如何用caffe解决线性回归解决什么问题问题
点击联系发帖人
时间：2016-04-26 03:40
使用caffe的HDF5数据完成回归任务-爱编程
使用caffe的HDF5数据完成回归任务
&&& 一直在研究如何用caffe做行人检测问题，然而参考那些经典结构比如faster-rcnn等，都是自定义的caffe层来完成的检测任务。这些都要求对caffe框架有一定程度的了解。最近看到了如何用caffe完成回归的任务，就想把检测问题当成回归问题来解决。
&&& 我们把行人检测问题当成回归来看待，就需要限制检出目标的个数，因为我们的输出个数是固定的。所以，这里我假定每张图片最多检出的目标个数为2。即每个目标用4个值来表示其位置信息（中心位置坐标x,y。BBox的宽和高），则网络的最后输出是8个值。
制作HDF5数据
&&& 这里我们使用HDF5格式的数据来完成我们的回归任务，那么首先我们需要的是制作h5格式的数据。这里以VOC数据集为例。下面是制作HDF5格式数据的python代码。
import h5py
import caffe
import xml.etree.ElementTree as ET
import cv2
import time
import math
from os.path import join, exists
import numpy as np
def convert(size, box):
dw = 1./size[0]
dh = 1./size[1]
x = (box[0] + box[1])/2.0
y = (box[2] + box[3])/2.0
w = box[1] - box[0]
h = box[3] - box[2]
return (x,y,w,h)
def shuffle_in_unison_scary(a, b):
rng_state = np.random.get_state()
np.random.shuffle(a)
np.random.set_state(rng_state)
np.random.shuffle(b)
def processImage(imgs):
imgs = imgs.astype(np.float32)
for i, img in enumerate(imgs):
m = img.mean()
s = img.std()
imgs[i] = (img - m) / s
return imgs
TrainImgDir = 'F:/GenerateHDF5/trainImage'
TrainLabelDir = 'F:/GenerateHDF5/trainLabels'
TestImgDir = 'F:/GenerateHDF5/testImg'
TestLabelDir = 'F:/GenerateHDF5/testLabels'
InImg = []
InBBox = []
for rootDir,dirs,files in os.walk(TestLabelDir):
for file in files:
file_name = file.split('.')[0]
full_file_name = '%s%s'%(file_name,'.jpg')
full_file_dir = '%s/%s'%(TestImgDir,full_file_name)
Img = cv2.imread(full_file_dir,cv2.CV_LOAD_IMAGE_GRAYSCALE)
xml_file = open(&%s/%s&%(rootDir,file))
tree = ET.parse(xml_file)
root = tree.getroot()
size = root.find('size')
w = int(size.find('width').text)
h = int(size.find('height').text)
landmark = np.zeros(8)
for obj in root.iter('object'):
count = count + 1
if count == 3:
xmlbox = obj.find('bndbox')
b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
bb = convert((w,h), b)
landmark[(count-1)*4+0]=bb[0]
landmark[(count-1)*4+1]=bb[1]
landmark[(count-1)*4+2]=bb[2]
landmark[(count-1)*4+3]=bb[3]
InBBox.append(landmark.reshape(8))
Img = cv2.resize(Img,(h,w))
InImg.append(Img.reshape((1,h,w)))
InImg, InBBox = np.asarray(InImg), np.asarray(InBBox)
InImg = processImage(InImg)
shuffle_in_unison_scary(InImg, InBBox)
outputDir = 'hdf5/'
HDF5_file_name = 'hdf5_test.h5'
if not os.path.exists(outputDir):
os.makedirs(outputDir)
output = join(outputDir,HDF5_file_name)
with h5py.File(output, 'w') as h5:
h5['data'] = InImg.astype(np.float32)
h5['labels'] = InBBox.astype(np.float32)
h5.close()
这里注意一点，所有的BBox数据都要做归一化操作，即所有坐标要除以图片对应的宽高。据说，这样做能使最后得到的结果更好。
制作好了HDF5数据后，注意每个H5文件大小不能超过2G（这是caffe的规定，如果一个文件超过2G，请分开制作多个）。然后建立一个TXT文件，文件里写上所有H5文件的绝对路径，比如我这里建立的文件是list_train.txt。然后我只有一个H5文件，即hdf5_train.h5。所以我的list_train.txt文件中的内容就是/home/XXX/caffe/model/hdf5/hdf5_train.h5
配置solver文件
接下来是caffe的solver文件，这个文件没有什么区别，
test_iter: 20
test_interval: 70
base_lr: 0.
display: 9
max_iter: 210000
lr_policy: &step&
gamma: 0.1
momentum: 0.9
weight_decay: 0.0001
stepsize: 700
snapshot: 500
snapshot_prefix: &snapshot&
solver_mode: GPU
net: &train_val.prototxt&
solver_type: SGD
配置train_val.prototxt文件
接下来是网络的train_val.prototxt文件。这是caffe的网络结构文件，我们这里以LeNet网络为例，我这里是这样的：
name: &LeNet&
name: &data&
type: &HDF5Data&
top: &data&
top: &labels&
phase: TRAIN
hdf5_data_param {
source: &list_train.txt&
batch_size: 50
name: &data&
type: &HDF5Data&
top: &data&
top: &labels&
phase: TEST
hdf5_data_param {
source: &list_test.txt&
batch_size: 50
name: &conv1&
type: &Convolution&
bottom: &scaled&
top: &conv1&
lr_mult: 1.0
lr_mult: 2.0
convolution_param {
num_output: 20
kernel_size: 5
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &pool1&
type: &Pooling&
bottom: &conv1&
top: &pool1&
pooling_param {
kernel_size: 2
name: &conv2&
type: &Convolution&
bottom: &pool1&
top: &conv2&
lr_mult: 1.0
lr_mult: 2.0
convolution_param {
num_output: 50
kernel_size: 5
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &pool2&
type: &Pooling&
bottom: &conv2&
top: &pool2&
pooling_param {
kernel_size: 2
name: &ip1&
type: &InnerProduct&
bottom: &pool2&
top: &ip1&
lr_mult: 1.0
lr_mult: 2.0
inner_product_param {
num_output: 500
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &relu1&
type: &ReLU&
bottom: &ip1&
top: &ip1&
name: &ip2&
type: &InnerProduct&
bottom: &ip1&
top: &ip2&
lr_mult: 1.0
lr_mult: 2.0
inner_product_param {
num_output: 8
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &error&
type: &EuclideanLoss&
bottom: &ip2&
bottom: &labels&
top: &error&
phase: TEST
name: &loss&
type: &EuclideanLoss&
bottom: &ip2&
bottom: &labels&
top: &loss&
phase: TRAIN
这里注意的是，最后的一层全连接层，输出的num_output应该是你label的维度，我这里是8。然后最后的loss计算，我使用的是欧氏距离的loss，也可以试着用其他类型的loss。
按照以上步骤配置好了，最后就是训练了。在控制台中输入以下指令来训练我们的数据：
./cafferoot/caffe/tools/caffe train --solver=solver.prototxt
可能是我数据源的问题，我的loss一开始非常大，然后一直降不下来。也有可能是LeNet本身网络性能就不好。关于网络的性能还需要另外再想办法提升。
版权所有 爱编程 (C) Copyright 2012. . All Rights Reserved.
闽ICP备号-3
微信扫一扫关注爱编程，每天为您推送一篇经典技术文章。1836人阅读
caffe（15）
虽然Caffe是设计给深度网络的，但它也可以表示“浅层”网络模型用于分类，例如逻辑回归。下面在合成数据上做简单的逻辑回归，生成并保存为HDF5，来向Caffe输入向量。模型完成后，添加层来提高准确率。Caffe完成的主要内容就是：定义模型，实验，应用。
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
os.chdir('..')
import sys
sys.path.insert(0, './python')
import caffe
import h5py
import shutil
import tempfile
import sklearn
import sklearn.datasets
import sklearn.linear_model
import pandas as pd
合成10,000个4向量的数据集，用于有2个有用特征和2个噪声特征的二值分类。
X, y = sklearn.datasets.make_classification(
n_samples=10000, n_features=4, n_redundant=0, n_informative=2,
n_clusters_per_class=2, hypercube=False, random_state=0
# Split into train and test
X, Xt, y, yt = sklearn.cross_validation.train_test_split(X, y)
# Visualize sample of the data
ind = np.random.permutation(X.shape[0])[:1000]
df = pd.DataFrame(X[ind])
_ = pd.scatter_matrix(df, figsize=(9, 9), diagonal='kde', marker='o', s=40, alpha=.4, c=y[ind])
用于学习scikit-learn逻辑回归的方法是随机梯度下降（SGD），下面检测分类器的准确性：
clf = sklearn.linear_model.SGDClassifier(
loss='log', n_iter=1000, penalty='l2', alpha=1e-3, class_weight='auto')
clf.fit(X, y)
yt_pred = clf.predict(Xt)
print('Accuracy: {:.3f}'.format(sklearn.metrics.accuracy_score(yt, yt_pred)))
Accuracy: 0.783
Accuracy: 0.783
Accuracy: 0.783
Accuracy: 0.783
1 loops, best of 3: 508 ms per loop
将这个数据集保存到HDF5中，方便Caffe载入：
dirname = os.path.abspath('./examples/hdf5_classification/data')
if not os.path.exists(dirname):
os.makedirs(dirname)
train_filename = os.path.join(dirname, 'train.h5')
test_filename = os.path.join(dirname, 'test.h5')
with h5py.File(train_filename, 'w') as f:
f['data'] = X
f['label'] = y.astype(np.float32)
with open(os.path.join(dirname, 'train.txt'), 'w') as f:
f.write(train_filename + '\n')
f.write(train_filename + '\n')
comp_kwargs = {'compression': 'gzip', 'compression_opts': 1}
with h5py.File(test_filename, 'w') as f:
f.create_dataset('data', data=Xt, **comp_kwargs)
f.create_dataset('label', data=yt.astype(np.float32), **comp_kwargs)
with open(os.path.join(dirname, 'test.txt'), 'w') as f:
f.write(test_filename + '\n')
通过Python网络规范来定义Caffe中的逻辑回归，如果不想手工编写protobuf模型的话这是一个好方法：
from caffe import layers as L
from caffe import params as P
def logreg(hdf5, batch_size):
n = caffe.NetSpec()
n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
n.ip1 = L.InnerProduct(n.data, num_output=2, weight_filler=dict(type='xavier'))
n.accuracy = L.Accuracy(n.ip1, n.label)
n.loss = L.SoftmaxWithLoss(n.ip1, n.label)
return n.to_proto()
with open('examples/hdf5_classification/logreg_auto_train.prototxt', 'w') as f:
f.write(str(logreg('examples/hdf5_classification/data/train.txt', 10)))
with open('examples/hdf5_classification/logreg_auto_test.prototxt', 'w') as f:
f.write(str(logreg('examples/hdf5_classification/data/test.txt', 10)))
现在开始学习与评估Python中的Caffe逻辑回归：
caffe.set_mode_cpu()
solver = caffe.get_solver('examples/hdf5_classification/solver.prototxt')
solver.solve()
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
solver.test_nets[0].forward()
accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
print("Accuracy: {:.3f}".format(accuracy))
Accuracy: 0.782
Accuracy: 0.782
Accuracy: 0.782
Accuracy: 0.782
1 loops, best of 3: 287 ms per loop
用命令行来运行，看一下具体的模型文件和求解过程：
!./build/tools/caffe train -solver examples/hdf5_classification/solver.prototxt
I0318 00:58:32.322571
caffe.cpp:117] Use CPU.
I0318 00:58:32.643163
caffe.cpp:121] Starting Optimization
I0318 00:58:32.643229
solver.cpp:32] Initializing solver from parameters:
train_net: "examples/hdf5_classification/logreg_auto_train.prototxt"
test_net: "examples/hdf5_classification/logreg_auto_test.prototxt"
test_iter: 250
test_interval: 1000
base_lr: 0.01
display: 1000
max_iter: 10000
lr_policy: "step"
gamma: 0.1
momentum: 0.9
weight_decay: 0.0005
stepsize: 5000
snapshot: 10000
snapshot_prefix: "examples/hdf5_classification/data/train"
solver_mode: CPU
I0318 00:58:32.643333
solver.cpp:61] Creating training net from train_net file: examples/hdf5_classification/logreg_auto_train.prototxt
I0318 00:58:32.643465
net.cpp:42] Initializing net from parameters:
phase: TRAIN
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/train.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip1"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip1"
bottom: "label"
top: "loss"
I0318 00:58:32.644197
layer_factory.hpp:74] Creating layer data
I0318 00:58:32.644219
net.cpp:84] Creating Layer data
I0318 00:58:32.644230
net.cpp:338] data -& data
I0318 00:58:32.644256
net.cpp:338] data -& label
I0318 00:58:32.644269
net.cpp:113] Setting up data
I0318 00:58:32.644278
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/train.txt
I0318 00:58:32.644327
hdf5_data_layer.cpp:80] Number of HDF5 files: 2
I0318 00:58:32.646458
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:32.646502
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646518
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:32.646538
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:32.646546
net.cpp:380] label_data_1_split &- label
I0318 00:58:32.646556
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:32.646569
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:32.646579
net.cpp:113] Setting up label_data_1_split
I0318 00:58:32.646586
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646595
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646601
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:32.646615
net.cpp:84] Creating Layer ip1
I0318 00:58:32.646622
net.cpp:380] ip1 &- data
I0318 00:58:32.646664
net.cpp:338] ip1 -& ip1
I0318 00:58:32.646689
net.cpp:113] Setting up ip1
I0318 00:58:32.652330
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652371
layer_factory.hpp:74] Creating layer ip1_ip1_0_split
I0318 00:58:32.652393
net.cpp:84] Creating Layer ip1_ip1_0_split
I0318 00:58:32.652407
net.cpp:380] ip1_ip1_0_split &- ip1
I0318 00:58:32.652421
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_0
I0318 00:58:32.652467
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_1
I0318 00:58:32.652480
net.cpp:113] Setting up ip1_ip1_0_split
I0318 00:58:32.652489
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652498
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652505
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:32.652521
net.cpp:84] Creating Layer accuracy
I0318 00:58:32.652534
net.cpp:380] accuracy &- ip1_ip1_0_split_0
I0318 00:58:32.652545
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:32.652562
net.cpp:338] accuracy -& accuracy
I0318 00:58:32.652577
net.cpp:113] Setting up accuracy
I0318 00:58:32.652590
net.cpp:120] Top shape: (1)
I0318 00:58:32.652642
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.652655
net.cpp:84] Creating Layer loss
I0318 00:58:32.652663
net.cpp:380] loss &- ip1_ip1_0_split_1
I0318 00:58:32.652672
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:32.652679
net.cpp:338] loss -& loss
I0318 00:58:32.652689
net.cpp:113] Setting up loss
I0318 00:58:32.652701
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.652716
net.cpp:120] Top shape: (1)
I0318 00:58:32.652724
net.cpp:122]
with loss weight 1
I0318 00:58:32.652740
net.cpp:167] loss needs backward computation.
I0318 00:58:32.652746
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:32.652753
net.cpp:167] ip1_ip1_0_split needs backward computation.
I0318 00:58:32.652760
net.cpp:167] ip1 needs backward computation.
I0318 00:58:32.652786
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:32.652801
net.cpp:169] data does not need backward computation.
I0318 00:58:32.652808
net.cpp:205] This network produces output accuracy
I0318 00:58:32.652815
net.cpp:205] This network produces output loss
I0318 00:58:32.652825
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:32.652833
net.cpp:217] Network initialization done.
I0318 00:58:32.652839
net.cpp:218] Memory required for data: 528
I0318 00:58:32.652964
solver.cpp:154] Creating test net (#0) specified by test_net file: examples/hdf5_classification/logreg_auto_test.prototxt
I0318 00:58:32.652986
net.cpp:42] Initializing net from parameters:
phase: TEST
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/test.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip1"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip1"
bottom: "label"
top: "loss"
I0318 00:58:32.653069
layer_factory.hpp:74] Creating layer data
I0318 00:58:32.653080
net.cpp:84] Creating Layer data
I0318 00:58:32.653090
net.cpp:338] data -& data
I0318 00:58:32.653128
net.cpp:338] data -& label
I0318 00:58:32.653146
net.cpp:113] Setting up data
I0318 00:58:32.653154
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/test.txt
I0318 00:58:32.653192
hdf5_data_layer.cpp:80] Number of HDF5 files: 1
I0318 00:58:32.654850
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:32.654897
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.654914
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:32.654933
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:32.654943
net.cpp:380] label_data_1_split &- label
I0318 00:58:32.654953
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:32.654966
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:32.654976
net.cpp:113] Setting up label_data_1_split
I0318 00:58:32.654985
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.654992
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.655000
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:32.655010
net.cpp:84] Creating Layer ip1
I0318 00:58:32.655017
net.cpp:380] ip1 &- data
I0318 00:58:32.655030
net.cpp:338] ip1 -& ip1
I0318 00:58:32.655041
net.cpp:113] Setting up ip1
I0318 00:58:32.655061
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655072
layer_factory.hpp:74] Creating layer ip1_ip1_0_split
I0318 00:58:32.655148
net.cpp:84] Creating Layer ip1_ip1_0_split
I0318 00:58:32.655159
net.cpp:380] ip1_ip1_0_split &- ip1
I0318 00:58:32.655170
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_0
I0318 00:58:32.655180
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_1
I0318 00:58:32.655190
net.cpp:113] Setting up ip1_ip1_0_split
I0318 00:58:32.655199
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655206
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655213
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:32.655223
net.cpp:84] Creating Layer accuracy
I0318 00:58:32.655230
net.cpp:380] accuracy &- ip1_ip1_0_split_0
I0318 00:58:32.655237
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:32.655251
net.cpp:338] accuracy -& accuracy
I0318 00:58:32.655259
net.cpp:113] Setting up accuracy
I0318 00:58:32.655267
net.cpp:120] Top shape: (1)
I0318 00:58:32.655340
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.655354
net.cpp:84] Creating Layer loss
I0318 00:58:32.655361
net.cpp:380] loss &- ip1_ip1_0_split_1
I0318 00:58:32.655369
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:32.655378
net.cpp:338] loss -& loss
I0318 00:58:32.655388
net.cpp:113] Setting up loss
I0318 00:58:32.655397
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.655414
net.cpp:120] Top shape: (1)
I0318 00:58:32.655422
net.cpp:122]
with loss weight 1
I0318 00:58:32.655438
net.cpp:167] loss needs backward computation.
I0318 00:58:32.655446
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:32.655455
net.cpp:167] ip1_ip1_0_split needs backward computation.
I0318 00:58:32.655462
net.cpp:167] ip1 needs backward computation.
I0318 00:58:32.655469
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:32.655477
net.cpp:169] data does not need backward computation.
I0318 00:58:32.655483
net.cpp:205] This network produces output accuracy
I0318 00:58:32.655489
net.cpp:205] This network produces output loss
I0318 00:58:32.655503
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:32.655511
net.cpp:217] Network initialization done.
I0318 00:58:32.655517
net.cpp:218] Memory required for data: 528
I0318 00:58:32.655547
solver.cpp:42] Solver scaffolding done.
I0318 00:58:32.655567
solver.cpp:222] Solving
I0318 00:58:32.655575
solver.cpp:223] Learning Rate Policy: step
I0318 00:58:32.655583
solver.cpp:266] Iteration 0, Testing net (#0)
I0318 00:58:32.683643
solver.cpp:315]
Test net output #0: accuracy = 0.3736
I0318 00:58:32.683686
solver.cpp:315]
Test net output #1: loss = 1.00555 (* 1 = 1.00555 loss)
I0318 00:58:32.683846
solver.cpp:189] Iteration 0, loss = 0.869394
I0318 00:58:32.683861
solver.cpp:204]
Train net output #0: accuracy = 0.3
I0318 00:58:32.683871
solver.cpp:204]
Train net output #1: loss = 0.869394 (* 1 = 0.869394 loss)
I0318 00:58:32.683883
solver.cpp:464] Iteration 0, lr = 0.01
I0318 00:58:32.698721
solver.cpp:266] Iteration 1000, Testing net (#0)
I0318 00:58:32.701917
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.701961
solver.cpp:315]
Test net output #1: loss = 0.590972 (* 1 = 0.590972 loss)
I0318 00:58:32.702014
solver.cpp:189] Iteration 1000, loss = 0.54742
I0318 00:58:32.702029
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:32.702041
solver.cpp:204]
Train net output #1: loss = 0.54742 (* 1 = 0.54742 loss)
I0318 00:58:32.702051
solver.cpp:464] Iteration 1000, lr = 0.01
I0318 00:58:32.718360
solver.cpp:266] Iteration 2000, Testing net (#0)
I0318 00:58:32.721529
solver.cpp:315]
Test net output #0: accuracy = 0.7696
I0318 00:58:32.721562
solver.cpp:315]
Test net output #1: loss = 0.593946 (* 1 = 0.593946 loss)
I0318 00:58:32.721593
solver.cpp:189] Iteration 2000, loss = 0.729569
I0318 00:58:32.721603
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.721613
solver.cpp:204]
Train net output #1: loss = 0.729569 (* 1 = 0.729569 loss)
I0318 00:58:32.721622
solver.cpp:464] Iteration 2000, lr = 0.01
I0318 00:58:32.740182
solver.cpp:266] Iteration 3000, Testing net (#0)
I0318 00:58:32.743494
solver.cpp:315]
Test net output #0: accuracy = 0.77
I0318 00:58:32.743544
solver.cpp:315]
Test net output #1: loss = 0.591229 (* 1 = 0.591229 loss)
I0318 00:58:32.744209
solver.cpp:189] Iteration 3000, loss = 0.406097
I0318 00:58:32.744231
solver.cpp:204]
Train net output #0: accuracy = 0.8
I0318 00:58:32.744249
solver.cpp:204]
Train net output #1: loss = 0.406096 (* 1 = 0.406096 loss)
I0318 00:58:32.744266
solver.cpp:464] Iteration 3000, lr = 0.01
I0318 00:58:32.764135
solver.cpp:266] Iteration 4000, Testing net (#0)
I0318 00:58:32.769110
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.769170
solver.cpp:315]
Test net output #1: loss = 0.590972 (* 1 = 0.590972 loss)
I0318 00:58:32.769223
solver.cpp:189] Iteration 4000, loss = 0.54742
I0318 00:58:32.769242
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:32.769255
solver.cpp:204]
Train net output #1: loss = 0.54742 (* 1 = 0.54742 loss)
I0318 00:58:32.769265
solver.cpp:464] Iteration 4000, lr = 0.01
I0318 00:58:32.785846
solver.cpp:266] Iteration 5000, Testing net (#0)
I0318 00:58:32.788722
solver.cpp:315]
Test net output #0: accuracy = 0.7696
I0318 00:58:32.788751
solver.cpp:315]
Test net output #1: loss = 0.593946 (* 1 = 0.593946 loss)
I0318 00:58:32.788811
solver.cpp:189] Iteration 5000, loss = 0.72957
I0318 00:58:32.788833
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.788846
solver.cpp:204]
Train net output #1: loss = 0.729569 (* 1 = 0.729569 loss)
I0318 00:58:32.788856
solver.cpp:464] Iteration 5000, lr = 0.001
I0318 00:58:32.804762
solver.cpp:266] Iteration 6000, Testing net (#0)
I0318 00:58:32.808061
solver.cpp:315]
Test net output #0: accuracy = 0.7856
I0318 00:58:32.808112
solver.cpp:315]
Test net output #1: loss = 0.59028 (* 1 = 0.59028 loss)
I0318 00:58:32.808732
solver.cpp:189] Iteration 6000, loss = 0.415444
I0318 00:58:32.808753
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:32.808773
solver.cpp:204]
Train net output #1: loss = 0.415444 (* 1 = 0.415444 loss)
I0318 00:58:32.808786
solver.cpp:464] Iteration 6000, lr = 0.001
I0318 00:58:32.827118
solver.cpp:266] Iteration 7000, Testing net (#0)
I0318 00:58:32.831614
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.831657
solver.cpp:315]
Test net output #1: loss = 0.589454 (* 1 = 0.589454 loss)
I0318 00:58:32.831707
solver.cpp:189] Iteration 7000, loss = 0.538038
I0318 00:58:32.831728
solver.cpp:204]
Train net output #0: accuracy = 0.8
I0318 00:58:32.831745
solver.cpp:204]
Train net output #1: loss = 0.538037 (* 1 = 0.538037 loss)
I0318 00:58:32.831759
solver.cpp:464] Iteration 7000, lr = 0.001
I0318 00:58:32.849634
solver.cpp:266] Iteration 8000, Testing net (#0)
I0318 00:58:32.852712
solver.cpp:315]
Test net output #0: accuracy = 0.7796
I0318 00:58:32.852748
solver.cpp:315]
Test net output #1: loss = 0.589365 (* 1 = 0.589365 loss)
I0318 00:58:32.852792
solver.cpp:189] Iteration 8000, loss = 0.684219
I0318 00:58:32.852840
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.852852
solver.cpp:204]
Train net output #1: loss = 0.684219 (* 1 = 0.684219 loss)
I0318 00:58:32.852861
solver.cpp:464] Iteration 8000, lr = 0.001
I0318 00:58:32.868440
solver.cpp:266] Iteration 9000, Testing net (#0)
I0318 00:58:32.871438
solver.cpp:315]
Test net output #0: accuracy = 0.7816
I0318 00:58:32.871461
solver.cpp:315]
Test net output #1: loss = 0.589656 (* 1 = 0.589656 loss)
I0318 00:58:32.872109
solver.cpp:189] Iteration 9000, loss = 0.421879
I0318 00:58:32.872131
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:32.872143
solver.cpp:204]
Train net output #1: loss = 0.421879 (* 1 = 0.421879 loss)
I0318 00:58:32.872153
solver.cpp:464] Iteration 9000, lr = 0.001
I0318 00:58:32.889981
solver.cpp:334] Snapshotting to examples/hdf5_classification/data/train_iter_10000.caffemodel
I0318 00:58:32.890224
solver.cpp:342] Snapshotting solver state to examples/hdf5_classification/data/train_iter_10000.solverstate
I0318 00:58:32.890362
solver.cpp:248] Iteration 10000, loss = 0.538933
I0318 00:58:32.890380
solver.cpp:266] Iteration 10000, Testing net (#0)
I0318 00:58:32.893728
solver.cpp:315]
Test net output #0: accuracy = 0.782
I0318 00:58:32.893757
solver.cpp:315]
Test net output #1: loss = 0.589366 (* 1 = 0.589366 loss)
I0318 00:58:32.893775
solver.cpp:253] Optimization Done.
I0318 00:58:32.893786
caffe.cpp:134] Optimization Done.
看一下上面的输出或者logreg_auto_train.prototxt，就会发现这个模型是简单的逻辑回归，可以通过引入输入权重和输出权重之间的非线性，使这个模型更好一点，变成两层的，把这个网络放在nonlinear_auto_train.prototxt中。新网络最终的准确率比逻辑回归要高：
from caffe import layers as L
from caffe import params as P
def nonlinear_net(hdf5, batch_size):
n = caffe.NetSpec()
n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
n.ip1 = L.InnerProduct(n.data, num_output=40, weight_filler=dict(type='xavier'))
n.relu1 = L.ReLU(n.ip1, in_place=True)
n.ip2 = L.InnerProduct(n.ip1, num_output=2, weight_filler=dict(type='xavier'))
n.accuracy = L.Accuracy(n.ip2, n.label)
n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
return n.to_proto()
with open('examples/hdf5_classification/nonlinear_auto_train.prototxt', 'w') as f:
f.write(str(nonlinear_net('examples/hdf5_classification/data/train.txt', 10)))
with open('examples/hdf5_classification/nonlinear_auto_test.prototxt', 'w') as f:
f.write(str(nonlinear_net('examples/hdf5_classification/data/test.txt', 10)))
caffe.set_mode_cpu()
solver = caffe.get_solver('examples/hdf5_classification/nonlinear_solver.prototxt')
solver.solve()
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
solver.test_nets[0].forward()
accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
print("Accuracy: {:.3f}".format(accuracy))
Accuracy: 0.832
Accuracy: 0.832
Accuracy: 0.832
Accuracy: 0.831
1 loops, best of 3: 386 ms per loop
同样也用命令行看一下模型文件和求解过程：
!./build/tools/caffe train -solver examples/hdf5_classification/nonlinear_solver.prototxt
I0318 00:58:43.336922
caffe.cpp:117] Use CPU.
I0318 00:58:43.654698
caffe.cpp:121] Starting Optimization
I0318 00:58:43.654747
solver.cpp:32] Initializing solver from parameters:
train_net: "examples/hdf5_classification/nonlinear_auto_train.prototxt"
test_net: "examples/hdf5_classification/nonlinear_auto_test.prototxt"
test_iter: 250
test_interval: 1000
base_lr: 0.01
display: 1000
max_iter: 10000
lr_policy: "step"
gamma: 0.1
momentum: 0.9
weight_decay: 0.0005
stepsize: 5000
snapshot: 10000
snapshot_prefix: "examples/hdf5_classification/data/train"
solver_mode: CPU
I0318 00:58:43.654855
solver.cpp:61] Creating training net from train_net file: examples/hdf5_classification/nonlinear_auto_train.prototxt
I0318 00:58:43.655004
net.cpp:42] Initializing net from parameters:
phase: TRAIN
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/train.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 40
weight_filler {
type: "xavier"
name: "relu1"
type: "ReLU"
bottom: "ip1"
top: "ip1"
name: "ip2"
type: "InnerProduct"
bottom: "ip1"
top: "ip2"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip2"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip2"
bottom: "label"
top: "loss"
I0318 00:58:43.655120
layer_factory.hpp:74] Creating layer data
I0318 00:58:43.655139
net.cpp:84] Creating Layer data
I0318 00:58:43.655264
net.cpp:338] data -& data
I0318 00:58:43.655297
net.cpp:338] data -& label
I0318 00:58:43.655310
net.cpp:113] Setting up data
I0318 00:58:43.655318
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/train.txt
I0318 00:58:43.655365
hdf5_data_layer.cpp:80] Number of HDF5 files: 2
I0318 00:58:43.657317
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:43.657342
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657356
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:43.657373
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:43.657384
net.cpp:380] label_data_1_split &- label
I0318 00:58:43.657395
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:43.657407
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:43.657418
net.cpp:113] Setting up label_data_1_split
I0318 00:58:43.657426
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657433
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657441
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:43.657451
net.cpp:84] Creating Layer ip1
I0318 00:58:43.657459
net.cpp:380] ip1 &- data
I0318 00:58:43.657467
net.cpp:338] ip1 -& ip1
I0318 00:58:43.657479
net.cpp:113] Setting up ip1
I0318 00:58:43.662454
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.662477
layer_factory.hpp:74] Creating layer relu1
I0318 00:58:43.662497
net.cpp:84] Creating Layer relu1
I0318 00:58:43.662508
net.cpp:380] relu1 &- ip1
I0318 00:58:43.662520
net.cpp:327] relu1 -& ip1 (in-place)
I0318 00:58:43.662530
net.cpp:113] Setting up relu1
I0318 00:58:43.662539
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.662546
layer_factory.hpp:74] Creating layer ip2
I0318 00:58:43.662555
net.cpp:84] Creating Layer ip2
I0318 00:58:43.662562
net.cpp:380] ip2 &- ip1
I0318 00:58:43.662571
net.cpp:338] ip2 -& ip2
I0318 00:58:43.662580
net.cpp:113] Setting up ip2
I0318 00:58:43.662595
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662606
layer_factory.hpp:74] Creating layer ip2_ip2_0_split
I0318 00:58:43.662654
net.cpp:84] Creating Layer ip2_ip2_0_split
I0318 00:58:43.662665
net.cpp:380] ip2_ip2_0_split &- ip2
I0318 00:58:43.662678
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_0
I0318 00:58:43.662689
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_1
I0318 00:58:43.662698
net.cpp:113] Setting up ip2_ip2_0_split
I0318 00:58:43.662706
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662714
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662722
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:43.662734
net.cpp:84] Creating Layer accuracy
I0318 00:58:43.662740
net.cpp:380] accuracy &- ip2_ip2_0_split_0
I0318 00:58:43.662749
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:43.662756
net.cpp:338] accuracy -& accuracy
I0318 00:58:43.662766
net.cpp:113] Setting up accuracy
I0318 00:58:43.662818
net.cpp:120] Top shape: (1)
I0318 00:58:43.662827
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.662839
net.cpp:84] Creating Layer loss
I0318 00:58:43.662847
net.cpp:380] loss &- ip2_ip2_0_split_1
I0318 00:58:43.662854
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:43.662863
net.cpp:338] loss -& loss
I0318 00:58:43.662873
net.cpp:113] Setting up loss
I0318 00:58:43.662883
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.662901
net.cpp:120] Top shape: (1)
I0318 00:58:43.662909
net.cpp:122]
with loss weight 1
I0318 00:58:43.662922
net.cpp:167] loss needs backward computation.
I0318 00:58:43.662930
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:43.662936
net.cpp:167] ip2_ip2_0_split needs backward computation.
I0318 00:58:43.662942
net.cpp:167] ip2 needs backward computation.
I0318 00:58:43.662976
net.cpp:167] relu1 needs backward computation.
I0318 00:58:43.662988
net.cpp:167] ip1 needs backward computation.
I0318 00:58:43.662997
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:43.663003
net.cpp:169] data does not need backward computation.
I0318 00:58:43.663009
net.cpp:205] This network produces output accuracy
I0318 00:58:43.663017
net.cpp:205] This network produces output loss
I0318 00:58:43.663028
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:43.663035
net.cpp:217] Network initialization done.
I0318 00:58:43.663041
net.cpp:218] Memory required for data: 3728
I0318 00:58:43.663158
solver.cpp:154] Creating test net (#0) specified by test_net file: examples/hdf5_classification/nonlinear_auto_test.prototxt
I0318 00:58:43.663179
net.cpp:42] Initializing net from parameters:
phase: TEST
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/test.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 40
weight_filler {
type: "xavier"
name: "relu1"
type: "ReLU"
bottom: "ip1"
top: "ip1"
name: "ip2"
type: "InnerProduct"
bottom: "ip1"
top: "ip2"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip2"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip2"
bottom: "label"
top: "loss"
I0318 00:58:43.663349
layer_factory.hpp:74] Creating layer data
I0318 00:58:43.663365
net.cpp:84] Creating Layer data
I0318 00:58:43.663373
net.cpp:338] data -& data
I0318 00:58:43.663385
net.cpp:338] data -& label
I0318 00:58:43.663396
net.cpp:113] Setting up data
I0318 00:58:43.663422
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/test.txt
I0318 00:58:43.663457
hdf5_data_layer.cpp:80] Number of HDF5 files: 1
I0318 00:58:43.664719
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:43.664739
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664754
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:43.664772
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:43.664783
net.cpp:380] label_data_1_split &- label
I0318 00:58:43.664791
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:43.664803
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:43.664813
net.cpp:113] Setting up label_data_1_split
I0318 00:58:43.664822
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664829
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664837
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:43.664846
net.cpp:84] Creating Layer ip1
I0318 00:58:43.664854
net.cpp:380] ip1 &- data
I0318 00:58:43.664862
net.cpp:338] ip1 -& ip1
I0318 00:58:43.664875
net.cpp:113] Setting up ip1
I0318 00:58:43.664901
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.664924
layer_factory.hpp:74] Creating layer relu1
I0318 00:58:43.664945
net.cpp:84] Creating Layer relu1
I0318 00:58:43.664958
net.cpp:380] relu1 &- ip1
I0318 00:58:43.664966
net.cpp:327] relu1 -& ip1 (in-place)
I0318 00:58:43.664975
net.cpp:113] Setting up relu1
I0318 00:58:43.664983
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.664990
layer_factory.hpp:74] Creating layer ip2
I0318 00:58:43.665000
net.cpp:84] Creating Layer ip2
I0318 00:58:43.665006
net.cpp:380] ip2 &- ip1
I0318 00:58:43.665015
net.cpp:338] ip2 -& ip2
I0318 00:58:43.665030
net.cpp:113] Setting up ip2
I0318 00:58:43.665052
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665066
layer_factory.hpp:74] Creating layer ip2_ip2_0_split
I0318 00:58:43.665077
net.cpp:84] Creating Layer ip2_ip2_0_split
I0318 00:58:43.665086
net.cpp:380] ip2_ip2_0_split &- ip2
I0318 00:58:43.665093
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_0
I0318 00:58:43.665103
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_1
I0318 00:58:43.665113
net.cpp:113] Setting up ip2_ip2_0_split
I0318 00:58:43.665122
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665128
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665137
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:43.665144
net.cpp:84] Creating Layer accuracy
I0318 00:58:43.665153
net.cpp:380] accuracy &- ip2_ip2_0_split_0
I0318 00:58:43.665168
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:43.665180
net.cpp:338] accuracy -& accuracy
I0318 00:58:43.665192
net.cpp:113] Setting up accuracy
I0318 00:58:43.665200
net.cpp:120] Top shape: (1)
I0318 00:58:43.665207
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.665216
net.cpp:84] Creating Layer loss
I0318 00:58:43.665223
net.cpp:380] loss &- ip2_ip2_0_split_1
I0318 00:58:43.665230
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:43.665241
net.cpp:338] loss -& loss
I0318 00:58:43.665251
net.cpp:113] Setting up loss
I0318 00:58:43.665259
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.665273
net.cpp:120] Top shape: (1)
I0318 00:58:43.665282
net.cpp:122]
with loss weight 1
I0318 00:58:43.665290
net.cpp:167] loss needs backward computation.
I0318 00:58:43.665338
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:43.665351
net.cpp:167] ip2_ip2_0_split needs backward computation.
I0318 00:58:43.665380
net.cpp:167] ip2 needs backward computation.
I0318 00:58:43.665387
net.cpp:167] relu1 needs backward computation.
I0318 00:58:43.665393
net.cpp:167] ip1 needs backward computation.
I0318 00:58:43.665400
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:43.665407
net.cpp:169] data does not need backward computation.
I0318 00:58:43.665415
net.cpp:205] This network produces output accuracy
I0318 00:58:43.665421
net.cpp:205] This network produces output loss
I0318 00:58:43.665431
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:43.665441
net.cpp:217] Network initialization done.
I0318 00:58:43.665446
net.cpp:218] Memory required for data: 3728
I0318 00:58:43.665534
solver.cpp:42] Solver scaffolding done.
I0318 00:58:43.665568
solver.cpp:222] Solving
I0318 00:58:43.665577
solver.cpp:223] Learning Rate Policy: step
I0318 00:58:43.665586
solver.cpp:266] Iteration 0, Testing net (#0)
I0318 00:58:43.683938
solver.cpp:315]
Test net output #0: accuracy = 0.5184
I0318 00:58:43.683981
solver.cpp:315]
Test net output #1: loss = 0.716141 (* 1 = 0.716141 loss)
I0318 00:58:43.684236
solver.cpp:189] Iteration 0, loss = 0.764954
I0318 00:58:43.684267
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:43.684285
solver.cpp:204]
Train net output #1: loss = 0.764954 (* 1 = 0.764954 loss)
I0318 00:58:43.684305
solver.cpp:464] Iteration 0, lr = 0.01
I0318 00:58:43.714700
solver.cpp:266] Iteration 1000, Testing net (#0)
I0318 00:58:43.721762
solver.cpp:315]
Test net output #0: accuracy = 0.8168
I0318 00:58:43.721818
solver.cpp:315]
Test net output #1: loss = 0.434918 (* 1 = 0.434918 loss)
I0318 00:58:43.721899
solver.cpp:189] Iteration 1000, loss = 0.282425
I0318 00:58:43.721917
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.721932
solver.cpp:204]
Train net output #1: loss = 0.282426 (* 1 = 0.282426 loss)
I0318 00:58:43.721942
solver.cpp:464] Iteration 1000, lr = 0.01
I0318 00:58:43.750509
solver.cpp:266] Iteration 2000, Testing net (#0)
I0318 00:58:43.754590
solver.cpp:315]
Test net output #0: accuracy = 0.8224
I0318 00:58:43.754621
solver.cpp:315]
Test net output #1: loss = 0.416874 (* 1 = 0.416874 loss)
I0318 00:58:43.754660
solver.cpp:189] Iteration 2000, loss = 0.51988
I0318 00:58:43.754672
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.754683
solver.cpp:204]
Train net output #1: loss = 0.51988 (* 1 = 0.51988 loss)
I0318 00:58:43.754690
solver.cpp:464] Iteration 2000, lr = 0.01
I0318 00:58:43.782609
solver.cpp:266] Iteration 3000, Testing net (#0)
I0318 00:58:43.789728
solver.cpp:315]
Test net output #0: accuracy = 0.8176
I0318 00:58:43.789777
solver.cpp:315]
Test net output #1: loss = 0.415907 (* 1 = 0.415907 loss)
I0318 00:58:43.790487
solver.cpp:189] Iteration 3000, loss = 0.5093
I0318 00:58:43.790510
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.790530
solver.cpp:204]
Train net output #1: loss = 0.509301 (* 1 = 0.509301 loss)
I0318 00:58:43.790544
solver.cpp:464] Iteration 3000, lr = 0.01
I0318 00:58:43.817451
solver.cpp:266] Iteration 4000, Testing net (#0)
I0318 00:58:43.821740
solver.cpp:315]
Test net output #0: accuracy = 0.8252
I0318 00:58:43.821770
solver.cpp:315]
Test net output #1: loss = 0.409124 (* 1 = 0.409124 loss)
I0318 00:58:43.821822
solver.cpp:189] Iteration 4000, loss = 0.284815
I0318 00:58:43.821835
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.821846
solver.cpp:204]
Train net output #1: loss = 0.284815 (* 1 = 0.284815 loss)
I0318 00:58:43.821890
solver.cpp:464] Iteration 4000, lr = 0.01
I0318 00:58:43.847015
solver.cpp:266] Iteration 5000, Testing net (#0)
I0318 00:58:43.852102
solver.cpp:315]
Test net output #0: accuracy = 0.8256
I0318 00:58:43.852145
solver.cpp:315]
Test net output #1: loss = 0.404445 (* 1 = 0.404445 loss)
I0318 00:58:43.852188
solver.cpp:189] Iteration 5000, loss = 0.511566
I0318 00:58:43.852200
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.852210
solver.cpp:204]
Train net output #1: loss = 0.511566 (* 1 = 0.511566 loss)
I0318 00:58:43.852219
solver.cpp:464] Iteration 5000, lr = 0.001
I0318 00:58:43.876060
solver.cpp:266] Iteration 6000, Testing net (#0)
I0318 00:58:43.880080
solver.cpp:315]
Test net output #0: accuracy = 0.8328
I0318 00:58:43.880105
solver.cpp:315]
Test net output #1: loss = 0.396847 (* 1 = 0.396847 loss)
I0318 00:58:43.880700
solver.cpp:189] Iteration 6000, loss = 0.397858
I0318 00:58:43.880718
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.880729
solver.cpp:204]
Train net output #1: loss = 0.397858 (* 1 = 0.397858 loss)
I0318 00:58:43.880738
solver.cpp:464] Iteration 6000, lr = 0.001
I0318 00:58:43.913795
solver.cpp:266] Iteration 7000, Testing net (#0)
I0318 00:58:43.917851
solver.cpp:315]
Test net output #0: accuracy = 0.8316
I0318 00:58:43.917876
solver.cpp:315]
Test net output #1: loss = 0.398135 (* 1 = 0.398135 loss)
I0318 00:58:43.917956
solver.cpp:189] Iteration 7000, loss = 0.243849
I0318 00:58:43.917971
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.917989
solver.cpp:204]
Train net output #1: loss = 0.243849 (* 1 = 0.243849 loss)
I0318 00:58:43.918002
solver.cpp:464] Iteration 7000, lr = 0.001
I0318 00:58:43.943681
solver.cpp:266] Iteration 8000, Testing net (#0)
I0318 00:58:43.947589
solver.cpp:315]
Test net output #0: accuracy = 0.8312
I0318 00:58:43.947615
solver.cpp:315]
Test net output #1: loss = 0.394763 (* 1 = 0.394763 loss)
I0318 00:58:43.947651
solver.cpp:189] Iteration 8000, loss = 0.513399
I0318 00:58:43.947664
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.947674
solver.cpp:204]
Train net output #1: loss = 0.513399 (* 1 = 0.513399 loss)
I0318 00:58:43.947682
solver.cpp:464] Iteration 8000, lr = 0.001
I0318 00:58:43.973080
solver.cpp:266] Iteration 9000, Testing net (#0)
I0318 00:58:43.977033
solver.cpp:315]
Test net output #0: accuracy = 0.834
I0318 00:58:43.977056
solver.cpp:315]
Test net output #1: loss = 0.395663 (* 1 = 0.395663 loss)
I0318 00:58:43.977710
solver.cpp:189] Iteration 9000, loss = 0.399341
I0318 00:58:43.977735
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.977746
solver.cpp:204]
Train net output #1: loss = 0.399342 (* 1 = 0.399342 loss)
I0318 00:58:43.977756
solver.cpp:464] Iteration 9000, lr = 0.001
I0318 00:58:44.003437
solver.cpp:334] Snapshotting to examples/hdf5_classification/data/train_iter_10000.caffemodel
I0318 00:58:44.003702
solver.cpp:342] Snapshotting solver state to examples/hdf5_classification/data/train_iter_10000.solverstate
I0318 00:58:44.003850
solver.cpp:248] Iteration 10000, loss = 0.244639
I0318 00:58:44.003871
solver.cpp:266] Iteration 10000, Testing net (#0)
I0318 00:58:44.008216
solver.cpp:315]
Test net output #0: accuracy = 0.8308
I0318 00:58:44.008252
solver.cpp:315]
Test net output #1: loss = 0.397291 (* 1 = 0.397291 loss)
I0318 00:58:44.008262
solver.cpp:253] Optimization Done.
I0318 00:58:44.008270
caffe.cpp:134] Optimization Done.
最后把这些数据删掉：
# Clean up (comment this out if you want to examine the hdf5_classification/data directory).
shutil.rmtree(dirname)
参考知识库
* 以上用户言论只代表其个人观点，不代表CSDN网站的观点或立场
访问：142348次
积分：1973
积分：1973
排名：第17934名
原创：36篇
评论：182条
(1)(7)(1)(1)(2)(1)(10)(2)(5)(7)(5)(1)}
使用caffe的HDF5数据完成回归任务
&&& 一直在研究如何用caffe做行人检测问题，然而参考那些经典结构比如faster-rcnn等，都是自定义的caffe层来完成的检测任务。这些都要求对caffe框架有一定程度的了解。最近看到了如何用caffe完成回归的任务，就想把检测问题当成回归问题来解决。
&&& 我们把行人检测问题当成回归来看待，就需要限制检出目标的个数，因为我们的输出个数是固定的。所以，这里我假定每张图片最多检出的目标个数为2。即每个目标用4个值来表示其位置信息（中心位置坐标x,y。BBox的宽和高），则网络的最后输出是8个值。
制作HDF5数据
&&& 这里我们使用HDF5格式的数据来完成我们的回归任务，那么首先我们需要的是制作h5格式的数据。这里以VOC数据集为例。下面是制作HDF5格式数据的python代码。
import h5py
import caffe
import xml.etree.ElementTree as ET
import cv2
import time
import math
from os.path import join, exists
import numpy as np
def convert(size, box):
dw = 1./size[0]
dh = 1./size[1]
x = (box[0] + box[1])/2.0
y = (box[2] + box[3])/2.0
w = box[1] - box[0]
h = box[3] - box[2]
return (x,y,w,h)
def shuffle_in_unison_scary(a, b):
rng_state = np.random.get_state()
np.random.shuffle(a)
np.random.set_state(rng_state)
np.random.shuffle(b)
def processImage(imgs):
imgs = imgs.astype(np.float32)
for i, img in enumerate(imgs):
m = img.mean()
s = img.std()
imgs[i] = (img - m) / s
return imgs
TrainImgDir = 'F:/GenerateHDF5/trainImage'
TrainLabelDir = 'F:/GenerateHDF5/trainLabels'
TestImgDir = 'F:/GenerateHDF5/testImg'
TestLabelDir = 'F:/GenerateHDF5/testLabels'
InImg = []
InBBox = []
for rootDir,dirs,files in os.walk(TestLabelDir):
for file in files:
file_name = file.split('.')[0]
full_file_name = '%s%s'%(file_name,'.jpg')
full_file_dir = '%s/%s'%(TestImgDir,full_file_name)
Img = cv2.imread(full_file_dir,cv2.CV_LOAD_IMAGE_GRAYSCALE)
xml_file = open(&%s/%s&%(rootDir,file))
tree = ET.parse(xml_file)
root = tree.getroot()
size = root.find('size')
w = int(size.find('width').text)
h = int(size.find('height').text)
landmark = np.zeros(8)
for obj in root.iter('object'):
count = count + 1
if count == 3:
xmlbox = obj.find('bndbox')
b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
bb = convert((w,h), b)
landmark[(count-1)*4+0]=bb[0]
landmark[(count-1)*4+1]=bb[1]
landmark[(count-1)*4+2]=bb[2]
landmark[(count-1)*4+3]=bb[3]
InBBox.append(landmark.reshape(8))
Img = cv2.resize(Img,(h,w))
InImg.append(Img.reshape((1,h,w)))
InImg, InBBox = np.asarray(InImg), np.asarray(InBBox)
InImg = processImage(InImg)
shuffle_in_unison_scary(InImg, InBBox)
outputDir = 'hdf5/'
HDF5_file_name = 'hdf5_test.h5'
if not os.path.exists(outputDir):
os.makedirs(outputDir)
output = join(outputDir,HDF5_file_name)
with h5py.File(output, 'w') as h5:
h5['data'] = InImg.astype(np.float32)
h5['labels'] = InBBox.astype(np.float32)
h5.close()
这里注意一点，所有的BBox数据都要做归一化操作，即所有坐标要除以图片对应的宽高。据说，这样做能使最后得到的结果更好。
制作好了HDF5数据后，注意每个H5文件大小不能超过2G（这是caffe的规定，如果一个文件超过2G，请分开制作多个）。然后建立一个TXT文件，文件里写上所有H5文件的绝对路径，比如我这里建立的文件是list_train.txt。然后我只有一个H5文件，即hdf5_train.h5。所以我的list_train.txt文件中的内容就是/home/XXX/caffe/model/hdf5/hdf5_train.h5
配置solver文件
接下来是caffe的solver文件，这个文件没有什么区别，
test_iter: 20
test_interval: 70
base_lr: 0.
display: 9
max_iter: 210000
lr_policy: &step&
gamma: 0.1
momentum: 0.9
weight_decay: 0.0001
stepsize: 700
snapshot: 500
snapshot_prefix: &snapshot&
solver_mode: GPU
net: &train_val.prototxt&
solver_type: SGD
配置train_val.prototxt文件
接下来是网络的train_val.prototxt文件。这是caffe的网络结构文件，我们这里以LeNet网络为例，我这里是这样的：
name: &LeNet&
name: &data&
type: &HDF5Data&
top: &data&
top: &labels&
phase: TRAIN
hdf5_data_param {
source: &list_train.txt&
batch_size: 50
name: &data&
type: &HDF5Data&
top: &data&
top: &labels&
phase: TEST
hdf5_data_param {
source: &list_test.txt&
batch_size: 50
name: &conv1&
type: &Convolution&
bottom: &scaled&
top: &conv1&
lr_mult: 1.0
lr_mult: 2.0
convolution_param {
num_output: 20
kernel_size: 5
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &pool1&
type: &Pooling&
bottom: &conv1&
top: &pool1&
pooling_param {
kernel_size: 2
name: &conv2&
type: &Convolution&
bottom: &pool1&
top: &conv2&
lr_mult: 1.0
lr_mult: 2.0
convolution_param {
num_output: 50
kernel_size: 5
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &pool2&
type: &Pooling&
bottom: &conv2&
top: &pool2&
pooling_param {
kernel_size: 2
name: &ip1&
type: &InnerProduct&
bottom: &pool2&
top: &ip1&
lr_mult: 1.0
lr_mult: 2.0
inner_product_param {
num_output: 500
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &relu1&
type: &ReLU&
bottom: &ip1&
top: &ip1&
name: &ip2&
type: &InnerProduct&
bottom: &ip1&
top: &ip2&
lr_mult: 1.0
lr_mult: 2.0
inner_product_param {
num_output: 8
weight_filler {
type: &xavier&
bias_filler {
type: &constant&
name: &error&
type: &EuclideanLoss&
bottom: &ip2&
bottom: &labels&
top: &error&
phase: TEST
name: &loss&
type: &EuclideanLoss&
bottom: &ip2&
bottom: &labels&
top: &loss&
phase: TRAIN
这里注意的是，最后的一层全连接层，输出的num_output应该是你label的维度，我这里是8。然后最后的loss计算，我使用的是欧氏距离的loss，也可以试着用其他类型的loss。
按照以上步骤配置好了，最后就是训练了。在控制台中输入以下指令来训练我们的数据：
./cafferoot/caffe/tools/caffe train --solver=solver.prototxt
可能是我数据源的问题，我的loss一开始非常大，然后一直降不下来。也有可能是LeNet本身网络性能就不好。关于网络的性能还需要另外再想办法提升。
版权所有 爱编程 (C) Copyright 2012. . All Rights Reserved.
闽ICP备号-3
微信扫一扫关注爱编程，每天为您推送一篇经典技术文章。1836人阅读
caffe（15）
虽然Caffe是设计给深度网络的，但它也可以表示“浅层”网络模型用于分类，例如逻辑回归。下面在合成数据上做简单的逻辑回归，生成并保存为HDF5，来向Caffe输入向量。模型完成后，添加层来提高准确率。Caffe完成的主要内容就是：定义模型，实验，应用。
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
os.chdir('..')
import sys
sys.path.insert(0, './python')
import caffe
import h5py
import shutil
import tempfile
import sklearn
import sklearn.datasets
import sklearn.linear_model
import pandas as pd
合成10,000个4向量的数据集，用于有2个有用特征和2个噪声特征的二值分类。
X, y = sklearn.datasets.make_classification(
n_samples=10000, n_features=4, n_redundant=0, n_informative=2,
n_clusters_per_class=2, hypercube=False, random_state=0
# Split into train and test
X, Xt, y, yt = sklearn.cross_validation.train_test_split(X, y)
# Visualize sample of the data
ind = np.random.permutation(X.shape[0])[:1000]
df = pd.DataFrame(X[ind])
_ = pd.scatter_matrix(df, figsize=(9, 9), diagonal='kde', marker='o', s=40, alpha=.4, c=y[ind])
用于学习scikit-learn逻辑回归的方法是随机梯度下降（SGD），下面检测分类器的准确性：
clf = sklearn.linear_model.SGDClassifier(
loss='log', n_iter=1000, penalty='l2', alpha=1e-3, class_weight='auto')
clf.fit(X, y)
yt_pred = clf.predict(Xt)
print('Accuracy: {:.3f}'.format(sklearn.metrics.accuracy_score(yt, yt_pred)))
Accuracy: 0.783
Accuracy: 0.783
Accuracy: 0.783
Accuracy: 0.783
1 loops, best of 3: 508 ms per loop
将这个数据集保存到HDF5中，方便Caffe载入：
dirname = os.path.abspath('./examples/hdf5_classification/data')
if not os.path.exists(dirname):
os.makedirs(dirname)
train_filename = os.path.join(dirname, 'train.h5')
test_filename = os.path.join(dirname, 'test.h5')
with h5py.File(train_filename, 'w') as f:
f['data'] = X
f['label'] = y.astype(np.float32)
with open(os.path.join(dirname, 'train.txt'), 'w') as f:
f.write(train_filename + '\n')
f.write(train_filename + '\n')
comp_kwargs = {'compression': 'gzip', 'compression_opts': 1}
with h5py.File(test_filename, 'w') as f:
f.create_dataset('data', data=Xt, **comp_kwargs)
f.create_dataset('label', data=yt.astype(np.float32), **comp_kwargs)
with open(os.path.join(dirname, 'test.txt'), 'w') as f:
f.write(test_filename + '\n')
通过Python网络规范来定义Caffe中的逻辑回归，如果不想手工编写protobuf模型的话这是一个好方法：
from caffe import layers as L
from caffe import params as P
def logreg(hdf5, batch_size):
n = caffe.NetSpec()
n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
n.ip1 = L.InnerProduct(n.data, num_output=2, weight_filler=dict(type='xavier'))
n.accuracy = L.Accuracy(n.ip1, n.label)
n.loss = L.SoftmaxWithLoss(n.ip1, n.label)
return n.to_proto()
with open('examples/hdf5_classification/logreg_auto_train.prototxt', 'w') as f:
f.write(str(logreg('examples/hdf5_classification/data/train.txt', 10)))
with open('examples/hdf5_classification/logreg_auto_test.prototxt', 'w') as f:
f.write(str(logreg('examples/hdf5_classification/data/test.txt', 10)))
现在开始学习与评估Python中的Caffe逻辑回归：
caffe.set_mode_cpu()
solver = caffe.get_solver('examples/hdf5_classification/solver.prototxt')
solver.solve()
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
solver.test_nets[0].forward()
accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
print("Accuracy: {:.3f}".format(accuracy))
Accuracy: 0.782
Accuracy: 0.782
Accuracy: 0.782
Accuracy: 0.782
1 loops, best of 3: 287 ms per loop
用命令行来运行，看一下具体的模型文件和求解过程：
!./build/tools/caffe train -solver examples/hdf5_classification/solver.prototxt
I0318 00:58:32.322571
caffe.cpp:117] Use CPU.
I0318 00:58:32.643163
caffe.cpp:121] Starting Optimization
I0318 00:58:32.643229
solver.cpp:32] Initializing solver from parameters:
train_net: "examples/hdf5_classification/logreg_auto_train.prototxt"
test_net: "examples/hdf5_classification/logreg_auto_test.prototxt"
test_iter: 250
test_interval: 1000
base_lr: 0.01
display: 1000
max_iter: 10000
lr_policy: "step"
gamma: 0.1
momentum: 0.9
weight_decay: 0.0005
stepsize: 5000
snapshot: 10000
snapshot_prefix: "examples/hdf5_classification/data/train"
solver_mode: CPU
I0318 00:58:32.643333
solver.cpp:61] Creating training net from train_net file: examples/hdf5_classification/logreg_auto_train.prototxt
I0318 00:58:32.643465
net.cpp:42] Initializing net from parameters:
phase: TRAIN
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/train.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip1"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip1"
bottom: "label"
top: "loss"
I0318 00:58:32.644197
layer_factory.hpp:74] Creating layer data
I0318 00:58:32.644219
net.cpp:84] Creating Layer data
I0318 00:58:32.644230
net.cpp:338] data -& data
I0318 00:58:32.644256
net.cpp:338] data -& label
I0318 00:58:32.644269
net.cpp:113] Setting up data
I0318 00:58:32.644278
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/train.txt
I0318 00:58:32.644327
hdf5_data_layer.cpp:80] Number of HDF5 files: 2
I0318 00:58:32.646458
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:32.646502
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646518
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:32.646538
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:32.646546
net.cpp:380] label_data_1_split &- label
I0318 00:58:32.646556
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:32.646569
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:32.646579
net.cpp:113] Setting up label_data_1_split
I0318 00:58:32.646586
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646595
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.646601
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:32.646615
net.cpp:84] Creating Layer ip1
I0318 00:58:32.646622
net.cpp:380] ip1 &- data
I0318 00:58:32.646664
net.cpp:338] ip1 -& ip1
I0318 00:58:32.646689
net.cpp:113] Setting up ip1
I0318 00:58:32.652330
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652371
layer_factory.hpp:74] Creating layer ip1_ip1_0_split
I0318 00:58:32.652393
net.cpp:84] Creating Layer ip1_ip1_0_split
I0318 00:58:32.652407
net.cpp:380] ip1_ip1_0_split &- ip1
I0318 00:58:32.652421
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_0
I0318 00:58:32.652467
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_1
I0318 00:58:32.652480
net.cpp:113] Setting up ip1_ip1_0_split
I0318 00:58:32.652489
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652498
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.652505
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:32.652521
net.cpp:84] Creating Layer accuracy
I0318 00:58:32.652534
net.cpp:380] accuracy &- ip1_ip1_0_split_0
I0318 00:58:32.652545
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:32.652562
net.cpp:338] accuracy -& accuracy
I0318 00:58:32.652577
net.cpp:113] Setting up accuracy
I0318 00:58:32.652590
net.cpp:120] Top shape: (1)
I0318 00:58:32.652642
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.652655
net.cpp:84] Creating Layer loss
I0318 00:58:32.652663
net.cpp:380] loss &- ip1_ip1_0_split_1
I0318 00:58:32.652672
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:32.652679
net.cpp:338] loss -& loss
I0318 00:58:32.652689
net.cpp:113] Setting up loss
I0318 00:58:32.652701
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.652716
net.cpp:120] Top shape: (1)
I0318 00:58:32.652724
net.cpp:122]
with loss weight 1
I0318 00:58:32.652740
net.cpp:167] loss needs backward computation.
I0318 00:58:32.652746
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:32.652753
net.cpp:167] ip1_ip1_0_split needs backward computation.
I0318 00:58:32.652760
net.cpp:167] ip1 needs backward computation.
I0318 00:58:32.652786
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:32.652801
net.cpp:169] data does not need backward computation.
I0318 00:58:32.652808
net.cpp:205] This network produces output accuracy
I0318 00:58:32.652815
net.cpp:205] This network produces output loss
I0318 00:58:32.652825
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:32.652833
net.cpp:217] Network initialization done.
I0318 00:58:32.652839
net.cpp:218] Memory required for data: 528
I0318 00:58:32.652964
solver.cpp:154] Creating test net (#0) specified by test_net file: examples/hdf5_classification/logreg_auto_test.prototxt
I0318 00:58:32.652986
net.cpp:42] Initializing net from parameters:
phase: TEST
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/test.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip1"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip1"
bottom: "label"
top: "loss"
I0318 00:58:32.653069
layer_factory.hpp:74] Creating layer data
I0318 00:58:32.653080
net.cpp:84] Creating Layer data
I0318 00:58:32.653090
net.cpp:338] data -& data
I0318 00:58:32.653128
net.cpp:338] data -& label
I0318 00:58:32.653146
net.cpp:113] Setting up data
I0318 00:58:32.653154
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/test.txt
I0318 00:58:32.653192
hdf5_data_layer.cpp:80] Number of HDF5 files: 1
I0318 00:58:32.654850
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:32.654897
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.654914
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:32.654933
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:32.654943
net.cpp:380] label_data_1_split &- label
I0318 00:58:32.654953
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:32.654966
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:32.654976
net.cpp:113] Setting up label_data_1_split
I0318 00:58:32.654985
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.654992
net.cpp:120] Top shape: 10 (10)
I0318 00:58:32.655000
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:32.655010
net.cpp:84] Creating Layer ip1
I0318 00:58:32.655017
net.cpp:380] ip1 &- data
I0318 00:58:32.655030
net.cpp:338] ip1 -& ip1
I0318 00:58:32.655041
net.cpp:113] Setting up ip1
I0318 00:58:32.655061
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655072
layer_factory.hpp:74] Creating layer ip1_ip1_0_split
I0318 00:58:32.655148
net.cpp:84] Creating Layer ip1_ip1_0_split
I0318 00:58:32.655159
net.cpp:380] ip1_ip1_0_split &- ip1
I0318 00:58:32.655170
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_0
I0318 00:58:32.655180
net.cpp:338] ip1_ip1_0_split -& ip1_ip1_0_split_1
I0318 00:58:32.655190
net.cpp:113] Setting up ip1_ip1_0_split
I0318 00:58:32.655199
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655206
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:32.655213
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:32.655223
net.cpp:84] Creating Layer accuracy
I0318 00:58:32.655230
net.cpp:380] accuracy &- ip1_ip1_0_split_0
I0318 00:58:32.655237
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:32.655251
net.cpp:338] accuracy -& accuracy
I0318 00:58:32.655259
net.cpp:113] Setting up accuracy
I0318 00:58:32.655267
net.cpp:120] Top shape: (1)
I0318 00:58:32.655340
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.655354
net.cpp:84] Creating Layer loss
I0318 00:58:32.655361
net.cpp:380] loss &- ip1_ip1_0_split_1
I0318 00:58:32.655369
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:32.655378
net.cpp:338] loss -& loss
I0318 00:58:32.655388
net.cpp:113] Setting up loss
I0318 00:58:32.655397
layer_factory.hpp:74] Creating layer loss
I0318 00:58:32.655414
net.cpp:120] Top shape: (1)
I0318 00:58:32.655422
net.cpp:122]
with loss weight 1
I0318 00:58:32.655438
net.cpp:167] loss needs backward computation.
I0318 00:58:32.655446
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:32.655455
net.cpp:167] ip1_ip1_0_split needs backward computation.
I0318 00:58:32.655462
net.cpp:167] ip1 needs backward computation.
I0318 00:58:32.655469
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:32.655477
net.cpp:169] data does not need backward computation.
I0318 00:58:32.655483
net.cpp:205] This network produces output accuracy
I0318 00:58:32.655489
net.cpp:205] This network produces output loss
I0318 00:58:32.655503
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:32.655511
net.cpp:217] Network initialization done.
I0318 00:58:32.655517
net.cpp:218] Memory required for data: 528
I0318 00:58:32.655547
solver.cpp:42] Solver scaffolding done.
I0318 00:58:32.655567
solver.cpp:222] Solving
I0318 00:58:32.655575
solver.cpp:223] Learning Rate Policy: step
I0318 00:58:32.655583
solver.cpp:266] Iteration 0, Testing net (#0)
I0318 00:58:32.683643
solver.cpp:315]
Test net output #0: accuracy = 0.3736
I0318 00:58:32.683686
solver.cpp:315]
Test net output #1: loss = 1.00555 (* 1 = 1.00555 loss)
I0318 00:58:32.683846
solver.cpp:189] Iteration 0, loss = 0.869394
I0318 00:58:32.683861
solver.cpp:204]
Train net output #0: accuracy = 0.3
I0318 00:58:32.683871
solver.cpp:204]
Train net output #1: loss = 0.869394 (* 1 = 0.869394 loss)
I0318 00:58:32.683883
solver.cpp:464] Iteration 0, lr = 0.01
I0318 00:58:32.698721
solver.cpp:266] Iteration 1000, Testing net (#0)
I0318 00:58:32.701917
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.701961
solver.cpp:315]
Test net output #1: loss = 0.590972 (* 1 = 0.590972 loss)
I0318 00:58:32.702014
solver.cpp:189] Iteration 1000, loss = 0.54742
I0318 00:58:32.702029
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:32.702041
solver.cpp:204]
Train net output #1: loss = 0.54742 (* 1 = 0.54742 loss)
I0318 00:58:32.702051
solver.cpp:464] Iteration 1000, lr = 0.01
I0318 00:58:32.718360
solver.cpp:266] Iteration 2000, Testing net (#0)
I0318 00:58:32.721529
solver.cpp:315]
Test net output #0: accuracy = 0.7696
I0318 00:58:32.721562
solver.cpp:315]
Test net output #1: loss = 0.593946 (* 1 = 0.593946 loss)
I0318 00:58:32.721593
solver.cpp:189] Iteration 2000, loss = 0.729569
I0318 00:58:32.721603
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.721613
solver.cpp:204]
Train net output #1: loss = 0.729569 (* 1 = 0.729569 loss)
I0318 00:58:32.721622
solver.cpp:464] Iteration 2000, lr = 0.01
I0318 00:58:32.740182
solver.cpp:266] Iteration 3000, Testing net (#0)
I0318 00:58:32.743494
solver.cpp:315]
Test net output #0: accuracy = 0.77
I0318 00:58:32.743544
solver.cpp:315]
Test net output #1: loss = 0.591229 (* 1 = 0.591229 loss)
I0318 00:58:32.744209
solver.cpp:189] Iteration 3000, loss = 0.406097
I0318 00:58:32.744231
solver.cpp:204]
Train net output #0: accuracy = 0.8
I0318 00:58:32.744249
solver.cpp:204]
Train net output #1: loss = 0.406096 (* 1 = 0.406096 loss)
I0318 00:58:32.744266
solver.cpp:464] Iteration 3000, lr = 0.01
I0318 00:58:32.764135
solver.cpp:266] Iteration 4000, Testing net (#0)
I0318 00:58:32.769110
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.769170
solver.cpp:315]
Test net output #1: loss = 0.590972 (* 1 = 0.590972 loss)
I0318 00:58:32.769223
solver.cpp:189] Iteration 4000, loss = 0.54742
I0318 00:58:32.769242
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:32.769255
solver.cpp:204]
Train net output #1: loss = 0.54742 (* 1 = 0.54742 loss)
I0318 00:58:32.769265
solver.cpp:464] Iteration 4000, lr = 0.01
I0318 00:58:32.785846
solver.cpp:266] Iteration 5000, Testing net (#0)
I0318 00:58:32.788722
solver.cpp:315]
Test net output #0: accuracy = 0.7696
I0318 00:58:32.788751
solver.cpp:315]
Test net output #1: loss = 0.593946 (* 1 = 0.593946 loss)
I0318 00:58:32.788811
solver.cpp:189] Iteration 5000, loss = 0.72957
I0318 00:58:32.788833
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.788846
solver.cpp:204]
Train net output #1: loss = 0.729569 (* 1 = 0.729569 loss)
I0318 00:58:32.788856
solver.cpp:464] Iteration 5000, lr = 0.001
I0318 00:58:32.804762
solver.cpp:266] Iteration 6000, Testing net (#0)
I0318 00:58:32.808061
solver.cpp:315]
Test net output #0: accuracy = 0.7856
I0318 00:58:32.808112
solver.cpp:315]
Test net output #1: loss = 0.59028 (* 1 = 0.59028 loss)
I0318 00:58:32.808732
solver.cpp:189] Iteration 6000, loss = 0.415444
I0318 00:58:32.808753
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:32.808773
solver.cpp:204]
Train net output #1: loss = 0.415444 (* 1 = 0.415444 loss)
I0318 00:58:32.808786
solver.cpp:464] Iteration 6000, lr = 0.001
I0318 00:58:32.827118
solver.cpp:266] Iteration 7000, Testing net (#0)
I0318 00:58:32.831614
solver.cpp:315]
Test net output #0: accuracy = 0.7848
I0318 00:58:32.831657
solver.cpp:315]
Test net output #1: loss = 0.589454 (* 1 = 0.589454 loss)
I0318 00:58:32.831707
solver.cpp:189] Iteration 7000, loss = 0.538038
I0318 00:58:32.831728
solver.cpp:204]
Train net output #0: accuracy = 0.8
I0318 00:58:32.831745
solver.cpp:204]
Train net output #1: loss = 0.538037 (* 1 = 0.538037 loss)
I0318 00:58:32.831759
solver.cpp:464] Iteration 7000, lr = 0.001
I0318 00:58:32.849634
solver.cpp:266] Iteration 8000, Testing net (#0)
I0318 00:58:32.852712
solver.cpp:315]
Test net output #0: accuracy = 0.7796
I0318 00:58:32.852748
solver.cpp:315]
Test net output #1: loss = 0.589365 (* 1 = 0.589365 loss)
I0318 00:58:32.852792
solver.cpp:189] Iteration 8000, loss = 0.684219
I0318 00:58:32.852840
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:32.852852
solver.cpp:204]
Train net output #1: loss = 0.684219 (* 1 = 0.684219 loss)
I0318 00:58:32.852861
solver.cpp:464] Iteration 8000, lr = 0.001
I0318 00:58:32.868440
solver.cpp:266] Iteration 9000, Testing net (#0)
I0318 00:58:32.871438
solver.cpp:315]
Test net output #0: accuracy = 0.7816
I0318 00:58:32.871461
solver.cpp:315]
Test net output #1: loss = 0.589656 (* 1 = 0.589656 loss)
I0318 00:58:32.872109
solver.cpp:189] Iteration 9000, loss = 0.421879
I0318 00:58:32.872131
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:32.872143
solver.cpp:204]
Train net output #1: loss = 0.421879 (* 1 = 0.421879 loss)
I0318 00:58:32.872153
solver.cpp:464] Iteration 9000, lr = 0.001
I0318 00:58:32.889981
solver.cpp:334] Snapshotting to examples/hdf5_classification/data/train_iter_10000.caffemodel
I0318 00:58:32.890224
solver.cpp:342] Snapshotting solver state to examples/hdf5_classification/data/train_iter_10000.solverstate
I0318 00:58:32.890362
solver.cpp:248] Iteration 10000, loss = 0.538933
I0318 00:58:32.890380
solver.cpp:266] Iteration 10000, Testing net (#0)
I0318 00:58:32.893728
solver.cpp:315]
Test net output #0: accuracy = 0.782
I0318 00:58:32.893757
solver.cpp:315]
Test net output #1: loss = 0.589366 (* 1 = 0.589366 loss)
I0318 00:58:32.893775
solver.cpp:253] Optimization Done.
I0318 00:58:32.893786
caffe.cpp:134] Optimization Done.
看一下上面的输出或者logreg_auto_train.prototxt，就会发现这个模型是简单的逻辑回归，可以通过引入输入权重和输出权重之间的非线性，使这个模型更好一点，变成两层的，把这个网络放在nonlinear_auto_train.prototxt中。新网络最终的准确率比逻辑回归要高：
from caffe import layers as L
from caffe import params as P
def nonlinear_net(hdf5, batch_size):
n = caffe.NetSpec()
n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
n.ip1 = L.InnerProduct(n.data, num_output=40, weight_filler=dict(type='xavier'))
n.relu1 = L.ReLU(n.ip1, in_place=True)
n.ip2 = L.InnerProduct(n.ip1, num_output=2, weight_filler=dict(type='xavier'))
n.accuracy = L.Accuracy(n.ip2, n.label)
n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
return n.to_proto()
with open('examples/hdf5_classification/nonlinear_auto_train.prototxt', 'w') as f:
f.write(str(nonlinear_net('examples/hdf5_classification/data/train.txt', 10)))
with open('examples/hdf5_classification/nonlinear_auto_test.prototxt', 'w') as f:
f.write(str(nonlinear_net('examples/hdf5_classification/data/test.txt', 10)))
caffe.set_mode_cpu()
solver = caffe.get_solver('examples/hdf5_classification/nonlinear_solver.prototxt')
solver.solve()
accuracy = 0
batch_size = solver.test_nets[0].blobs['data'].num
test_iters = int(len(Xt) / batch_size)
for i in range(test_iters):
solver.test_nets[0].forward()
accuracy += solver.test_nets[0].blobs['accuracy'].data
accuracy /= test_iters
print("Accuracy: {:.3f}".format(accuracy))
Accuracy: 0.832
Accuracy: 0.832
Accuracy: 0.832
Accuracy: 0.831
1 loops, best of 3: 386 ms per loop
同样也用命令行看一下模型文件和求解过程：
!./build/tools/caffe train -solver examples/hdf5_classification/nonlinear_solver.prototxt
I0318 00:58:43.336922
caffe.cpp:117] Use CPU.
I0318 00:58:43.654698
caffe.cpp:121] Starting Optimization
I0318 00:58:43.654747
solver.cpp:32] Initializing solver from parameters:
train_net: "examples/hdf5_classification/nonlinear_auto_train.prototxt"
test_net: "examples/hdf5_classification/nonlinear_auto_test.prototxt"
test_iter: 250
test_interval: 1000
base_lr: 0.01
display: 1000
max_iter: 10000
lr_policy: "step"
gamma: 0.1
momentum: 0.9
weight_decay: 0.0005
stepsize: 5000
snapshot: 10000
snapshot_prefix: "examples/hdf5_classification/data/train"
solver_mode: CPU
I0318 00:58:43.654855
solver.cpp:61] Creating training net from train_net file: examples/hdf5_classification/nonlinear_auto_train.prototxt
I0318 00:58:43.655004
net.cpp:42] Initializing net from parameters:
phase: TRAIN
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/train.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 40
weight_filler {
type: "xavier"
name: "relu1"
type: "ReLU"
bottom: "ip1"
top: "ip1"
name: "ip2"
type: "InnerProduct"
bottom: "ip1"
top: "ip2"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip2"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip2"
bottom: "label"
top: "loss"
I0318 00:58:43.655120
layer_factory.hpp:74] Creating layer data
I0318 00:58:43.655139
net.cpp:84] Creating Layer data
I0318 00:58:43.655264
net.cpp:338] data -& data
I0318 00:58:43.655297
net.cpp:338] data -& label
I0318 00:58:43.655310
net.cpp:113] Setting up data
I0318 00:58:43.655318
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/train.txt
I0318 00:58:43.655365
hdf5_data_layer.cpp:80] Number of HDF5 files: 2
I0318 00:58:43.657317
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:43.657342
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657356
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:43.657373
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:43.657384
net.cpp:380] label_data_1_split &- label
I0318 00:58:43.657395
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:43.657407
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:43.657418
net.cpp:113] Setting up label_data_1_split
I0318 00:58:43.657426
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657433
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.657441
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:43.657451
net.cpp:84] Creating Layer ip1
I0318 00:58:43.657459
net.cpp:380] ip1 &- data
I0318 00:58:43.657467
net.cpp:338] ip1 -& ip1
I0318 00:58:43.657479
net.cpp:113] Setting up ip1
I0318 00:58:43.662454
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.662477
layer_factory.hpp:74] Creating layer relu1
I0318 00:58:43.662497
net.cpp:84] Creating Layer relu1
I0318 00:58:43.662508
net.cpp:380] relu1 &- ip1
I0318 00:58:43.662520
net.cpp:327] relu1 -& ip1 (in-place)
I0318 00:58:43.662530
net.cpp:113] Setting up relu1
I0318 00:58:43.662539
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.662546
layer_factory.hpp:74] Creating layer ip2
I0318 00:58:43.662555
net.cpp:84] Creating Layer ip2
I0318 00:58:43.662562
net.cpp:380] ip2 &- ip1
I0318 00:58:43.662571
net.cpp:338] ip2 -& ip2
I0318 00:58:43.662580
net.cpp:113] Setting up ip2
I0318 00:58:43.662595
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662606
layer_factory.hpp:74] Creating layer ip2_ip2_0_split
I0318 00:58:43.662654
net.cpp:84] Creating Layer ip2_ip2_0_split
I0318 00:58:43.662665
net.cpp:380] ip2_ip2_0_split &- ip2
I0318 00:58:43.662678
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_0
I0318 00:58:43.662689
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_1
I0318 00:58:43.662698
net.cpp:113] Setting up ip2_ip2_0_split
I0318 00:58:43.662706
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662714
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.662722
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:43.662734
net.cpp:84] Creating Layer accuracy
I0318 00:58:43.662740
net.cpp:380] accuracy &- ip2_ip2_0_split_0
I0318 00:58:43.662749
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:43.662756
net.cpp:338] accuracy -& accuracy
I0318 00:58:43.662766
net.cpp:113] Setting up accuracy
I0318 00:58:43.662818
net.cpp:120] Top shape: (1)
I0318 00:58:43.662827
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.662839
net.cpp:84] Creating Layer loss
I0318 00:58:43.662847
net.cpp:380] loss &- ip2_ip2_0_split_1
I0318 00:58:43.662854
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:43.662863
net.cpp:338] loss -& loss
I0318 00:58:43.662873
net.cpp:113] Setting up loss
I0318 00:58:43.662883
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.662901
net.cpp:120] Top shape: (1)
I0318 00:58:43.662909
net.cpp:122]
with loss weight 1
I0318 00:58:43.662922
net.cpp:167] loss needs backward computation.
I0318 00:58:43.662930
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:43.662936
net.cpp:167] ip2_ip2_0_split needs backward computation.
I0318 00:58:43.662942
net.cpp:167] ip2 needs backward computation.
I0318 00:58:43.662976
net.cpp:167] relu1 needs backward computation.
I0318 00:58:43.662988
net.cpp:167] ip1 needs backward computation.
I0318 00:58:43.662997
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:43.663003
net.cpp:169] data does not need backward computation.
I0318 00:58:43.663009
net.cpp:205] This network produces output accuracy
I0318 00:58:43.663017
net.cpp:205] This network produces output loss
I0318 00:58:43.663028
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:43.663035
net.cpp:217] Network initialization done.
I0318 00:58:43.663041
net.cpp:218] Memory required for data: 3728
I0318 00:58:43.663158
solver.cpp:154] Creating test net (#0) specified by test_net file: examples/hdf5_classification/nonlinear_auto_test.prototxt
I0318 00:58:43.663179
net.cpp:42] Initializing net from parameters:
phase: TEST
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/hdf5_classification/data/test.txt"
batch_size: 10
name: "ip1"
type: "InnerProduct"
bottom: "data"
top: "ip1"
inner_product_param {
num_output: 40
weight_filler {
type: "xavier"
name: "relu1"
type: "ReLU"
bottom: "ip1"
top: "ip1"
name: "ip2"
type: "InnerProduct"
bottom: "ip1"
top: "ip2"
inner_product_param {
num_output: 2
weight_filler {
type: "xavier"
name: "accuracy"
type: "Accuracy"
bottom: "ip2"
bottom: "label"
top: "accuracy"
name: "loss"
type: "SoftmaxWithLoss"
bottom: "ip2"
bottom: "label"
top: "loss"
I0318 00:58:43.663349
layer_factory.hpp:74] Creating layer data
I0318 00:58:43.663365
net.cpp:84] Creating Layer data
I0318 00:58:43.663373
net.cpp:338] data -& data
I0318 00:58:43.663385
net.cpp:338] data -& label
I0318 00:58:43.663396
net.cpp:113] Setting up data
I0318 00:58:43.663422
hdf5_data_layer.cpp:66] Loading list of HDF5 filenames from: examples/hdf5_classification/data/test.txt
I0318 00:58:43.663457
hdf5_data_layer.cpp:80] Number of HDF5 files: 1
I0318 00:58:43.664719
net.cpp:120] Top shape: 10 4 (40)
I0318 00:58:43.664739
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664754
layer_factory.hpp:74] Creating layer label_data_1_split
I0318 00:58:43.664772
net.cpp:84] Creating Layer label_data_1_split
I0318 00:58:43.664783
net.cpp:380] label_data_1_split &- label
I0318 00:58:43.664791
net.cpp:338] label_data_1_split -& label_data_1_split_0
I0318 00:58:43.664803
net.cpp:338] label_data_1_split -& label_data_1_split_1
I0318 00:58:43.664813
net.cpp:113] Setting up label_data_1_split
I0318 00:58:43.664822
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664829
net.cpp:120] Top shape: 10 (10)
I0318 00:58:43.664837
layer_factory.hpp:74] Creating layer ip1
I0318 00:58:43.664846
net.cpp:84] Creating Layer ip1
I0318 00:58:43.664854
net.cpp:380] ip1 &- data
I0318 00:58:43.664862
net.cpp:338] ip1 -& ip1
I0318 00:58:43.664875
net.cpp:113] Setting up ip1
I0318 00:58:43.664901
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.664924
layer_factory.hpp:74] Creating layer relu1
I0318 00:58:43.664945
net.cpp:84] Creating Layer relu1
I0318 00:58:43.664958
net.cpp:380] relu1 &- ip1
I0318 00:58:43.664966
net.cpp:327] relu1 -& ip1 (in-place)
I0318 00:58:43.664975
net.cpp:113] Setting up relu1
I0318 00:58:43.664983
net.cpp:120] Top shape: 10 40 (400)
I0318 00:58:43.664990
layer_factory.hpp:74] Creating layer ip2
I0318 00:58:43.665000
net.cpp:84] Creating Layer ip2
I0318 00:58:43.665006
net.cpp:380] ip2 &- ip1
I0318 00:58:43.665015
net.cpp:338] ip2 -& ip2
I0318 00:58:43.665030
net.cpp:113] Setting up ip2
I0318 00:58:43.665052
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665066
layer_factory.hpp:74] Creating layer ip2_ip2_0_split
I0318 00:58:43.665077
net.cpp:84] Creating Layer ip2_ip2_0_split
I0318 00:58:43.665086
net.cpp:380] ip2_ip2_0_split &- ip2
I0318 00:58:43.665093
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_0
I0318 00:58:43.665103
net.cpp:338] ip2_ip2_0_split -& ip2_ip2_0_split_1
I0318 00:58:43.665113
net.cpp:113] Setting up ip2_ip2_0_split
I0318 00:58:43.665122
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665128
net.cpp:120] Top shape: 10 2 (20)
I0318 00:58:43.665137
layer_factory.hpp:74] Creating layer accuracy
I0318 00:58:43.665144
net.cpp:84] Creating Layer accuracy
I0318 00:58:43.665153
net.cpp:380] accuracy &- ip2_ip2_0_split_0
I0318 00:58:43.665168
net.cpp:380] accuracy &- label_data_1_split_0
I0318 00:58:43.665180
net.cpp:338] accuracy -& accuracy
I0318 00:58:43.665192
net.cpp:113] Setting up accuracy
I0318 00:58:43.665200
net.cpp:120] Top shape: (1)
I0318 00:58:43.665207
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.665216
net.cpp:84] Creating Layer loss
I0318 00:58:43.665223
net.cpp:380] loss &- ip2_ip2_0_split_1
I0318 00:58:43.665230
net.cpp:380] loss &- label_data_1_split_1
I0318 00:58:43.665241
net.cpp:338] loss -& loss
I0318 00:58:43.665251
net.cpp:113] Setting up loss
I0318 00:58:43.665259
layer_factory.hpp:74] Creating layer loss
I0318 00:58:43.665273
net.cpp:120] Top shape: (1)
I0318 00:58:43.665282
net.cpp:122]
with loss weight 1
I0318 00:58:43.665290
net.cpp:167] loss needs backward computation.
I0318 00:58:43.665338
net.cpp:169] accuracy does not need backward computation.
I0318 00:58:43.665351
net.cpp:167] ip2_ip2_0_split needs backward computation.
I0318 00:58:43.665380
net.cpp:167] ip2 needs backward computation.
I0318 00:58:43.665387
net.cpp:167] relu1 needs backward computation.
I0318 00:58:43.665393
net.cpp:167] ip1 needs backward computation.
I0318 00:58:43.665400
net.cpp:169] label_data_1_split does not need backward computation.
I0318 00:58:43.665407
net.cpp:169] data does not need backward computation.
I0318 00:58:43.665415
net.cpp:205] This network produces output accuracy
I0318 00:58:43.665421
net.cpp:205] This network produces output loss
I0318 00:58:43.665431
net.cpp:447] Collecting Learning Rate and Weight Decay.
I0318 00:58:43.665441
net.cpp:217] Network initialization done.
I0318 00:58:43.665446
net.cpp:218] Memory required for data: 3728
I0318 00:58:43.665534
solver.cpp:42] Solver scaffolding done.
I0318 00:58:43.665568
solver.cpp:222] Solving
I0318 00:58:43.665577
solver.cpp:223] Learning Rate Policy: step
I0318 00:58:43.665586
solver.cpp:266] Iteration 0, Testing net (#0)
I0318 00:58:43.683938
solver.cpp:315]
Test net output #0: accuracy = 0.5184
I0318 00:58:43.683981
solver.cpp:315]
Test net output #1: loss = 0.716141 (* 1 = 0.716141 loss)
I0318 00:58:43.684236
solver.cpp:189] Iteration 0, loss = 0.764954
I0318 00:58:43.684267
solver.cpp:204]
Train net output #0: accuracy = 0.5
I0318 00:58:43.684285
solver.cpp:204]
Train net output #1: loss = 0.764954 (* 1 = 0.764954 loss)
I0318 00:58:43.684305
solver.cpp:464] Iteration 0, lr = 0.01
I0318 00:58:43.714700
solver.cpp:266] Iteration 1000, Testing net (#0)
I0318 00:58:43.721762
solver.cpp:315]
Test net output #0: accuracy = 0.8168
I0318 00:58:43.721818
solver.cpp:315]
Test net output #1: loss = 0.434918 (* 1 = 0.434918 loss)
I0318 00:58:43.721899
solver.cpp:189] Iteration 1000, loss = 0.282425
I0318 00:58:43.721917
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.721932
solver.cpp:204]
Train net output #1: loss = 0.282426 (* 1 = 0.282426 loss)
I0318 00:58:43.721942
solver.cpp:464] Iteration 1000, lr = 0.01
I0318 00:58:43.750509
solver.cpp:266] Iteration 2000, Testing net (#0)
I0318 00:58:43.754590
solver.cpp:315]
Test net output #0: accuracy = 0.8224
I0318 00:58:43.754621
solver.cpp:315]
Test net output #1: loss = 0.416874 (* 1 = 0.416874 loss)
I0318 00:58:43.754660
solver.cpp:189] Iteration 2000, loss = 0.51988
I0318 00:58:43.754672
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.754683
solver.cpp:204]
Train net output #1: loss = 0.51988 (* 1 = 0.51988 loss)
I0318 00:58:43.754690
solver.cpp:464] Iteration 2000, lr = 0.01
I0318 00:58:43.782609
solver.cpp:266] Iteration 3000, Testing net (#0)
I0318 00:58:43.789728
solver.cpp:315]
Test net output #0: accuracy = 0.8176
I0318 00:58:43.789777
solver.cpp:315]
Test net output #1: loss = 0.415907 (* 1 = 0.415907 loss)
I0318 00:58:43.790487
solver.cpp:189] Iteration 3000, loss = 0.5093
I0318 00:58:43.790510
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.790530
solver.cpp:204]
Train net output #1: loss = 0.509301 (* 1 = 0.509301 loss)
I0318 00:58:43.790544
solver.cpp:464] Iteration 3000, lr = 0.01
I0318 00:58:43.817451
solver.cpp:266] Iteration 4000, Testing net (#0)
I0318 00:58:43.821740
solver.cpp:315]
Test net output #0: accuracy = 0.8252
I0318 00:58:43.821770
solver.cpp:315]
Test net output #1: loss = 0.409124 (* 1 = 0.409124 loss)
I0318 00:58:43.821822
solver.cpp:189] Iteration 4000, loss = 0.284815
I0318 00:58:43.821835
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.821846
solver.cpp:204]
Train net output #1: loss = 0.284815 (* 1 = 0.284815 loss)
I0318 00:58:43.821890
solver.cpp:464] Iteration 4000, lr = 0.01
I0318 00:58:43.847015
solver.cpp:266] Iteration 5000, Testing net (#0)
I0318 00:58:43.852102
solver.cpp:315]
Test net output #0: accuracy = 0.8256
I0318 00:58:43.852145
solver.cpp:315]
Test net output #1: loss = 0.404445 (* 1 = 0.404445 loss)
I0318 00:58:43.852188
solver.cpp:189] Iteration 5000, loss = 0.511566
I0318 00:58:43.852200
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.852210
solver.cpp:204]
Train net output #1: loss = 0.511566 (* 1 = 0.511566 loss)
I0318 00:58:43.852219
solver.cpp:464] Iteration 5000, lr = 0.001
I0318 00:58:43.876060
solver.cpp:266] Iteration 6000, Testing net (#0)
I0318 00:58:43.880080
solver.cpp:315]
Test net output #0: accuracy = 0.8328
I0318 00:58:43.880105
solver.cpp:315]
Test net output #1: loss = 0.396847 (* 1 = 0.396847 loss)
I0318 00:58:43.880700
solver.cpp:189] Iteration 6000, loss = 0.397858
I0318 00:58:43.880718
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.880729
solver.cpp:204]
Train net output #1: loss = 0.397858 (* 1 = 0.397858 loss)
I0318 00:58:43.880738
solver.cpp:464] Iteration 6000, lr = 0.001
I0318 00:58:43.913795
solver.cpp:266] Iteration 7000, Testing net (#0)
I0318 00:58:43.917851
solver.cpp:315]
Test net output #0: accuracy = 0.8316
I0318 00:58:43.917876
solver.cpp:315]
Test net output #1: loss = 0.398135 (* 1 = 0.398135 loss)
I0318 00:58:43.917956
solver.cpp:189] Iteration 7000, loss = 0.243849
I0318 00:58:43.917971
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.917989
solver.cpp:204]
Train net output #1: loss = 0.243849 (* 1 = 0.243849 loss)
I0318 00:58:43.918002
solver.cpp:464] Iteration 7000, lr = 0.001
I0318 00:58:43.943681
solver.cpp:266] Iteration 8000, Testing net (#0)
I0318 00:58:43.947589
solver.cpp:315]
Test net output #0: accuracy = 0.8312
I0318 00:58:43.947615
solver.cpp:315]
Test net output #1: loss = 0.394763 (* 1 = 0.394763 loss)
I0318 00:58:43.947651
solver.cpp:189] Iteration 8000, loss = 0.513399
I0318 00:58:43.947664
solver.cpp:204]
Train net output #0: accuracy = 0.7
I0318 00:58:43.947674
solver.cpp:204]
Train net output #1: loss = 0.513399 (* 1 = 0.513399 loss)
I0318 00:58:43.947682
solver.cpp:464] Iteration 8000, lr = 0.001
I0318 00:58:43.973080
solver.cpp:266] Iteration 9000, Testing net (#0)
I0318 00:58:43.977033
solver.cpp:315]
Test net output #0: accuracy = 0.834
I0318 00:58:43.977056
solver.cpp:315]
Test net output #1: loss = 0.395663 (* 1 = 0.395663 loss)
I0318 00:58:43.977710
solver.cpp:189] Iteration 9000, loss = 0.399341
I0318 00:58:43.977735
solver.cpp:204]
Train net output #0: accuracy = 0.9
I0318 00:58:43.977746
solver.cpp:204]
Train net output #1: loss = 0.399342 (* 1 = 0.399342 loss)
I0318 00:58:43.977756
solver.cpp:464] Iteration 9000, lr = 0.001
I0318 00:58:44.003437
solver.cpp:334] Snapshotting to examples/hdf5_classification/data/train_iter_10000.caffemodel
I0318 00:58:44.003702
solver.cpp:342] Snapshotting solver state to examples/hdf5_classification/data/train_iter_10000.solverstate
I0318 00:58:44.003850
solver.cpp:248] Iteration 10000, loss = 0.244639
I0318 00:58:44.003871
solver.cpp:266] Iteration 10000, Testing net (#0)
I0318 00:58:44.008216
solver.cpp:315]
Test net output #0: accuracy = 0.8308
I0318 00:58:44.008252
solver.cpp:315]
Test net output #1: loss = 0.397291 (* 1 = 0.397291 loss)
I0318 00:58:44.008262
solver.cpp:253] Optimization Done.
I0318 00:58:44.008270
caffe.cpp:134] Optimization Done.
最后把这些数据删掉：
# Clean up (comment this out if you want to examine the hdf5_classification/data directory).
shutil.rmtree(dirname)
参考知识库
* 以上用户言论只代表其个人观点，不代表CSDN网站的观点或立场
访问：142348次
积分：1973
积分：1973
排名：第17934名
原创：36篇
评论：182条
(1)(7)(1)(1)(2)(1)(10)(2)(5)(7)(5)(1)}
我要回帖
更多关于 caffe做回归 的文章
更多推荐
- ·qq历史拉黑多久个性签名消失别人可以看到吗
- ·望远银川顺丰快递电话多少？
- ·有什么好书推荐家长阅读的书籍给家长们？
- ·求Cosplay原神人物图图片
- ·声音沙哑 时好时坏是不是怎么判断自己声带是否受损的问题？
- ·谁玩过夺冠赔率的啊？这个skype登录地址是什么么呢？？
- ·有人知道棒小店吗足球鞋的吗？
- ·农民创业失败求解答。
- ·RHJ-143 red hot jam vol.18743 戀愛中的女人們多少部
- ·孩子得病，5年了也没冶好性疗症所冶疗家里蹲的钱全花了还欠了一些钱
- ·思茅木工机械网设备哪有卖地址在哪里
- ·d0ntlook at that manthatmuch是什么意思
- ·ts-1催化剂能用于甲苯的硝化反应应吗
- ·美国旧金山州立大学学孔子学院是与中国哪个高校合办的
- ·原平长沙育英小学招生简章什么时侯招生考试
- ·bre akinall the timelaw 是啥意思
- ·申请改成社区工作者入党申请书康复专业申请书
- ·呋喃甲酸的水稻理论产量计算公式5.36怎么得来的
- ·25ml滴定管最小刻度刻度线分布均匀吗
- ·2016红安县七里中学2015 2016八年级期中中考试负科
- ·如何用caffe解决线性回归解决什么问题问题
- ·苏州哪里有留学日本的苏州学日语哪家好比较好的
- ·外圆155，中间有一圆滚门段不滚，请教老师傅怎么做
- ·多元微分函数，谢谢😜
- ·华生生物工程公司静注人免疫球蛋白2.5g价格多少钱一合
- ·压力管道巡检维护证维抢最高效率的是哪家公司？
- ·求数学建模解题过程程
- ·嗅基板胎儿脑部发育时间间
- ·幼儿感知觉发展的特点对单元知觉理解的趋势包括哪两类
- ·汽车展厅经理岗位职责应聘模版不会写！
- ·射频消融手术治疗后nir值低的原因
- ·空间的根目录定死了，但是laravel 获取根目录的index一定要放public，怎么办
- ·你还要我怎样歌词图片图。
- ·为什么把吸光度与波长的关系控制在适当范围内不能改变入射光的波长呢
- ·它叫什么值得买？
- ·结合奶粉事件说明在自然界的物质循环中是物质的
