Skip to first unread message
包青平
Nov 17, 2015, 6:43:57 AM11/17/15
to Caffe Users
I tried to do a multi-label classifiction with hdf5 dataset .there are 2 labels. one label could be labeld 0 or 1 , the other could be labeled 0,1,2,3,4. But during training iterations, accuracy kept the same, namely accuracy_1 was kept 0.5 without changing and accuracy_2 was kept 0.2. I have checked the code for many times , but still don't know why. Would someone give me some advice.
this is the code to create hdf5 dataset
image_dir = 'test/'
HDF5_FILE_TRAIN = 'hdf5_test.h5'
LABELS = dict(
kinds_0 = (0, 0),
kinds_1 = (0, 1),
kinds_2 = (0, 2),
kinds_3 = (0, 3),
kinds_4 = (0, 4),
kinds_5 = (1, 0),
kinds_6 = (1, 1),
kinds_7 = (1, 2),
kinds_8 = (1, 3),
kinds_9 = (1, 4),
)
file_list = os.listdir(image_dir)
random.shuffle(file_list)
datas = np.zeros((len(file_list), 3, 256, 256))
labels = np.zeros((len(file_list), 2)).astype(np.int)
with open('test.txt', 'r') as T:
lines = T.readlines()
labelDict = dict()
for i,l in enumerate(lines):
sp = l.strip().split(' ')
fileName = sp[0]
label = sp[1]
labelDict[fileName] = label
for j,_file in enumerate(file_list):
image = np.array(Image.open(image_dir + _file))
image = np.transpose(image, (2,0,1))
datas[j, :, :, :] = np.array(image).astype(np.float32) / 255
labels[j,:] = np.array(LABELS[labelDict[_file]]).astype(np.int)
with h5py.File(HDF5_FILE_TRAIN, 'w') as f:
f['data'] = datas
f['labels'] = labels
f.close()
in text.txt are lines as follows:
long_plaid_001.jpg kinds_0
long_plaid_002.jpg kinds_0
long_plaid_003.jpg kinds_0
......
hdf5 dataset for training is created the same way.
the definition of network:
name: "CaffeNet"
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "labels"
include {
phase: TRAIN
}
hdf5_data_param {
source: "/home/percy/caffe/models/multilabel/list_train.txt"
batch_size: 50
}
}
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "labels"
include {
phase: TEST
}
hdf5_data_param {
source: "/home/percy/caffe/models/multilabel/list_test.txt"
batch_size: 50
}
}
layer {
name: "slicers"
type: "Slice"
bottom: "labels"
top: "label_1"
top: "label_2"
slice_param {
axis: 1
slice_point: 1
}
}
.......
Conv Pool ...
.......
layer {
name: "fc6"
type: "InnerProduct"
bottom: "pool5"
top: "fc6"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 512
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 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: "InnerProduct"
bottom: "fc6"
top: "fc7"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 512
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 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: "fc8"
type: "InnerProduct"
bottom: "fc7"
top: "fc8"
param {
lr_mult: 10
decay_mult: 1
}
param {
lr_mult: 20
decay_mult: 0
}
inner_product_param {
num_output: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "accuracy_1"
type: "Accuracy"
bottom: "fc8"
bottom: "label_1"
top: "accuracy_1"
include {
phase: TEST
}
}
layer {
name: "loss_1"
type: "SoftmaxWithLoss"
bottom: "fc8"
bottom: "label_1"
top: "loss_1"
}
layer {
name: "fc9"
type: "InnerProduct"
bottom: "pool5"
top: "fc9"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 512
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 1
}
}
}
layer {
name: "relu9"
type: "ReLU"
bottom: "fc9"
top: "fc9"
}
layer {
name: "drop9"
type: "Dropout"
bottom: "fc9"
top: "fc9"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc10"
type: "InnerProduct"
bottom: "fc9"
top: "fc10"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 512
weight_filler {
type: "gaussian"
std: 0.005
}
bias_filler {
type: "constant"
value: 1
}
}
}
layer {
name: "relu10"
type: "ReLU"
bottom: "fc10"
top: "fc10"
}
layer {
name: "drop10"
type: "Dropout"
bottom: "fc10"
top: "fc10"
dropout_param {
dropout_ratio: 0.5
}
}
layer {
name: "fc11"
type: "InnerProduct"
bottom: "fc10"
top: "fc11"
param {
lr_mult: 10
decay_mult: 1
}
param {
lr_mult: 20
decay_mult: 0
}
inner_product_param {
num_output: 5
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "accuracy_2"
type: "Accuracy"
bottom: "fc11"
bottom: "label_2"
top: "aaccuracy_2"
include {
phase: TEST
}
}
layer {
name: "loss_2"
type: "SoftmaxWithLoss"
bottom: "fc11"
bottom: "label_2"
top: "loss_2"
}
solver.prototxt:
net: "/home/percy/caffe/models/multilabel/train_val.prototxt"
test_iter: 1000
test_interval: 1000
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 10000
display: 50
max_iter: 40000
momentum: 0.9
weight_decay: 0.0005
snapshot: 4000
snapshot_prefix: "caffenet_train"
solver_mode: GPU
Reply all
Reply to author
Forward
0 new messages