backbone.py

import torch.nn as nn
import torch
import torch.nn.functional as F
import pdb
import math
from utils import l2_norm
from torch.nn.utils.weight_norm import WeightNorm
from torch.distributions import Bernoulli
# This ResNet network was designed following the practice of the following papers:
# TADAM: Task dependent adaptive metric for improved few-shot learning (Oreshkin et al., in NIPS 2018) and
# A Simple Neural Attentive Meta-Learner (Mishra et al., in ICLR 2018).

def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)

class Flatten(nn.Module):
    def __init__(self):
        super(Flatten, self).__init__()
        
    def forward(self, x):        
        return x.view(x.size(0), -1)

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None, drop_rate=0.0, drop_block=False, block_size=1):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes)
        self.bn1 = nn.BatchNorm2d(planes)
        self.relu = nn.LeakyReLU(0.1)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = conv3x3(planes, planes)
        self.bn3 = nn.BatchNorm2d(planes)
        self.maxpool = nn.MaxPool2d(stride)
        self.downsample = downsample
        self.stride = stride
        self.drop_rate = drop_rate
        self.num_batches_tracked = 0
        self.drop_block = drop_block
        self.block_size = block_size
        self.DropBlock = DropBlock(block_size=self.block_size)

    def forward(self, x):
        self.num_batches_tracked += 1

        residual = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            residual = self.downsample(x)
        out += residual
        out = self.relu(out)
        out = self.maxpool(out)
        
        if self.drop_rate > 0:
            if self.drop_block == True:
                feat_size = out.size()[2]
                keep_rate = max(1.0 - self.drop_rate / (20*2000) * (self.num_batches_tracked), 1.0 - self.drop_rate)
                gamma = (1 - keep_rate) / self.block_size**2 * feat_size**2 / (feat_size - self.block_size + 1)**2
                out = self.DropBlock(out, gamma=gamma)
            else:
                out = F.dropout(out, p=self.drop_rate, training=self.training, inplace=True)

        return out

class DropBlock(nn.Module):
    def __init__(self, block_size):
        super(DropBlock, self).__init__()

        self.block_size = block_size
        #self.gamma = gamma
        #self.bernouli = Bernoulli(gamma)

    def forward(self, x, gamma):
        # shape: (bsize, channels, height, width)

        if self.training:
            batch_size, channels, height, width = x.shape
            
            bernoulli = Bernoulli(gamma)
            mask = bernoulli.sample((batch_size, channels, height - (self.block_size - 1), width - (self.block_size - 1))).cuda()
            #print((x.sample[-2], x.sample[-1]))
            block_mask = self._compute_block_mask(mask)
            #print (block_mask.size())
            #print (x.size())
            countM = block_mask.size()[0] * block_mask.size()[1] * block_mask.size()[2] * block_mask.size()[3]
            count_ones = block_mask.sum()

            return block_mask * x * (countM / count_ones)
        else:
            return x

    def _compute_block_mask(self, mask):
        left_padding = int((self.block_size-1) / 2)
        right_padding = int(self.block_size / 2)
        
        batch_size, channels, height, width = mask.shape
        #print ("mask", mask[0][0])
        non_zero_idxs = mask.nonzero()
        nr_blocks = non_zero_idxs.shape[0]

        offsets = torch.stack(
            [
                torch.arange(self.block_size).view(-1, 1).expand(self.block_size, self.block_size).reshape(-1), # - left_padding,
                torch.arange(self.block_size).repeat(self.block_size), #- left_padding
            ]
        ).t().cuda()
        offsets = torch.cat((torch.zeros(self.block_size**2, 2).cuda().long(), offsets.long()), 1)
        
        if nr_blocks > 0:
            non_zero_idxs = non_zero_idxs.repeat(self.block_size ** 2, 1)
            offsets = offsets.repeat(nr_blocks, 1).view(-1, 4)
            offsets = offsets.long()

            block_idxs = non_zero_idxs + offsets
            #block_idxs += left_padding
            padded_mask = F.pad(mask, (left_padding, right_padding, left_padding, right_padding))
            padded_mask[block_idxs[:, 0], block_idxs[:, 1], block_idxs[:, 2], block_idxs[:, 3]] = 1.
        else:
            padded_mask = F.pad(mask, (left_padding, right_padding, left_padding, right_padding))
            
        block_mask = 1 - padded_mask#[:height, :width]
        return block_mask


class ResNet(nn.Module):

    def __init__(self, block, keep_prob=1.0, avg_pool=False, drop_rate=0.0, dropblock_size=5, flatten=False):
        self.inplanes = 3
        super(ResNet, self).__init__()
        self.final_feat_dim = 640
        self.layer1 = self._make_layer(block, 64, stride=2, drop_rate=drop_rate)
        self.layer2 = self._make_layer(block, 160, stride=2, drop_rate=drop_rate)
        self.layer3 = self._make_layer(block, 320, stride=2, drop_rate=drop_rate, drop_block=True, block_size=dropblock_size)
        self.layer4 = self._make_layer(block, 640, stride=2, drop_rate=drop_rate, drop_block=True, block_size=dropblock_size)
        if avg_pool:
            self.avgpool = nn.Sequential(
                  nn.AvgPool2d(5, stride=1),
                  Flatten()
            )
        self.keep_prob = keep_prob
        self.flatten = flatten
        self.keep_avg_pool = avg_pool
        self.dropout = nn.Dropout(p=1 - self.keep_prob, inplace=False)
        self.drop_rate = drop_rate
        
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='leaky_relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, planes, stride=1, drop_rate=0.0, drop_block=False, block_size=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=1, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, drop_rate, drop_block, block_size))
        self.inplanes = planes * block.expansion

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        if self.keep_avg_pool:
            x = self.avgpool(x)
        return x

class ConvNet(nn.Module):
    def __init__(self, depth, flatten = True):
        super(ConvNet,self).__init__()
        trunk = []
        for i in range(depth):
            indim = 3 if i == 0 else 64
            outdim = 64
            B = ConvBlock(indim, outdim, pool = ( i <4 ) ) #only pooling for fist 4 layers
            trunk.append(B)

        if flatten:
            trunk.append(Flatten())

        self.trunk = nn.Sequential(*trunk)
        self.final_feat_dim = 640

    def forward(self,x):
        out = self.trunk(x)
        return out


def resnet12(keep_prob=1.0, avg_pool=False, flatten=True, **kwargs):
    """Constructs a ResNet-12 model.
    """
    model = ResNet(BasicBlock, keep_prob=keep_prob, drop_rate=0.1, dropblock_size=2, avg_pool=avg_pool, flatten=flatten)
    return model


def Conv4(avg_pool=True, flatten=True):
    return ConvNet(4, flatten=flatten)

class distLinear(nn.Module):
    def __init__(self, indim, outdim):
        super(distLinear, self).__init__()
        self.L = nn.Linear( indim, outdim, bias = False)
        self.class_wise_learnable_norm = True  #See the issue#4&8 in the github 
        if self.class_wise_learnable_norm:      
            WeightNorm.apply(self.L, 'weight', dim=0) #split the weight update component to direction and norm      

        if outdim <=200:
            self.scale_factor = 2; #a fixed scale factor to scale the output of cos value into a reasonably large input for softmax, for to reproduce the result of CUB with ResNet10, use 4. see the issue#31 in the github 
        else:
            self.scale_factor = 10; #in omniglot, a larger scale factor is required to handle >1000 output classes.

    def forward(self, x, y=None):
        x_norm = torch.norm(x, p=2, dim =1).unsqueeze(1).expand_as(x)
        x_normalized = x.div(x_norm+ 0.00001)
        if not self.class_wise_learnable_norm:
            L_norm = torch.norm(self.L.weight.data, p=2, dim =1).unsqueeze(1).expand_as(self.L.weight.data)
            self.L.weight.data = self.L.weight.data.div(L_norm + 0.00001)
        cos_dist = self.L(x_normalized) #matrix product by forward function, but when using WeightNorm, this also multiply the cosine distance by a class-wise learnable norm, see the issue#4&8 in the github
        scores = self.scale_factor* (cos_dist) 

        return scores


class NormLinear(nn.Module):
    # implementation of additive margin softmax loss in https://arxiv.org/abs/1801.05599    
    def __init__(self, embedding_size=512, classnum=51332, radius=10, pretrained=None):
        super(NormLinear, self).__init__()
        self.classnum = classnum
        self.s = radius
        # initial kernel
        self.weight = nn.Parameter(torch.Tensor(embedding_size, classnum))
        self.reset_parameters()

    def reset_parameters(self):
        stdv = 1./math.sqrt(self.weight.size(0))
        self.weight.data.uniform_(-stdv, stdv)

    def forward(self, embeddings, y=None):
        # weights norm
        nB = len(embeddings)
        x_len, embeddings = l2_norm(embeddings)
        kernel_len, kernel_norm = l2_norm(self.weight,axis=0)
        # cos(theta+m)
        cos_theta = torch.mm(embeddings,kernel_norm)
        theta = embeddings.mm(kernel_norm.detach())
        output = cos_theta * 1.0 # a little bit hacky way to prevent in_place operation on cos_theta

        if self.s:
            output*=self.s # scale up in order to make softmax work, first introduced in normface
        else:
            output*=x_len

       
        return output