, deps):
super(VANet, self).__init__()
sq = builder.Sequential()
sq.add_module(
'conv1',
builder.Conv2dBNReLU(in_channels=3,
out_channels=deps[0],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv2',
builder.Conv2dBNReLU(in_channels=deps[0],
out_channels=deps[1],
kernel_size=3,
stride=1,
padding=1))
sq.add_module('maxpool1', builder.Maxpool2d(kernel_size=2))
sq.add_module(
'conv3',
builder.Conv2dBNReLU(in_channels=deps[1],
out_channels=deps[2],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv4',
builder.Conv2dBNReLU(in_channels=deps[2],
out_channels=deps[3],
kernel_size=3,
stride=1,
padding=1))
sq.add_module('maxpool2', builder.Maxpool2d(kernel_size=2))
sq.add_module(
'conv5',
builder.Conv2dBNReLU(in_channels=deps[3],
out_channels=deps[4],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv6',
builder.Conv2dBNReLU(in_channels=deps[4],
out_channels=deps[5],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv7',
builder.Conv2dBNReLU(in_channels=deps[5],
out_channels=deps[6],
kernel_size=3,
stride=1,
padding=1))
sq.add_module('maxpool3', builder.Maxpool2d(kernel_size=2))
sq.add_module(
'conv8',
builder.Conv2dBNReLU(in_channels=deps[6],
out_channels=deps[7],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv9',
builder.Conv2dBNReLU(in_channels=deps[7],
out_channels=deps[8],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv10',
builder.Conv2dBNReLU(in_channels=deps[8],
out_channels=deps[9],
kernel_size=3,
stride=1,
padding=1))
sq.add_module('maxpool4', builder.Maxpool2d(kernel_size=2))
sq.add_module(
'conv11',
builder.Conv2dBNReLU(in_channels=deps[9],
out_channels=deps[10],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv12',
builder.Conv2dBNReLU(in_channels=deps[10],
out_channels=deps[11],
kernel_size=3,
stride=1,
padding=1))
sq.add_module(
'conv13',
builder.Conv2dBNReLU(in_channels=deps[11],
out_channels=deps[12],
kernel_size=3,
stride=1,
padding=1))
sq.add_module('maxpool5', builder.Maxpool2d(kernel_size=2))
self.stem = sq
self.flatten = builder.Flatten()
self.linear1 = builder.IntermediateLinear(in_features=deps[12],
out_features=512)
self.relu = builder.ReLU()
self.linear2 = builder.Linear(in_features=512,
out_features=num_classes)