Skip to first unread message
Wu Huan
Oct 7, 2017, 10:36:01 PM10/7/17
to Caffe Users
In traditional, the label is single integer number. During the runtime, the label is stretched out into a vector. Assume the label is 3, then the output is `(0, 0, 0, 1, 0, 0, 0, 0, 0, 0)`. But now I have a probability vector, such as `(0.1, 0.1, 0.01, 0.78, 0.01, 0, 0, 0, 0, 0)`.
Because of the float number, I should use hdf5 instead of lmdb. But the `SoftmaxWithLoss` layer only support single integer label rather than float vector
label.
**How do I train the model in Caffe?**
**train_prototxt**
name: "LeNet"
layer {
name: "feature"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
hdf5_data_param {
source: "h5_list.txt"
batch_size: 1
}
}
layer {
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 250
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "relu1"
type: "ReLU"
bottom: "ip1"
top: "ip1"
}
layer {
name: "ip2"
type: "InnerProduct"
bottom: "ip1"
top: "ip2"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 10
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip2"
bottom: "label"
top: "loss"
}
****
Reply all
Reply to author
Forward
0 new messages