包青平
Nov 25, 2015, 9:36:47 PM11/25/15
to Caffe Users
Hello all. I try to learn a multi-task CNN model. During training iterations, loss_1 and loss_2 decrease, accuracy_1 increases, but accuracy_2 changes a little. i.e. label_2 has 5 categories, accuracy_2 is around 0.2. Actually accuracy_2 decrease from 0.2 to 0.18 after 20000 iterations.
train_val.prototxt : I make two lmdb datasets with the same images but different labels. loss_weight of loss_1 and loss_2 are both set 1.0. Is there any errors in the prototxt?
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
mirror: true
crop_size: 227
mean_file: "/home/percy/caffe/data/clothing_multilabel/imagenet_mean.binaryproto"
}
data_param {
source: "/home/percy/caffe/data/clothing_multilabel/clothing_train1_lmdb"
batch_size: 50
backend: LMDB
}
}
layer {
name: "data2"
type: "Data"
top: "data2"
top: "label2"
include {
phase: TRAIN
}
transform_param {
mirror: true
crop_size: 227
mean_file: "/home/percy/caffe/data/clothing_multilabel/imagenet_mean.binaryproto"
}
data_param {
source: "/home/percy/caffe/data/clothing_multilabel/clothing_train2_lmdb"
batch_size: 50
backend: LMDB
}
}
layer {
name: "data"
type: "Data"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
mirror: true
crop_size: 227
mean_file: "/home/percy/caffe/data/clothing_multilabel/imagenet_mean.binaryproto"
}
data_param {
source: "/home/percy/caffe/data/clothing_multilabel/clothing_val1_lmdb"
batch_size: 50
backend: LMDB
}
}
layer {
name: "data2"
type: "Data"
top: "data2"
top: "label2"
include {
phase: TEST
}
transform_param {
mirror: true
crop_size: 227
mean_file: "/home/percy/caffe/data/clothing_multilabel/imagenet_mean.binaryproto"
}
data_param {
source: "/home/percy/caffe/data/clothing_multilabel/clothing_val2_lmdb"
batch_size: 50
backend: LMDB
}
}
……
……
layer {
name: "accuracy_1"
type: "Accuracy"
bottom: "fc8"
bottom: "label"
top: "accuracy_1"
include {
phase: TEST
}
}
layer {
name: "loss_1"
type: "SoftmaxWithLoss"
bottom: "fc8"
bottom: "label"
top: "loss_1"
loss_weight:1
}
layer {
name: "loss2_1"
type: "SoftmaxWithLoss"
bottom: "fc2_6"
bottom: "label2"
top: "loss2_1"
loss_weight:0.0
}
……
……
layer {
name: "accuracy_2"
type: "Accuracy"
bottom: "fc11"
bottom: "label2"
top: "accuracy_2"
include {
phase: TEST
}
}
layer {
name: "loss_2"
type: "SoftmaxWithLoss"
bottom: "fc11"
bottom: "label2"
top: "loss_2"
loss_weight:1
}
Reply all
Reply to author
Forward
0 new messages