当步幅为2时，pytorch和tensorflow的ResNet模型给出了不同的结果

Question

class BasicBlock(nn.Module):
    def __init__(self, in_planes, out_planes, stride, dropRate=0.0):
        super(BasicBlock, self).__init__()
        self.bn1 = nn.BatchNorm2d(in_planes)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                               padding=1, bias=False) # 1
        self.bn2 = nn.BatchNorm2d(out_planes)
        self.relu2 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1,
                               padding=1, bias=False)
        self.droprate = dropRate
        self.equalInOut = (in_planes == out_planes)
        self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride,
                                                                padding=0, bias=False) or None
    def forward(self, x):
        if not self.equalInOut:
            x = self.relu1(self.bn1(x))
        else:
            out = self.relu1(self.bn1(x))
        out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x)))
        if self.droprate > 0:
            out = F.dropout(out, p=self.droprate, training=self.training)
        out = self.conv2(out)
        if self.convShortcut is not None:
        return torch.add(x if self.equalInOut else self.convShortcut(x), out)


class NetworkBlock(nn.Module):
    def __init__(self, nb_layers, in_planes, out_planes, block, stride, dropRate=0.0):
        super(NetworkBlock, self).__init__()
        self.layer = self._make_layer(block, in_planes, out_planes, nb_layers, stride, dropRate)

    def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, dropRate):
        layers = []
        for i in range(int(nb_layers)):
            layers.append(block(i == 0 and in_planes or out_planes, out_planes, i == 0 and stride or 1, dropRate))
        return nn.Sequential(*layers)

    def forward(self, x):
        return self.layer(x)

class WideResNet(nn.Module):
    def __init__(self, depth=34, num_classes=10, widen_factor=10, dropRate=0.0):
        super(WideResNet, self).__init__()
        nChannels = [16, 16 * widen_factor, 32 * widen_factor, 64 * widen_factor]
        assert ((depth - 4) % 6 == 0)
        n = (depth - 4) / 6
        block = BasicBlock
        # 1st conv before any network block
        self.conv1 = nn.Conv2d(3, nChannels[0], kernel_size=3, stride=1,
                               padding=1, bias=False)
        # 1st block
        self.block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate)
        # 1st sub-block
        self.sub_block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate)
        # 2nd block
        self.block2 = NetworkBlock(n, nChannels[1], nChannels[2], block, 2, dropRate) # 2
        # 3rd block
        self.block3 = NetworkBlock(n, nChannels[2], nChannels[3], block, 2, dropRate) # 2
        # global average pooling and classifier
        self.bn1 = nn.BatchNorm2d(nChannels[3])
        self.relu = nn.ReLU(inplace=True)
        self.fc = nn.Linear(nChannels[3], num_classes)
        self.nChannels = nChannels[3]

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                m.bias.data.zero_()

    def forward(self, x):
        out = self.conv1(x)
        out = self.block1(out)
        out = self.block2(out)
        out = self.block3(out)
        out = self.relu(self.bn1(out))
        out = F.avg_pool2d(out, 8)
        out = out.view(-1, self.nChannels)
        return self.fc(out)

  def _conv(self, name, x, filter_size, in_filters, out_filters, strides, padding='SAME'):
    """Convolution."""
    with tf.variable_scope(name):
      n = filter_size * filter_size * out_filters
      kernel = tf.get_variable(
          'DW', [filter_size, filter_size, in_filters, out_filters],
          tf.float32, initializer=tf.random_normal_initializer(
              stddev=np.sqrt(2.0/n)))
      return tf.nn.conv2d(x, kernel, strides, padding=padding)

  def _residual(self, x, in_filter, out_filter, stride,
                activate_before_residual=False, is_log=False):
    """Residual unit with 2 sub layers."""
    if activate_before_residual:
        x = self._batch_norm('bn1', x)
        x = self._relu(x)
        orig_x = x
    else:
        orig_x = x
        x = self._batch_norm('bn1', x)
        x = self._relu(x)

    x = self._conv('conv1', x, 3, in_filter, out_filter, stride)
    x = self._batch_norm('bn2', x)
    x = self._relu(x)
    x = self._conv('conv2', x, 3, out_filter, out_filter, [1, 1, 1, 1])
    if in_filter != out_filter:
        orig_x = self._conv('shortcut_conv', orig_x, filter_size=1, in_filters=in_filter, out_filters=out_filter,
                            strides=stride, padding="VALID")
    x += orig_x
    return x

  def _build_model(self):
    assert self.mode == 'train' or self.mode == 'eval'
    with tf.variable_scope('input'):

      self.x_input = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
      self.y_input = tf.placeholder(tf.float32, shape=[None, 10])
      self.is_training = tf.placeholder(tf.bool, shape=None)
    x = self._conv('conv1.weight', self.x_input, 3, 3, 16, self._stride_arr(1))

    strides = [1, 2, 2]
    activate_before_residual = [True, True, True]
    res_func = self._residual
    # wide residual network (https://arxiv.org/abs/1605.07146v1)
    filters = [16, 160, 320, 640]

    with tf.variable_scope('block1.layer.0'):
      x = res_func(x, filters[0], filters[1], self._stride_arr(strides[0]),
                   activate_before_residual[0])
    for i in range(1, 5):
      with tf.variable_scope('block1.layer.%d' % i):
        x = res_func(x, filters[1], filters[1], self._stride_arr(1), False)

    with tf.variable_scope('block2.layer.0'):
      x = res_func(x, filters[1], filters[2], self._stride_arr(strides[1]),
                   activate_before_residual[1], is_log=True)
    for i in range(1, 5):
      with tf.variable_scope('block2.layer.%d' % i):
        x = res_func(x, filters[2], filters[2], self._stride_arr(1), False)

    with tf.variable_scope('block3.layer.0'):
      x = res_func(x, filters[2], filters[3], self._stride_arr(strides[2]),
                   activate_before_residual[2])
    for i in range(1, 5):
      with tf.variable_scope('block3.layer.%d' % i):
        x = res_func(x, filters[3], filters[3], self._stride_arr(1), False)

    x = self._batch_norm('bn1', x)
    x = self._relu(x)
    x = self._global_avg_pool(x)

    with tf.variable_scope('fc'):
      self.pre_softmax = self._fully_connected(x, 10)

我正在做“对抗性防御”实验，我检查了Pytork和tensorflow在相同权重下的性能是否不同（我将其导出为numpy并加载到Pytork和tensorflow）。我打印了WideResNet34的每个结果，并计算每个输出的差异，然后，下图的上面输出出来

结果开始与block2不同。然后，我只将每个块的步幅更改为全部1（块2和块3的步幅），上面图像的下面输出就出来了

所有层的差异都可以忽略不计，因此我认为只有当步幅=2时，差异才会出现。我不知道为什么步幅=1时没有区别，但步幅=2时不同。。。谁知道这件事

Answer 1

我最终发现问题在于“填充”。当对称填充导致奇数时，Tensorflow的“相同”填充不对称地零填充（左=0，右=1，上=0，下=1）。。。然而，pytorch在nn.conv2d中不支持不对称填充，因此它对称地将填充归零（左=1，右=1，上=1，下=1）

所以，我认为当输入大小=8，过滤器大小=3，步幅=2时，tensorflow中过滤器左上角的索引将是0,2,4,6，但在pytorch中它将是-1（零垫），1,3,5。。。我检查过，当我使用nn.zero-pad2d对焊盘进行非对称归零时，它给出了几乎相同的结果（2-norm diff<；1e-2）

class BasicBlock(nn.Module):
    def __init__(self, in_planes, out_planes, stride, dropRate=0.0):
        super(BasicBlock, self).__init__()
        self.bn1 = nn.BatchNorm2d(in_planes)
        self.relu1 = nn.ReLU(inplace=True)
        if stride==1:
            self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) # 1
        else:
            self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=0, bias=False)  # 1
        self.pad1 = nn.ZeroPad2d((0,1,0,1)) # 0,1,0,1
        self.stride = stride
        self.bn2 = nn.BatchNorm2d(out_planes)
        self.relu2 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1,
                               padding=1, bias=False)
        self.droprate = dropRate
        self.equalInOut = (in_planes == out_planes)
        self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride,
                                                                padding=0, bias=False) or None

    def forward(self, x):
        if not self.equalInOut:
            x = self.relu1(self.bn1(x))
        else:
            out = self.relu1(self.bn1(x))
        if self.stride==1:
            out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x)))
        else:
            out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else self.pad1(x))))
        if self.droprate > 0:
            out = F.dropout(out, p=self.droprate, training=self.training)
        out = self.conv2(out)
        return torch.add(x if self.equalInOut else self.convShortcut(x), out)