Skip to content

Latest commit

 

History

History
344 lines (298 loc) · 11.7 KB

README.md

File metadata and controls

344 lines (298 loc) · 11.7 KB
Raw

Training-Tricks-for-Binarized-Neural-Networks

A collection of training tricks of binarized neural networks from previously published/pre-print work on binary networks. larq further provides an open-source deep learning library for training neural networks with extremely low precision weights and activations, such as Binarized Neural Networks (BNNs).

1. Modified ResNet Block Structure

class BinActiveF(torch.autograd.Function):
    def forward(self, input):
        self.save_for_backward(input)
        input = input.sign()
        return input

    def backward(self, grad_output):
        input, = self.saved_tensors
        grad_output[input.ge(1.0)] = 0.
        grad_output[input.le(-1.0)] = 0.
        return grad_output

class BinActive(nn.Module):
    def __init__(self, bin=True):
        super(BinActive, self).__init__()
        self.bin = bin
    def forward(self, x):
        if self.bin:
            x = BinActiveF()(x)
        else:
            x = F.relu(x, inplace=True)
        return x
        
class BasicBlock(nn.Module):
    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.bn = nn.BatchNorm2d(inplanes)
        self.ba = BinActive()
        self.conv = nn.conv2d(inplanes, planes, 3, stride)
        self.prelu = nn.PReLU(planes)
        self.downsample = downsample

    def forward(self, x):
        residual = x

        out = self.bn(x)
        out = self.ba(out)
        out = self.conv(out)
        
        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        out = self.prelu(out)

        return out

class Bottleneck(nn.Module):
    def __init__(self, inplanes, planes, stride=1, has_branch=True):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.conv2d(inplanes, planes, kernel_size=1, stride=stride, padding=0)
        self.conv2 = nn.conv2d(planes, planes, kernel_size=3, stride=1, padding=1)
        self.conv3 = nn.conv2d(planes, planes*4, kernel_size=1, stride=1, padding=0)

        self.bn1 = nn.BatchNorm2d(inplanes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.bn3 = nn.BatchNorm2d(planes)
        self.bn4 = nn.BatchNorm2d(planes*4)
        
        self.ba1 = BinActive()

        self.has_branch = has_branch
        self.stride = stride

        if self.has_branch:
            if self.stride == 1:
                self.bn_bran1 = nn.Sequential(
                    nn.Conv2d(inplanes, planes*4, kernel_size=1, stride=1, padding=0, bias=False),
                    nn.BatchNorm2d(planes*4, eps=1e-4, momentum=0.1, affine=True),
                    nn.AvgPool2d(kernel_size=3, stride=1, padding=1))
                self.prelu = nn.PReLU(planes*4)
            else:
                self.branch1 = nn.Sequential(
                    nn.Conv2d(inplanes, inplanes*2, kernel_size=1, stride=1, padding=0, bias=False),
                    nn.BatchNorm2d(inplanes*2, eps=1e-4, momentum=0.1, affine=True),
                    nn.AvgPool2d(kernel_size=2, stride=2))
                self.prelu = nn.PReLU(inplanes*2)

    def forward(self, x):
        if self.stride == 2:
            short_cut = self.branch1(x)
        else:
            if self.has_branch:
                short_cut = self.bn_bran1(x)
            else:
                short_cut = x

        out = self.bn1(x)
        out = self.ba1(out)
        out = self.conv1(out)
        add = out
        out = self.bn2(out)

        out = self.ba1(out)
        out = self.conv2(out)
        out += add
        out = self.bn3(out)

        out = self.ba1(out)
        out = self.conv3(out)
        out = self.bn4(out)
        out += short_cut
        
        if self.has_branch:
            out = self.prelu(out)
        return out

2. PReLU Activation

Please refer to the above structures.

3. Double Skip Connections

Replace the original basic block in ResNet18 with two BasicBlock mentioned above.

4. Full Precision Downsampling Layers

# v1 (recommended)
downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, \
                                 stride=1, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
                nn.AvgPool2d(kernel_size=2, stride=2)
            )
# v2
downsample = nn.Sequential(
                nn.AvgPool2d(kernel_size=2, stride=2)
                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, \
                                 stride=1, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )
# v3
downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, \
                                 stride=1, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
                nn.AvgPool2d(kernel_size=3, stride=2, padd=1)
            )

5. Two-stage Training Strategy

  • Full-precision weights with binarized activations./ Full-precision activations with binarized weights.
  • Using the first stage model as initialization, then train 1-bit networks.

6. Weight Decay Setting

  • 1e-5 for stage 1.
  • 0.0 for stage 2.

7. Optimizer

  • Adam with the default beta settings.
optimizer = torch.optim.Adam(model.parameters(), args.lr, weight_decay=args.weight_decay)

8. Learning Rate

  • Two-stage setting:
    • 1e-3 for stage 1.
    • 2e-4 for stage 2.
    • CIFAR-100 : decay ratio 0.2 at 150th, 250th, 320th epochs. End at 350 epoch.
    • ImageNet : decay ratio 0.1 at 40th, 60th, 70th epochs. End at 75 epoch. 
      parser.add_argument('--schedule', type=int, nargs='+', default=[150, 225],
                    help='Decrease learning rate at these epochs.')
parser.add_argument('--gammas', type=float, nargs='+', default=[0.1, 0.1],
                    help='LR is multiplied by gamma on schedule, number of gammas should be equal to schedule')

def adjust_learning_rate(optimizer, epoch, gammas, schedule):
    """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
    lr = args.lr
    assert len(gammas) == len(schedule), "length of gammas and schedule should be equal"
    for (gamma, step) in zip(gammas, schedule):
        if (epoch >= step):
            lr = lr * gamma
        else:
            break
    print('learning rate : %.6f.' % lr)
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return lr
  • One-stage setting (~+0.5% Top-1 Acc. on ImageNet)
    • Warm-up 5 epochs with lr=0.001.
    • Increase lr to 0.004 for training based on Adam.

9. Data Augmentation

  • CIFAR-100: random crop, random horizontal flip, random rotation (+/-15 degree), mix-up/auto augmentation.
transforms.Compose([
            transforms.RandomCrop(32, padding=4),
            transforms.RandomHorizontalFlip(),
            transforms.RandomRotation(15),
            transforms.ToTensor(),
            normalize,
        ]
  • ImageNet: random crop, random flip, colour jitter (only in first stage, disabled for stage 2).
transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.1)
            transforms.ToTensor(),
            normalize,
        ]
  • ImageNet: Lighting (+~0.3% Top-1 on ImageNet).
#lighting data augmentation
imagenet_pca = {
    'eigval': np.asarray([0.2175, 0.0188, 0.0045]),
    'eigvec': np.asarray([
        [-0.5675, 0.7192, 0.4009],
        [-0.5808, -0.0045, -0.8140],
        [-0.5836, -0.6948, 0.4203],
    ])
}

class Lighting(object):
    def __init__(self, alphastd,
                 eigval=imagenet_pca['eigval'],
                 eigvec=imagenet_pca['eigvec']):
        self.alphastd = alphastd
        assert eigval.shape == (3,)
        assert eigvec.shape == (3, 3)
        self.eigval = eigval
        self.eigvec = eigvec

    def __call__(self, img):
        if self.alphastd == 0.:
            return img
        rnd = np.random.randn(3) * self.alphastd
        rnd = rnd.astype('float32')
        v = rnd
        old_dtype = np.asarray(img).dtype
        v = v * self.eigval
        v = v.reshape((3, 1))
        inc = np.dot(self.eigvec, v).reshape((3,))
        img = np.add(img, inc)
        if old_dtype == np.uint8:
            img = np.clip(img, 0, 255)
        img = Image.fromarray(img.astype(old_dtype), 'RGB')
        return img

    def __repr__(self):
        return self.__class__.__name__ + '()'

lighting_param = 0.1
train_transforms = transforms.Compose([
        transforms.RandomResizedCrop(224, scale=(0.08, 1.0)),
        Lighting(lighting_param),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        normalize])

10. Momentum in Batch Normalization Layers

  • Set momentum to 0.2 (marginal improvements to accuracy).
nn.BatchNorm2d(128, momentum=0.2, affine=True),

11. Reorder Pooling Block

From Conv+BN+ReLU+Pooling to Conv+Pooling+BN+ReLU.

12. Knowledge-distillation

  • KL divergence matching.
  • Feature-map matching after L2 normalization, e.g., equation.
  • Label refinery (recommended).

13. Channel-attention

x = BN(x)
out = x.sign()
out = conv(out)
out *= SE(x) # SE() generates [batchsize x C x 1 x 1] attention tensor
out = prelu(out)

where SE could be any channel attention module, such as SE-Net, CGD, CBAM, BAM, etc.

14. Auxiliary Loss Function

  • Center loss for stage 2 (marginal improvements).
  • Cross Entropy loss with labelsmooth (~+0.5% Top-1).
criterion_smooth = CrossEntropyLabelSmooth(num_classes=1000, epsilon=0.1).cuda()

class CrossEntropyLabelSmooth(nn.Module):

  def __init__(self, num_classes, epsilon):
    super(CrossEntropyLabelSmooth, self).__init__()
    self.num_classes = num_classes
    self.epsilon = epsilon
    self.logsoftmax = nn.LogSoftmax(dim=1)

  def forward(self, inputs, targets):
    log_probs = self.logsoftmax(inputs)
    targets = torch.zeros_like(log_probs).scatter_(1, targets.unsqueeze(1), 1)
    targets = (1 - self.epsilon) * targets + self.epsilon / self.num_classes
    loss = (-targets * log_probs).mean(0).sum()
    return loss

15. Double/Treble Channel Number

  • Using 3x3 group convolution layers to reduce BOPs.

16. Full-precision Pre-training

  • step 1. replace relu with the following leaky-clip
index = x.abs()>1.
x[index] = x[index]*0.1+x[index].sign()*0.9
  • step 2. replace leaky-clip with x.clamp_(-1,1)

17. Gradient Centralization

18. Image Scale Setting

train_transforms = transforms.Compose([
        transforms.Resize(256), # transforms.Resize(int(224*1.15)),
        transforms.RandomCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        normalize])
        
test_transforms = transforms.Compose([
        transforms.Resize(int(224*1.35)),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        normalize])

Cite:

If you find this repo useful, please cite

@misc{tricks4BNN,
  author =       {Shuan},
  title =        {Training-Tricks-for-Binarized-Neural-Networks},
  howpublished = {\url{https://github.com/HolmesShuan/Training-Tricks-for-Binarized-Neural-Networks}},
  year =         {2019}
}