为什么 caffe 的 net.forward() 将我的数据变成负值？

Question

我对 caffe 很陌生，正在尝试针对我的具体情况使用别人的代码。

我把它缩小到这个最小的例子：

import caffe
import numpy as np

model_path = "mnist_model_rgb/caffe_model/mnist_model_rgb.caffemodel"
model_definition = 'mnist_model_rgb/caffe_model/mnist_model_rgb.prototxt'

net = caffe.Classifier(model_definition, model_path, mean=np.float32([0.131, 0.131, 0.131]), channel_swap=(2, 1, 0))

w = h = 28

start_image = np.random.normal(np.float32([175.0, 175.0, 175.0]), 8, (w, h, 3))
start_image = np.float32(np.rollaxis(start_image, 2)[::-1]) - net.transformer.mean['data']

src = net.blobs['data']
src.reshape(1, 3, h, w)
src.data[0] = start_image

layer = 'flat'
print("before forward: ", net.blobs[layer].data)
net.forward()
print("after forward: ", net.blobs[layer].data)

这是我的模型描述：

layer {
  name: "data"
  type: "Input"
  top: "data"
  input_param {
    shape {
      dim: 1
      dim: 3
      dim: 28
      dim: 28
    }
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 32
    kernel_size: 3
  }
}
layer {
  name: "max1"
  type: "Pooling"
  bottom: "conv1"
  top: "mp1"
  pooling_param {
    pool: MAX
    kernel_size: 2
  }
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "mp1"
  top: "conv2"
  convolution_param {
    num_output: 64
    kernel_size: 3
  }
}
layer {
  name: "max2"
  type: "Pooling"
  bottom: "conv2"
  top: "mp2"
  pooling_param {
    pool: MAX
    kernel_size: 2
  }
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "mp2"
  top: "conv3"
  convolution_param {
    num_output: 64
    kernel_size: 3
  }
}
layer {
  name: "flat"
  type: "Flatten"
  bottom: "conv3"
  top: "flat"
}
layer {
  name: "dense1"
  type: "InnerProduct"
  bottom: "flat"
  top: "dense1"
  inner_product_param {
    num_output: 64
  }
}
layer {
  name: "dense2"
  type: "InnerProduct"
  bottom: "dense1"
  top: "dense2"
  inner_product_param {
    num_output: 10
  }
}

这是我得到的输出：

before forward:  [[0. 0. 0. ... 0. 0. 0.]]
after forward:  [[-350974.78 -351307.06 -350903.56 ... -296046.62 -295620.03 -297597.94]]

net.forward()

之后的值对我来说似乎不正确。

当我使用作者预期的模型 (https://www.robots.ox.ac.uk/~vgg/software/vgg_face/) 运行相同的脚本时，我得到了这个结果，这似乎更合理：

before forward:  [[0. 0. 0. ... 0. 0. 0.]]
after forward:  [[0.04778102 0.         0.18989444 ... 2.3401992  0.         0.        ]]

有人可以向我解释一下，为什么我的模型的输出与示例模型如此不同？我应该期待得到的结果吗？或者我只是做错了什么，我怎样才能做得更好？

任何解释或帮助表示赞赏！