methods/backbone.py

# This code is modified from https://github.com/facebookresearch/low-shot-shrink-hallucinate

import torch
import torch.nn as nn
import math
import torch.nn.functional as F
from torch.nn.utils import weight_norm

# --- gaussian initialize ---
def init_layer(L):
  # Initialization using fan-in
  if isinstance(L, nn.Conv2d):
    n = L.kernel_size[0]*L.kernel_size[1]*L.out_channels
    L.weight.data.normal_(0,math.sqrt(2.0/float(n)))
  elif isinstance(L, nn.BatchNorm2d):
    L.weight.data.fill_(1)
    L.bias.data.fill_(0)

class distLinear(nn.Module):
  def __init__(self, indim, outdim):
    super(distLinear, self).__init__()
    self.L = weight_norm(nn.Linear(indim, outdim, bias=False), name='weight', dim=0)
    self.relu = nn.ReLU()

  def forward(self, x):
    x_norm = torch.norm(x, p=2, dim =1).unsqueeze(1).expand_as(x)
    x_normalized = x.div(x_norm + 0.00001)
    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)
    scores = 10 * cos_dist
    return scores

# --- flatten tensor ---
class Flatten(nn.Module):
  def __init__(self):
    super(Flatten, self).__init__()

  def forward(self, x):
    return x.view(x.size(0), -1)

# --- LSTMCell module for matchingnet ---
class LSTMCell(nn.Module):
  maml = False
  def __init__(self, input_size, hidden_size, bias=True):
    super(LSTMCell, self).__init__()
    self.input_size = input_size
    self.hidden_size = hidden_size
    self.bias = bias
    if self.maml:
      self.x2h = Linear_fw(input_size, 4 * hidden_size, bias=bias)
      self.h2h = Linear_fw(hidden_size, 4 * hidden_size, bias=bias)
    else:
      self.x2h = nn.Linear(input_size, 4 * hidden_size, bias=bias)
      self.h2h = nn.Linear(hidden_size, 4 * hidden_size, bias=bias)
    self.reset_parameters()

  def reset_parameters(self):
    std = 1.0 / math.sqrt(self.hidden_size)
    for w in self.parameters():
      w.data.uniform_(-std, std)

  def forward(self, x, hidden=None):
    if hidden is None:
      hx = torch.zeors_like(x)
      cx = torch.zeros_like(x)
    else:
      hx, cx = hidden

    gates = self.x2h(x) + self.h2h(hx)
    ingate, forgetgate, cellgate, outgate = torch.split(gates, self.hidden_size, dim=1)

    ingate = torch.sigmoid(ingate)
    forgetgate = torch.sigmoid(forgetgate)
    cellgate = torch.tanh(cellgate)
    outgate = torch.sigmoid(outgate)

    cy = torch.mul(cx, forgetgate) +  torch.mul(ingate, cellgate)
    hy = torch.mul(outgate, torch.tanh(cy))
    return (hy, cy)

# --- LSTM module for matchingnet ---
class LSTM(nn.Module):
  def __init__(self, input_size, hidden_size, num_layers=1, bias=True, batch_first=False, bidirectional=False):
    super(LSTM, self).__init__()

    self.input_size = input_size
    self.hidden_size = hidden_size
    self.num_layers = num_layers
    self.bias = bias
    self.batch_first = batch_first
    self.num_directions = 2 if bidirectional else 1
    assert(self.num_layers == 1)

    self.lstm = LSTMCell(input_size, hidden_size, self.bias)

  def forward(self, x, hidden=None):
    # swap axis if batch first
    if self.batch_first:
      x = x.permute(1, 0 ,2)

    # hidden state
    if hidden is None:
      h0 = torch.zeros(self.num_directions, x.size(1), self.hidden_size, dtype=x.dtype, device=x.device)
      c0 = torch.zeros(self.num_directions, x.size(1), self.hidden_size, dtype=x.dtype, device=x.device)
    else:
      h0, c0 = hidden

    # forward
    outs = []
    hn = h0[0]
    cn = c0[0]
    for seq in range(x.size(0)):
      hn, cn = self.lstm(x[seq], (hn, cn))
      outs.append(hn.unsqueeze(0))
    outs = torch.cat(outs, dim=0)

    # reverse foward
    if self.num_directions == 2:
      outs_reverse = []
      hn = h0[1]
      cn = c0[1]
      for seq in range(x.size(0)):
        seq = x.size(1) - 1 - seq
        hn, cn = self.lstm(x[seq], (hn, cn))
        outs_reverse.append(hn.unsqueeze(0))
      outs_reverse = torch.cat(outs_reverse, dim=0)
      outs = torch.cat([outs, outs_reverse], dim=2)

    # swap axis if batch first
    if self.batch_first:
      outs = outs.permute(1, 0, 2)
    return outs

# --- Linear module ---
class Linear_fw(nn.Linear): #used in MAML to forward input with fast weight
  def __init__(self, in_features, out_features, bias=True):
    super(Linear_fw, self).__init__(in_features, out_features, bias=bias)
    self.weight.fast = None #Lazy hack to add fast weight link
    self.bias.fast = None

  def forward(self, x):
    if self.weight.fast is not None and self.bias.fast is not None:
      out = F.linear(x, self.weight.fast, self.bias.fast)
    else:
      out = super(Linear_fw, self).forward(x)
    return out

# --- Conv2d module ---
class Conv2d_fw(nn.Conv2d): #used in MAML to forward input with fast weight
  def __init__(self, in_channels, out_channels, kernel_size, stride=1,padding=0, bias = True):
    super(Conv2d_fw, self).__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, bias=bias)
    self.weight.fast = None
    if not self.bias is None:
      self.bias.fast = None

  def forward(self, x):
    if self.bias is None:
      if self.weight.fast is not None:
        out = F.conv2d(x, self.weight.fast, None, stride= self.stride, padding=self.padding)
      else:
        out = super(Conv2d_fw, self).forward(x)
    else:
      if self.weight.fast is not None and self.bias.fast is not None:
        out = F.conv2d(x, self.weight.fast, self.bias.fast, stride= self.stride, padding=self.padding)
      else:
        out = super(Conv2d_fw, self).forward(x)
    return out

# --- softplus module ---
def softplus(x):
  return torch.nn.functional.softplus(x, beta=100)

# --- feature-wise transformation layer ---
class FeatureWiseTransformation2d_fw(nn.BatchNorm2d):
  feature_augment = False
  def __init__(self, num_features, momentum=0.1, track_running_stats=True):
    super(FeatureWiseTransformation2d_fw, self).__init__(num_features, momentum=momentum, track_running_stats=track_running_stats)
    self.weight.fast = None
    self.bias.fast = None
    if self.track_running_stats:
      self.register_buffer('running_mean', torch.zeros(num_features))
      self.register_buffer('running_var', torch.zeros(num_features))
    if self.feature_augment: # initialize {gamma, beta} with {0.3, 0.5}
      self.gamma = torch.nn.Parameter(torch.ones(1, num_features, 1, 1)*0.3)
      self.beta  = torch.nn.Parameter(torch.ones(1, num_features, 1, 1)*0.5)
    self.reset_parameters()

  def reset_running_stats(self):
    if self.track_running_stats:
      self.running_mean.zero_()
      self.running_var.fill_(1)

  def forward(self, x, step=0):
    if self.weight.fast is not None and self.bias.fast is not None:
      weight = self.weight.fast
      bias = self.bias.fast
    else:
      weight = self.weight
      bias = self.bias
    if self.track_running_stats:
      out = F.batch_norm(x, self.running_mean, self.running_var, weight, bias, training=self.training, momentum=self.momentum)
    else:
      out = F.batch_norm(x, torch.zeros_like(x), torch.ones_like(x), weight, bias, training=True, momentum=1)

    # apply feature-wise transformation
    if self.feature_augment and self.training:
      gamma = (1 + torch.randn(1, self.num_features, 1, 1, dtype=self.gamma.dtype, device=self.gamma.device)*softplus(self.gamma)).expand_as(out)
      beta = (torch.randn(1, self.num_features, 1, 1, dtype=self.beta.dtype, device=self.beta.device)*softplus(self.beta)).expand_as(out)
      out = gamma*out + beta
    return out

# --- BatchNorm2d ---
class BatchNorm2d_fw(nn.BatchNorm2d):
  def __init__(self, num_features, momentum=0.1, track_running_stats=True):
    super(BatchNorm2d_fw, self).__init__(num_features, momentum=momentum, track_running_stats=track_running_stats)
    self.weight.fast = None
    self.bias.fast = None
    if self.track_running_stats:
      self.register_buffer('running_mean', torch.zeros(num_features))
      self.register_buffer('running_var', torch.zeros(num_features))
    self.reset_parameters()

  def reset_running_stats(self):
    if self.track_running_stats:
      self.running_mean.zero_()
      self.running_var.fill_(1)

  def forward(self, x, step=0):
    if self.weight.fast is not None and self.bias.fast is not None:
      weight = self.weight.fast
      bias = self.bias.fast
    else:
      weight = self.weight
      bias = self.bias
    if self.track_running_stats:
      out = F.batch_norm(x, self.running_mean, self.running_var, weight, bias, training=self.training, momentum=self.momentum)
    else:
      out = F.batch_norm(x, torch.zeros(x.size(1), dtype=x.dtype, device=x.device), torch.ones(x.size(1), dtype=x.dtype, device=x.device), weight, bias, training=True, momentum=1)
    return out

# --- BatchNorm1d ---
class BatchNorm1d_fw(nn.BatchNorm1d):
  def __init__(self, num_features, momentum=0.1, track_running_stats=True):
    super(BatchNorm1d_fw, self).__init__(num_features, momentum=momentum, track_running_stats=track_running_stats)
    self.weight.fast = None
    self.bias.fast = None
    if self.track_running_stats:
      self.register_buffer('running_mean', torch.zeros(num_features))
      self.register_buffer('running_var', torch.zeros(num_features))
    self.reset_parameters()

  def reset_running_stats(self):
    if self.track_running_stats:
      self.running_mean.zero_()
      self.running_var.fill_(1)

  def forward(self, x, step=0):
    if self.weight.fast is not None and self.bias.fast is not None:
      weight = self.weight.fast
      bias = self.bias.fast
    else:
      weight = self.weight
      bias = self.bias
    if self.track_running_stats:
      out = F.batch_norm(x, self.running_mean, self.running_var, weight, bias, training=self.training, momentum=self.momentum)
    else:
      out = F.batch_norm(x, torch.zeros(x.size(1), dtype=x.dtype, device=x.device), torch.ones(x.size(1), dtype=x.dtype, device=x.device), weight, bias, training=True, momentum=1)
    return out

# --- Simple Conv Block ---
class ConvBlock(nn.Module):
  maml = False
  def __init__(self, indim, outdim, pool = True, padding = 1):
    super(ConvBlock, self).__init__()
    self.indim  = indim
    self.outdim = outdim
    if self.maml:
      self.C = Conv2d_fw(indim, outdim, 3, padding = padding)
      self.BN = FeatureWiseTransformation2d_fw(outdim)
    else:
      self.C = nn.Conv2d(indim, outdim, 3, padding= padding)
      self.BN = nn.BatchNorm2d(outdim)
    self.relu = nn.ReLU(inplace=True)

    self.parametrized_layers = [self.C, self.BN, self.relu]
    if pool:
      self.pool = nn.MaxPool2d(2)
      self.parametrized_layers.append(self.pool)

    for layer in self.parametrized_layers:
      init_layer(layer)
    self.trunk = nn.Sequential(*self.parametrized_layers)

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

# --- Simple ResNet Block ---
class SimpleBlock(nn.Module):
  maml = False
  def __init__(self, indim, outdim, half_res, leaky=False):
    super(SimpleBlock, self).__init__()
    self.indim = indim
    self.outdim = outdim
    if self.maml:
      self.C1 = Conv2d_fw(indim, outdim, kernel_size=3, stride=2 if half_res else 1, padding=1, bias=False)
      self.BN1 = BatchNorm2d_fw(outdim)
      self.C2 = Conv2d_fw(outdim, outdim,kernel_size=3, padding=1,bias=False)
      self.BN2 = FeatureWiseTransformation2d_fw(outdim) # feature-wise transformation at the end of each residual block
    else:
      self.C1 = nn.Conv2d(indim, outdim, kernel_size=3, stride=2 if half_res else 1, padding=1, bias=False)
      self.BN1 = nn.BatchNorm2d(outdim)
      self.C2 = nn.Conv2d(outdim, outdim,kernel_size=3, padding=1,bias=False)
      self.BN2 = nn.BatchNorm2d(outdim)
    self.relu1 = nn.ReLU(inplace=True) if not leaky else nn.LeakyReLU(0.2, inplace=True)
    self.relu2 = nn.ReLU(inplace=True) if not leaky else nn.LeakyReLU(0.2, inplace=True)

    self.parametrized_layers = [self.C1, self.C2, self.BN1, self.BN2]

    self.half_res = half_res

    # if the input number of channels is not equal to the output, then need a 1x1 convolution
    if indim!=outdim:
      if self.maml:
        self.shortcut = Conv2d_fw(indim, outdim, 1, 2 if half_res else 1, bias=False)
        self.BNshortcut = FeatureWiseTransformation2d_fw(outdim)
      else:
        self.shortcut = nn.Conv2d(indim, outdim, 1, 2 if half_res else 1, bias=False)
        self.BNshortcut = nn.BatchNorm2d(outdim)

      self.parametrized_layers.append(self.shortcut)
      self.parametrized_layers.append(self.BNshortcut)
      self.shortcut_type = '1x1'
    else:
      self.shortcut_type = 'identity'

    for layer in self.parametrized_layers:
      init_layer(layer)

  def forward(self, x):
    out = self.C1(x)
    out = self.BN1(out)
    out = self.relu1(out)
    out = self.C2(out)
    out = self.BN2(out)
    short_out = x if self.shortcut_type == 'identity' else self.BNshortcut(self.shortcut(x))
    out = out + short_out
    out = self.relu2(out)
    return out

# --- ConvNet module ---
class ConvNet(nn.Module):
  def __init__(self, depth, flatten = True):
    super(ConvNet,self).__init__()
    self.grads = []
    self.fmaps = []
    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 = 1600

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

# --- ConvNetNopool module ---
class ConvNetNopool(nn.Module): #Relation net use a 4 layer conv with pooling in only first two layers, else no pooling
  def __init__(self, depth):
    super(ConvNetNopool,self).__init__()
    self.grads = []
    self.fmaps = []
    trunk = []
    for i in range(depth):
      indim = 3 if i == 0 else 64
      outdim = 64
      B = ConvBlock(indim, outdim, pool = ( i in [0,1] ), padding = 0 if i in[0,1] else 1  ) #only first two layer has pooling and no padding
      trunk.append(B)

    self.trunk = nn.Sequential(*trunk)
    self.final_feat_dim = [64,19,19]

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

# --- ResNet module ---
class ResNet(nn.Module):
  maml = False
  print('backbone:', 'maml:', maml)
  def __init__(self,block,list_of_num_layers, list_of_out_dims, flatten=True, leakyrelu=False):
    # list_of_num_layers specifies number of layers in each stage
    # list_of_out_dims specifies number of output channel for each stage
    super(ResNet,self).__init__()
    self.grads = []
    self.fmaps = []
    assert len(list_of_num_layers)==4, 'Can have only four stages'
    if self.maml:
      conv1 = Conv2d_fw(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
      bn1 = BatchNorm2d_fw(64)
    else:
      conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
      bn1 = nn.BatchNorm2d(64)

    relu = nn.ReLU(inplace=True) if not leakyrelu else nn.LeakyReLU(0.2, inplace=True)
    pool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

    init_layer(conv1)
    init_layer(bn1)

    trunk = [conv1, bn1, relu, pool1]

    indim = 64
    for i in range(4):
      for j in range(list_of_num_layers[i]):
        half_res = (i>=1) and (j==0)
        B = block(indim, list_of_out_dims[i], half_res, leaky=leakyrelu)
        trunk.append(B)
        indim = list_of_out_dims[i]

    if flatten:
      avgpool = nn.AvgPool2d(7)
      trunk.append(avgpool)
      trunk.append(Flatten())
      self.final_feat_dim = indim
    else:
      self.final_feat_dim = [ indim, 7, 7]

    self.trunk = nn.Sequential(*trunk)

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

# --- Conv networks ---
def Conv4():
    return ConvNet(4)
def Conv6():
    return ConvNet(6)
def Conv4NP():
    return ConvNetNopool(4)
def Conv6NP():
    return ConvNetNopool(6)

# --- ResNet networks ---
def ResNet10(flatten=True, leakyrelu=False):
    return ResNet(SimpleBlock, [1,1,1,1],[64,128,256,512], flatten, leakyrelu)
def ResNet18(flatten=True, leakyrelu=False):
    return ResNet(SimpleBlock, [2,2,2,2],[64,128,256,512], flatten, leakyrelu)
def ResNet34(flatten=True, leakyrelu=False):
    return ResNet(SimpleBlock, [3,4,6,3],[64,128,256,512], flatten, leakyrelu)

model_dict = dict(Conv4 = Conv4,
                  Conv6 = Conv6,
                  ResNet10 = ResNet10,
                  ResNet18 = ResNet18,
                  ResNet34 = ResNet34)