Question

现在很长一段时间（差不多两个月）我正在使用FCN32进行单通道图像的语义分割。我玩不同的学习率，甚至添加BatchNormalization图层。但是，我甚至没有看到任何输出成功。我没有任何选择，除非立即在这里寻求帮助。我真的不知道我做错了什么。

我正在批量发送一张图片到网络。这是火车损失曲线LR=1e-9和lr_policy="fixed"：

我将学习率提高到1e-4（下图）。似乎损失正在下降，然而，学习曲线并不正常。

我减少了原始FCN的层次，如下所示：（1）Conv64 - ReLU - Conv64 - ReLU - MaxPool

（2）Conv128 - ReLU - Conv128 - ReLU - MaxPool

（3）Conv256 - ReLU - Conv256 - ReLU - MaxPool

（4）Conv4096 - ReLU - Dropout0.5

（5）Conv4096 - ReLU - Dropout0.5

（6）Conv2

（7）Deconv32x - 裁剪

（8）SoftmaxWithLoss

layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TRAIN
  }
  transform_param {
    mean_file: "/jjj/FCN32_mean.binaryproto"
  }

  data_param {
    source: "/jjj/train_lmdb/"
    batch_size: 1
    backend: LMDB
  }
}
layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TRAIN
  }
  data_param {
    source: "/jjj/train_label_lmdb/"
    batch_size: 1
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TEST
  }
  transform_param {
    mean_file: "/jjj/FCN32_mean.binaryproto"
  }
  data_param {
    source: "/jjj/val_lmdb/"
    batch_size: 1
    backend: LMDB
  }
}
layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TEST
  }
  data_param {
    source: "/jjj/val_label_lmdb/"
    batch_size: 1
    backend: LMDB
  }
}

layer {
  name: "conv1_1"
  type: "Convolution"
  bottom: "data"
  top: "conv1_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 100
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu1_1"
  type: "ReLU"
  bottom: "conv1_1"
  top: "conv1_1"
}
layer {
  name: "conv1_2"
  type: "Convolution"
  bottom: "conv1_1"
  top: "conv1_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu1_2"
  type: "ReLU"
  bottom: "conv1_2"
  top: "conv1_2"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1_2"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2_1"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu2_1"
  type: "ReLU"
  bottom: "conv2_1"
  top: "conv2_1"
}
layer {
  name: "conv2_2"
  type: "Convolution"
  bottom: "conv2_1"
  top: "conv2_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu2_2"
  type: "ReLU"
  bottom: "conv2_2"
  top: "conv2_2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2_2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv3_1"
  type: "Convolution"
  bottom: "pool2"
  top: "conv3_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu3_1"
  type: "ReLU"
  bottom: "conv3_1"
  top: "conv3_1"
}
layer {
  name: "conv3_2"
  type: "Convolution"
  bottom: "conv3_1"
  top: "conv3_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu3_2"
  type: "ReLU"
  bottom: "conv3_2"
  top: "conv3_2"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3_2"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "fc6"
  type: "Convolution"
  bottom: "pool3"
  top: "fc6"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4096
    pad: 0
    kernel_size: 7
    stride: 1
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  bottom: "fc6"
  top: "fc6"
}
layer {
  name: "drop6"
  type: "Dropout"
  bottom: "fc6"
  top: "fc6"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "fc7"
  type: "Convolution"
  bottom: "fc6"
  top: "fc7"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4096
    pad: 0
    kernel_size: 1
    stride: 1
  }
}
layer {
  name: "relu7"
  type: "ReLU"
  bottom: "fc7"
  top: "fc7"
}
layer {
  name: "drop7"
  type: "Dropout"
  bottom: "fc7"
  top: "fc7"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "score_fr"
  type: "Convolution"
  bottom: "fc7"
  top: "score_fr"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 5 #21
    pad: 0
    kernel_size: 1
    weight_filler {
        type: "xavier"
    }   
    bias_filler {
        type: "constant"
     } 
  }
}
layer {
  name: "upscore"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 5 #21
    bias_term: false
    kernel_size: 64
    stride: 32
    group: 5 #2
    weight_filler: { 
         type: "bilinear" 
     }
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore"
  bottom: "data"
  top: "score"
  crop_param {
    axis: 2
    offset: 19
  }
}
layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "score"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TRAIN
  }
}

layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "score"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TEST
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "score"
  bottom: "label"
  top: "loss"
  loss_param {
    ignore_label: 255
    normalize: true
  }
}

这是求解器的定义：

net: "train_val.prototxt"
#test_net: "val.prototxt"
test_iter: 736
# make test net, but don't invoke it from the solver itself
test_interval: 2000 #1000000
display: 50
average_loss: 50
lr_policy: "step" #"fixed"
stepsize: 2000  #+
gamma: 0.1  #+
# lr for unnormalized softmax
base_lr: 0.0001 
# high momentum
momentum: 0.99
# no gradient accumulation
iter_size: 1
max_iter: 10000
weight_decay: 0.0005
snapshot: 2000
snapshot_prefix: "snapshot/NET1"
test_initialization: false
solver_mode: GPU

一开始，损失开始下降，但经过一些迭代后，它再次没有表现出良好的学习行为：

我是深度学习的初学者，caffe。我真的不明白为什么会这样。我非常感谢那些有专业知识的人，请看一下模型定义，如果你帮助我，我将非常感激。

Answer 1

问题在于你是从头开始训练。

阅读FCN paper会告诉您他们总是使用在ImageNet上预训练的网络，如果您从头开始训练它将 NOT 工作，它必须从预训练中进行微调网络。如果从随机权重训练的优化问题不会收敛。

我在这个语义细分中做错了什么？

1 个答案: