我正在尝试在窗口上的pycaffe中训练一个卷积神经网络，用于像素分类。
在每次迭代期间，损失和准确度保持不变。 当我使用快照中的权重时，输出只是黑色。
我的问题是：我该如何改变？ 到目前为止，我已经学习了学习率（0.01到0.0000001），批量大小（1-20）和权重初始化（高斯与xavier）。
我应该增加图像数量吗？
训练数据集由HDF5文件中的100个图像对（通过减去中位数预处理）组成，如下所示：
[u'data'], [u'label']
尺寸为：
100 x 1 x 584 x 584数据（火车）
标签（火车）100 x 1 x 139 x 139
和
10 x 1 x 584 x 584用于数据（测试）
10 x 1 x 139 x 139用于标签（测试）
solver.prototxt如下所示：
net: "train.prototxt"
test_initialization: true
test_iter: 10
test_interval: 100
base_lr: 0.0000001
momentum: 0.9
weight_decay: 0.0005
lr_policy: "step"
stepsize: 1000
gamma: 0.1
display: 100
iter_size: 5
max_iter: 50000
snapshot: 5000
snapshot_prefix: "models/train"
solver_mode: GPU
type: "SGD"
它调用的网络发布在下面：
# data: 20 1 584 584 (6821120)
# label: 20 1 139 139 (386420)
# conv1: 20 32 582 582 (216783360)
# relu1: 20 32 582 582 (216783360)
# norm1: 20 32 582 582 (216783360)
# pool1: 20 32 290 290 (53824000)
# conv2: 20 32 288 288 (53084160)
# relu2: 20 32 288 288 (53084160)
# norm2: 20 32 288 288 (53084160)
# pool2: 20 32 144 144 (13271040)
# fc6-conv: 20 64 139 139 (24730880)
# relu6: 20 64 139 139 (24730880)
# drop6: 20 64 139 139 (24730880)
# fc7-conv: 20 64 139 139 (24730880)
# relu7: 20 64 139 139 (24730880)
# drop7: 20 64 139 139 (24730880)
# fc8-conv: 20 1 139 139 (386420)
# prob: 20 1 139 139 (386420)
# loss: (1)
#
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
hdf5_data_param {
source: "../train/train.txt"
batch_size: 20
}
}
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TEST
}
hdf5_data_param {
source: "../train/test.txt"
batch_size: 1
}
}
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 2.0
decay_mult: 0.0
}
convolution_param {
num_output: 32
kernel_size: 3
stride: 1
weight_filler {
type: "xavier"
#std: 0.01
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "relu1"
type: "ReLU"
bottom: "conv1"
top: "conv1"
}
layer {
name: "norm1"
type: "LRN"
bottom: "conv1"
top: "norm1"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool1"
type: "Pooling"
bottom: "norm1"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 4
stride: 2
}
}
layer {
name: "conv2"
type: "Convolution"
bottom: "pool1"
top: "conv2"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 2.0
decay_mult: 0.0
}
convolution_param {
num_output: 32
kernel_size: 3
weight_filler {
type: "xavier"
#std: 0.01
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "relu2"
type: "ReLU"
bottom: "conv2"
top: "conv2"
}
layer {
name: "norm2"
type: "LRN"
bottom: "conv2"
top: "norm2"
lrn_param {
local_size: 5
alpha: 0.0001
beta: 0.75
}
}
layer {
name: "pool2"
type: "Pooling"
bottom: "norm2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "fc6-conv"
type: "Convolution"
bottom: "pool2"
top: "fc6-conv"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 2.0
decay_mult: 0.0
}
convolution_param{
kernel_size: 6
num_output: 64
weight_filler {
type: "xavier"
#std: 0.005
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "relu6"
type: "ReLU"
bottom: "fc6-conv"
top: "fc6-conv"
}
layer {
name: "drop6"
type: "Dropout"
bottom: "fc6-conv"
top: "fc6-conv"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc7-conv"
type: "Convolution"
bottom: "fc6-conv"
top: "fc7-conv"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 2.0
decay_mult: 0.0
}
convolution_param {
kernel_size: 1
num_output: 64
weight_filler {
type: "xavier"
#std: 0.005
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "relu7"
type: "ReLU"
bottom: "fc7-conv"
top: "fc7-conv"
}
layer {
name: "drop7"
type: "Dropout"
bottom: "fc7-conv"
top: "fc7-conv"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc8-conv"
type: "Convolution"
bottom: "fc7-conv"
top: "fc8-conv"
param {
lr_mult: 1.0
decay_mult: 1.0
}
param {
lr_mult: 2.0
decay_mult: 0.0
}
convolution_param{
kernel_size: 1
num_output: 1
weight_filler {
type: "xavier"
#std: 0.01
}
bias_filler {
type: "constant"
value: 0.2
}
}
}
layer {
name: "prob"
type: "Softmax"
bottom: "fc8-conv"
top: "prob"
}
layer {
name: "accuracy"
type: "Accuracy"
bottom: "prob"
bottom: "label"
top: "accuracy"
include {
phase: TEST
}
}
layer {
name: "loss"
type: "SigmoidCrossEntropyLoss"
bottom: "prob"
bottom: "label"
top: "loss"
loss_weight: 1
include {
phase: TRAIN
}
}
非常感谢帮助！
编辑： 似乎向后传球没有做任何事情。 在上面发布的网络文件中，net.backward只给出{}。 将输入图层重写为：
input: "data"
input_dim: 1
input_dim: 1
input_dim: 584
input_dim: 584
input: "label"
input_dim: 1
input_dim: 1
input_dim: 139
input_dim: 139
并设置force_backward：true
给出一个零数组。