VoxNet对比实验
测试准确率92%
1h53min
Exp1数据集:levelsize=20 ringsize=99
46800 times making data
47000 times making data
47200 times making data
47400 times making data
47600 times making data
47800 times making data
congratuations!! We succeed!!
the total time is : 1.38630362113
004.desk_000001046.004
(20, 99, 3)
实验结果
准确率在20%左右 不理想 而且loss没有变过 不知道为什么 可能是数据的问题
换种数据组织形式
Exp2数据集 levelsize=15 ringsize=99
46600 times making data
46800 times making data
47000 times making data
47200 times making data
47400 times making data
47600 times making data
47800 times making data
congratuations!! We succeed!!
the total time is : 1.34806551887 hours
(15, 99, 3)
修改网络结构发现出现了这个问题
ValueError: dimension mismatch in x,y_idx arguments
发现可能是卷积维度出了问题
测试文件用10个数据测了filtersize 1*3 stride 1*1的卷积
得到的尺寸:
(150, 32, 99, 1)
在dimension上增加了尺度 并不是想象中的那样 明天调一下应该ok
睡了晚安
input : batches channels1 rings coors
conv output : batches filters(new_features) (rings features)
dimshuffle output : batches 1 rings new_features
l_conv4=lasagne.layers.Conv2DLayer(
incoming=l_conv3,
num_filters=128,
filter_size=(1, 1),
stride=(1,1),
pad='valid',
name='conv4',
# nonlinearity=activations.leaky_relu_001
)
# 得到150,128,99,1
l_ds1=lasagne.layers.DimshuffleLayer(
incoming=l_conv4,
pattern=(0,3,2,1),
name='ds1'
)
# 得到150,1,99,128
l_pool1=lasagne.layers.MaxPool2DLayer(
incoming=l_ds1,
pool_size=(n_rings,1),
name='pool1'
)
# 得到150,1,128
调通了
Exp3 修改网络层的卷积和维度结构
然后网络结果。。。更烂了(烂的恐怖。。。。。基本就是随机猜了
epoch: 79, itr: 119159, loss: 2.3556740284, acc: 0.15625
epoch: 79, itr: 119223, loss: 2.35532474518, acc: 0.09375
epoch: 79, itr: 119287, loss: 2.35581302643, acc: 0.0625
epoch: 79, itr: 119351, loss: 2.35496091843, acc: 0.1875
epoch: 79, itr: 119415, loss: 2.35812425613, acc: 0.125
epoch: 79, itr: 119479, loss: 2.35301709175, acc: 0.15625
epoch: 79, itr: 119543, loss: 2.35525465012, acc: 0.03125
epoch: 79, itr: 119607, loss: 2.35557365417, acc: 0.09375
epoch: 79, itr: 119671, loss: 2.35606193542, acc: 0.34375
epoch: 79, itr: 119735, loss: 2.35537862778, acc: 0.09375
epoch: 79, itr: 119760, loss: 2.35481595993, acc: 0.125
training done
running time : 10454.2010741
这次调整的是jitter和levelsize
随之调整的是网络结构 取消了层与层之间的关系 单个提取了特征 应用最大池化选出特征最明显的那个
接下来取消jitter看看是不是网络对于抖动的适应性不强
Exp4 取消jitter调整网络架构
之后结果好了很多
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 3.05266571045, acc: 0.125
epoch: 0, itr: 128, loss: 3.03939914703, acc: 0.21875
epoch: 0, itr: 192, loss: 3.02359199524, acc: 0.15625
epoch: 0, itr: 256, loss: 3.01317501068, acc: 0.21875
epoch: 0, itr: 320, loss: 3.00327014923, acc: 0.28125
epoch: 0, itr: 384, loss: 2.99204850197, acc: 0.3125
epoch: 0, itr: 448, loss: 2.97815012932, acc: 0.21875
epoch: 0, itr: 512, loss: 2.97294998169, acc: 0.09375
epoch: 0, itr: 576, loss: 2.96256351471, acc: 0.34375
epoch: 0, itr: 640, loss: 2.95148897171, acc: 0.1875
epoch: 0, itr: 704, loss: 2.96142339706, acc: 0.1875
epoch: 0, itr: 768, loss: 2.94087266922, acc: 0.3125
epoch: 0, itr: 832, loss: 2.93827581406, acc: 0.21875
epoch: 0, itr: 896, loss: 2.92367386818, acc: 0.21875
epoch: 0, itr: 960, loss: 2.91369652748, acc: 0.28125
epoch: 0, itr: 1024, loss: 2.92283964157, acc: 0.125
epoch: 0, itr: 1088, loss: 2.90435504913, acc: 0.15625
epoch: 0, itr: 1152, loss: 2.90478110313, acc: 0.4375
epoch: 0, itr: 1216, loss: 2.90733265877, acc: 0.125
epoch: 0, itr: 1280, loss: 2.90865659714, acc: 0.25
end
epoch: 79, itr: 118903, loss: 0.921702980995, acc: 0.625
epoch: 79, itr: 118967, loss: 0.951644897461, acc: 0.5625
epoch: 79, itr: 119031, loss: 0.889996409416, acc: 0.75
epoch: 79, itr: 119095, loss: 0.910210728645, acc: 0.625
epoch: 79, itr: 119159, loss: 0.948296189308, acc: 0.53125
epoch: 79, itr: 119223, loss: 0.902329266071, acc: 0.75
epoch: 79, itr: 119287, loss: 0.964941382408, acc: 0.59375
epoch: 79, itr: 119351, loss: 0.925312876701, acc: 0.625
epoch: 79, itr: 119415, loss: 0.967159569263, acc: 0.65625
epoch: 79, itr: 119479, loss: 0.930582284927, acc: 0.65625
epoch: 79, itr: 119543, loss: 0.946759283543, acc: 0.84375
epoch: 79, itr: 119607, loss: 0.963017344475, acc: 0.6875
epoch: 79, itr: 119671, loss: 0.949140667915, acc: 0.71875
epoch: 79, itr: 119735, loss: 0.970282793045, acc: 0.5625
epoch: 79, itr: 119760, loss: 0.958107590675, acc: 0.53125
training done
running time : 10373.8875451
再接再厉 这是对你成果的一种鼓励
你可以把他调的更好的 明天见
8月30日
早上好呀 现在是早上7点32分
我还在啃着早饭 昨天吃多了 今天轻断食
然后趁早上比较闲 先跑一个
Exp5 把level间特征提取的卷基层的filtersize改成了3*1
看看增强level间的联系会不会使结果有所改善
好啦我吃饭去啦
结果又回到20%了
所以证明增加level间的特征提取并没啥用呀
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 3.04191064835, acc: 0.09375
epoch: 0, itr: 128, loss: 3.0387802124, acc: 0.1875
epoch: 0, itr: 192, loss: 3.03397274017, acc: 0.25
epoch: 0, itr: 256, loss: 3.0263967514, acc: 0.15625
epoch: 0, itr: 320, loss: 3.01532649994, acc: 0.1875
epoch: 0, itr: 384, loss: 3.0037651062, acc: 0.3125
epoch: 0, itr: 448, loss: 2.98882770538, acc: 0.21875
epoch: 0, itr: 512, loss: 2.97912979126, acc: 0.09375
epoch: 0, itr: 576, loss: 2.97171258926, acc: 0.34375
end
epoch: 79, itr: 119287, loss: 2.22515869141, acc: 0.125
epoch: 79, itr: 119351, loss: 2.20100402832, acc: 0.15625
epoch: 79, itr: 119415, loss: 2.21867132187, acc: 0.4375
epoch: 79, itr: 119479, loss: 2.21990919113, acc: 0.125
epoch: 79, itr: 119543, loss: 2.22997045517, acc: 0.25
epoch: 79, itr: 119607, loss: 2.21950268745, acc: 0.21875
epoch: 79, itr: 119671, loss: 2.22106313705, acc: 0.09375
epoch: 79, itr: 119735, loss: 2.23824357986, acc: 0.21875
epoch: 79, itr: 119760, loss: 2.24561977386, acc: 0.25
training done
running time : 2.91176632583 hours
(这一早上磨磨唧唧也没干啥)
Exp6 取消层内的conv层 直接用每层提取的环内的512个特征进行全连接
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 3.62573289871, acc: 0.125
epoch: 0, itr: 128, loss: 3.61146068573, acc: 0.21875
epoch: 0, itr: 192, loss: 3.59484577179, acc: 0.15625
epoch: 0, itr: 256, loss: 3.58192825317, acc: 0.21875
epoch: 0, itr: 320, loss: 3.57144451141, acc: 0.28125
epoch: 0, itr: 384, loss: 3.56091451645, acc: 0.3125
epoch: 0, itr: 448, loss: 3.54156970978, acc: 0.21875
epoch: 0, itr: 512, loss: 3.53366374969, acc: 0.09375
epoch: 0, itr: 576, loss: 3.52067279816, acc: 0.34375
epoch: 0, itr: 640, loss: 3.49756240845, acc: 0.1875
epoch: 0, itr: 704, loss: 3.50591802597, acc: 0.1875
end
epoch: 79, itr: 119287, loss: 0.587748169899, acc: 0.9375
epoch: 79, itr: 119351, loss: 0.582167387009, acc: 0.96875
epoch: 79, itr: 119415, loss: 0.592884719372, acc: 0.96875
epoch: 79, itr: 119479, loss: 0.566123723984, acc: 0.875
epoch: 79, itr: 119543, loss: 0.576935708523, acc: 0.9375
epoch: 79, itr: 119607, loss: 0.592402815819, acc: 0.75
epoch: 79, itr: 119671, loss: 0.571764111519, acc: 0.9375
epoch: 79, itr: 119735, loss: 0.57043492794, acc: 0.875
epoch: 79, itr: 119760, loss: 0.59311568737, acc: 0.8125
training done
running time : 3.26074210193 hours
我的天啊 效果出奇的好 基本平均准确度在80%左右了
可能层间整合特征不是很靠谱
还是需要一层一层的比对特征
但是随着全连接层单元数变多 网络速度慢了一倍
(不知道是不是因为我一边做数据集一边跑实验的原因,感觉应该不是,网络不应该走的是gpu嘛)
减少全连接的神经元个数试一下
Exp7 去掉了1024的fc层
start
Training...
epoch: 0, itr: 64, loss: 3.1758480072, acc: 0.125
epoch: 0, itr: 128, loss: 3.14145350456, acc: 0.21875
epoch: 0, itr: 192, loss: 3.05617046356, acc: 0.15625
epoch: 0, itr: 256, loss: 3.05490922928, acc: 0.21875
epoch: 0, itr: 320, loss: 3.05752849579, acc: 0.28125
epoch: 0, itr: 384, loss: 3.04178166389, acc: 0.3125
epoch: 0, itr: 448, loss: 3.03817391396, acc: 0.21875
epoch: 0, itr: 512, loss: 3.03374910355, acc: 0.09375
epoch: 0, itr: 576, loss: 3.04026508331, acc: 0.34375
epoch: 0, itr: 640, loss: 3.02300786972, acc: 0.1875
epoch: 0, itr: 704, loss: 3.05015778542, acc: 0.1875
epoch: 0, itr: 768, loss: 3.01740646362, acc: 0.3125
epoch: 0, itr: 832, loss: 3.02174949646, acc: 0.21875
end
epoch: 79, itr: 118903, loss: 0.625308871269, acc: 0.84375
epoch: 79, itr: 118967, loss: 0.678461551666, acc: 0.84375
epoch: 79, itr: 119031, loss: 0.631180644035, acc: 0.8125
epoch: 79, itr: 119095, loss: 0.649296939373, acc: 0.84375
epoch: 79, itr: 119159, loss: 0.661121845245, acc: 0.71875
epoch: 79, itr: 119223, loss: 0.642716765404, acc: 0.875
epoch: 79, itr: 119287, loss: 0.656843662262, acc: 0.875
epoch: 79, itr: 119351, loss: 0.661109030247, acc: 0.65625
epoch: 79, itr: 119415, loss: 0.698909878731, acc: 0.78125
epoch: 79, itr: 119479, loss: 0.65883654356, acc: 0.8125
epoch: 79, itr: 119543, loss: 0.70613193512, acc: 0.96875
epoch: 79, itr: 119607, loss: 0.676484048367, acc: 0.9375
epoch: 79, itr: 119671, loss: 0.668956637383, acc: 0.84375
epoch: 79, itr: 119735, loss: 0.658791065216, acc: 0.75
epoch: 79, itr: 119760, loss: 0.635891675949, acc: 0.8125
training done
running time : 3.18637537029 hours
Exp8 将z轴取消
start
Training...
epoch: 0, itr: 64, loss: 3.6291282177, acc: 0.15625
epoch: 0, itr: 128, loss: 3.62130594254, acc: 0.125
epoch: 0, itr: 192, loss: 3.61307263374, acc: 0.0625
epoch: 0, itr: 256, loss: 3.60816526413, acc: 0.21875
epoch: 0, itr: 320, loss: 3.59987401962, acc: 0.0625
epoch: 0, itr: 384, loss: 3.58944916725, acc: 0.09375
epoch: 0, itr: 448, loss: 3.58728194237, acc: 0.21875
epoch: 0, itr: 512, loss: 3.57982206345, acc: 0.21875
epoch: 0, itr: 576, loss: 3.57056283951, acc: 0.0625
epoch: 0, itr: 640, loss: 3.56605672836, acc: 0.21875
epoch: 0, itr: 704, loss: 3.55911779404, acc: 0.03125
epoch: 0, itr: 768, loss: 3.55160093307, acc: 0.09375
epoch: 0, itr: 832, loss: 3.53931260109, acc: 0.1875
epoch: 0, itr: 896, loss: 3.54640316963, acc: 0.0
epoch: 0, itr: 960, loss: 3.53443574905, acc: 0.1875
epoch: 0, itr: 1024, loss: 3.52865743637, acc: 0.125
epoch: 0, itr: 1088, loss: 3.52027368546, acc: 0.09375
epoch: 0, itr: 1152, loss: 3.51587772369, acc: 0.0
epoch: 0, itr: 1216, loss: 3.5110938549, acc: 0.125
epoch: 0, itr: 1280, loss: 3.50110864639, acc: 0.3125
epoch: 0, itr: 1344, loss: 3.49929952621, acc: 0.0625
epoch: 0, itr: 1408, loss: 3.49026918411, acc: 0.125
end
epoch: 79, itr: 118903, loss: 1.24063241482, acc: 0.5
epoch: 79, itr: 118967, loss: 1.28016114235, acc: 0.46875
epoch: 79, itr: 119031, loss: 1.23755991459, acc: 0.40625
epoch: 79, itr: 119095, loss: 1.28409016132, acc: 0.4375
epoch: 79, itr: 119159, loss: 1.27584254742, acc: 0.4375
epoch: 79, itr: 119223, loss: 1.26372897625, acc: 0.375
epoch: 79, itr: 119287, loss: 1.2584913969, acc: 0.4375
epoch: 79, itr: 119351, loss: 1.23385000229, acc: 0.5625
epoch: 79, itr: 119415, loss: 1.25886106491, acc: 0.71875
epoch: 79, itr: 119479, loss: 1.23001623154, acc: 0.5625
epoch: 79, itr: 119543, loss: 1.23659217358, acc: 0.59375
epoch: 79, itr: 119607, loss: 1.28406059742, acc: 0.4375
epoch: 79, itr: 119671, loss: 1.25786757469, acc: 0.53125
epoch: 79, itr: 119735, loss: 1.22539961338, acc: 0.46875
epoch: 79, itr: 119760, loss: 1.24615871906, acc: 0.34375
training done
running time : 3.22871028834 hours
实验结果变回40-50
我觉着 有可能是因为最大池化把相关序列信息给混乱了
那我修改一下池化的维度试一下
Exp 9 修改最大池化的作用维度
修改l_pool1
原:
l_ds1=lasagne.layers.DimshuffleLayer(
incoming=l_conv5,
pattern=(0,3,2,1),
name='ds1'
)
l_pool1=lasagne.layers.MaxPool2DLayer(
incoming=l_ds1,
pool_size=(n_rings,1),
name='pool1'
)
# -------------shape n*1*512------------------------
现:
l_ds1=lasagne.layers.DimshuffleLayer(
incoming=l_conv5,
pattern=(0,3,1,2),
name='ds1'
)
l_pool1=lasagne.layers.MaxPool2DLayer(
incoming=l_ds1,
pool_size=(512,1),
name='pool1'
)
# -------------shape n*1*nrings------------------------
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 2.98683166504, acc: 0.125
epoch: 0, itr: 128, loss: 2.96195411682, acc: 0.21875
epoch: 0, itr: 192, loss: 2.93373084068, acc: 0.15625
epoch: 0, itr: 256, loss: 2.91701793671, acc: 0.21875
epoch: 0, itr: 320, loss: 2.9011182785, acc: 0.28125
epoch: 0, itr: 384, loss: 2.8788895607, acc: 0.3125
epoch: 0, itr: 448, loss: 2.84255361557, acc: 0.21875
epoch: 0, itr: 512, loss: 2.82576656342, acc: 0.09375
epoch: 0, itr: 576, loss: 2.82654047012, acc: 0.34375
epoch: 0, itr: 640, loss: 2.80454111099, acc: 0.1875
epoch: 0, itr: 704, loss: 2.83079338074, acc: 0.1875
epoch: 0, itr: 768, loss: 2.7900967598, acc: 0.3125
epoch: 0, itr: 832, loss: 2.7781457901, acc: 0.21875
epoch: 0, itr: 896, loss: 2.69097137451, acc: 0.3125
epoch: 0, itr: 960, loss: 2.55469369888, acc: 0.4375
epoch: 0, itr: 1024, loss: 2.39997696877, acc: 0.5
epoch: 0, itr: 1088, loss: 2.20239520073, acc: 0.40625
epoch: 0, itr: 1152, loss: 2.06093478203, acc: 0.71875
epoch: 0, itr: 1216, loss: 1.8644349575, acc: 0.71875
epoch: 0, itr: 1280, loss: 1.89491319656, acc: 0.875
epoch: 0, itr: 1344, loss: 1.88246119022, acc: 0.59375
epoch: 0, itr: 1408, loss: 1.76130223274, acc: 0.65625
epoch: 0, itr: 1472, loss: 1.7661318779, acc: 0.5625
epoch: 0, itr: 1497, loss: 1.74149274826, acc: 0.625
epoch: 1, itr: 1561, loss: 1.68512618542, acc: 0.53125
epoch: 1, itr: 1625, loss: 1.66278815269, acc: 0.8125
epoch: 1, itr: 1689, loss: 1.5877494812, acc: 0.65625
epoch: 1, itr: 1753, loss: 1.63686072826, acc: 0.65625
end
epoch: 78, itr: 117406, loss: 0.204402521253, acc: 0.96875
epoch: 79, itr: 118327, loss: 0.200673252344, acc: 0.96875
epoch: 79, itr: 118391, loss: 0.210909336805, acc: 0.96875
epoch: 79, itr: 118455, loss: 0.202990919352, acc: 0.96875
epoch: 79, itr: 118519, loss: 0.203096389771, acc: 0.96875
epoch: 79, itr: 118583, loss: 0.202306494117, acc: 1.0
epoch: 79, itr: 118647, loss: 0.205301970243, acc: 1.0
epoch: 79, itr: 118711, loss: 0.205429598689, acc: 1.0
epoch: 79, itr: 118775, loss: 0.208796769381, acc: 1.0
epoch: 79, itr: 118839, loss: 0.201861843467, acc: 1.0
epoch: 79, itr: 118903, loss: 0.201220005751, acc: 0.96875
epoch: 79, itr: 118967, loss: 0.205331593752, acc: 1.0
epoch: 79, itr: 119031, loss: 0.204278469086, acc: 1.0
epoch: 79, itr: 119095, loss: 0.209433943033, acc: 1.0
epoch: 79, itr: 119159, loss: 0.210401251912, acc: 0.96875
epoch: 79, itr: 119223, loss: 0.207351177931, acc: 1.0
epoch: 79, itr: 119287, loss: 0.203929051757, acc: 1.0
epoch: 79, itr: 119351, loss: 0.200749695301, acc: 1.0
epoch: 79, itr: 119415, loss: 0.203493878245, acc: 1.0
epoch: 79, itr: 119479, loss: 0.202249914408, acc: 1.0
epoch: 79, itr: 119543, loss: 0.205056667328, acc: 0.96875
epoch: 79, itr: 119607, loss: 0.204173788428, acc: 0.96875
epoch: 79, itr: 119671, loss: 0.199122384191, acc: 1.0
epoch: 79, itr: 119735, loss: 0.198002070189, acc: 0.96875
epoch: 79, itr: 119760, loss: 0.205145686865, acc: 0.96875
training done
running time : 1.92804625644 hours
准确率一度达到95+,时间缩短了1个多小时
意外的收获
我决定加上jitter试一下
Exp 10 修改的网络 2D数据 加 jitter
效果竟然也很好 90+, 而且loss最后没降到一定程度,可以提高学习率
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 2.99788355827, acc: 0.1875
epoch: 0, itr: 128, loss: 2.98973274231, acc: 0.1875
epoch: 0, itr: 192, loss: 2.97995328903, acc: 0.25
epoch: 0, itr: 256, loss: 2.9726293087, acc: 0.15625
epoch: 0, itr: 320, loss: 2.96896266937, acc: 0.1875
epoch: 0, itr: 384, loss: 2.95817780495, acc: 0.21875
epoch: 0, itr: 448, loss: 2.95817995071, acc: 0.15625
epoch: 0, itr: 512, loss: 2.94537591934, acc: 0.21875
epoch: 0, itr: 576, loss: 2.95091867447, acc: 0.1875
epoch: 0, itr: 640, loss: 2.94266986847, acc: 0.21875
epoch: 0, itr: 704, loss: 2.93309593201, acc: 0.28125
epoch: 0, itr: 768, loss: 2.92532396317, acc: 0.15625
epoch: 0, itr: 832, loss: 2.92647719383, acc: 0.125
epoch: 0, itr: 896, loss: 2.92631554604, acc: 0.09375
epoch: 0, itr: 960, loss: 2.91295337677, acc: 0.1875
epoch: 0, itr: 1024, loss: 2.91445064545, acc: 0.28125
epoch: 0, itr: 1088, loss: 2.91036009789, acc: 0.1875
epoch: 0, itr: 1152, loss: 2.91043424606, acc: 0.125
epoch: 0, itr: 1216, loss: 2.90271663666, acc: 0.1875
epoch: 0, itr: 1280, loss: 2.90355014801, acc: 0.21875
epoch: 0, itr: 1344, loss: 2.90614938736, acc: 0.28125
epoch: 0, itr: 1408, loss: 2.89301490784, acc: 0.125
epoch: 0, itr: 1472, loss: 2.88181209564, acc: 0.15625
epoch: 0, itr: 1497, loss: 2.91098237038, acc: 0.0937
end
epoch: 79, itr: 118327, loss: 0.342211484909, acc: 0.9375
epoch: 79, itr: 118391, loss: 0.342197805643, acc: 1.0
epoch: 79, itr: 118455, loss: 0.325636744499, acc: 1.0
epoch: 79, itr: 118519, loss: 0.361558914185, acc: 0.9375
epoch: 79, itr: 118583, loss: 0.350892186165, acc: 1.0
epoch: 79, itr: 118647, loss: 0.346374809742, acc: 1.0
epoch: 79, itr: 118711, loss: 0.329656600952, acc: 0.96875
epoch: 79, itr: 118775, loss: 0.358119547367, acc: 0.84375
epoch: 79, itr: 118839, loss: 0.352772593498, acc: 0.96875
epoch: 79, itr: 118903, loss: 0.304788649082, acc: 0.96875
epoch: 79, itr: 118967, loss: 0.354173123837, acc: 0.9375
epoch: 79, itr: 119031, loss: 0.337402403355, acc: 0.9375
epoch: 79, itr: 119095, loss: 0.357786595821, acc: 1.0
epoch: 79, itr: 119159, loss: 0.352255821228, acc: 0.875
epoch: 79, itr: 119223, loss: 0.343419969082, acc: 0.96875
epoch: 79, itr: 119287, loss: 0.345054745674, acc: 0.90625
epoch: 79, itr: 119351, loss: 0.360839188099, acc: 0.96875
epoch: 79, itr: 119415, loss: 0.357647120953, acc: 1.0
epoch: 79, itr: 119479, loss: 0.354044675827, acc: 0.9375
epoch: 79, itr: 119543, loss: 0.346770763397, acc: 1.0
epoch: 79, itr: 119607, loss: 0.349215626717, acc: 0.96875
epoch: 79, itr: 119671, loss: 0.348800957203, acc: 0.90625
epoch: 79, itr: 119735, loss: 0.328034698963, acc: 0.9375
epoch: 79, itr: 119760, loss: 0.353965312243, acc: 0.96875
training done
running time : 1.93263997999 hours
提高学习率 重新训练
epoch: 79, itr: 118327, loss: 0.21359808743, acc: 0.96875
epoch: 79, itr: 118391, loss: 0.218373626471, acc: 0.96875
epoch: 79, itr: 118455, loss: 0.214376941323, acc: 0.96875
epoch: 79, itr: 118519, loss: 0.215929538012, acc: 0.96875
epoch: 79, itr: 118583, loss: 0.218056619167, acc: 1.0
epoch: 79, itr: 118647, loss: 0.220117151737, acc: 1.0
epoch: 79, itr: 118711, loss: 0.214310333133, acc: 1.0
epoch: 79, itr: 118775, loss: 0.219554305077, acc: 0.9375
epoch: 79, itr: 118839, loss: 0.21261459589, acc: 1.0
epoch: 79, itr: 118903, loss: 0.211480095983, acc: 0.96875
epoch: 79, itr: 118967, loss: 0.219197958708, acc: 1.0
epoch: 79, itr: 119031, loss: 0.217068359256, acc: 1.0
epoch: 79, itr: 119095, loss: 0.22190079093, acc: 1.0
epoch: 79, itr: 119159, loss: 0.218642815948, acc: 1.0
epoch: 79, itr: 119223, loss: 0.218100041151, acc: 1.0
epoch: 79, itr: 119287, loss: 0.211290866137, acc: 1.0
epoch: 79, itr: 119351, loss: 0.212612181902, acc: 1.0
epoch: 79, itr: 119415, loss: 0.213988184929, acc: 1.0
epoch: 79, itr: 119479, loss: 0.21365454793, acc: 1.0
epoch: 79, itr: 119543, loss: 0.219817101955, acc: 0.96875
epoch: 79, itr: 119607, loss: 0.214001744986, acc: 1.0
epoch: 79, itr: 119671, loss: 0.216216236353, acc: 1.0
epoch: 79, itr: 119735, loss: 0.211886689067, acc: 0.96875
epoch: 79, itr: 119760, loss: 0.225625216961, acc: 0.96875
training done
running time : 1.93658612384 hours
尝试跑了一下测试集
效果居然提升了!!!!!!!
Preparing data...
Compiling theano functions...
Testing...
normal acc = 0.992149001921, macro_recall = 0.983889798209
running time : 0.00640085392528 hours
Exp 11
修改的网络 1D数据 加 jitter
start
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 2.98787975311, acc: 0.125
epoch: 0, itr: 128, loss: 2.96497440338, acc: 0.21875
epoch: 0, itr: 192, loss: 2.93989515305, acc: 0.15625
epoch: 0, itr: 256, loss: 2.92350435257, acc: 0.21875
epoch: 0, itr: 320, loss: 2.9126701355, acc: 0.28125
epoch: 0, itr: 384, loss: 2.9028263092, acc: 0.3125
epoch: 0, itr: 448, loss: 2.88080263138, acc: 0.21875
epoch: 0, itr: 512, loss: 2.87811326981, acc: 0.09375
epoch: 0, itr: 576, loss: 2.87193989754, acc: 0.34375
epoch: 0, itr: 640, loss: 2.85681152344, acc: 0.1875
epoch: 0, itr: 704, loss: 2.87573051453, acc: 0.1875
epoch: 0, itr: 768, loss: 2.85074710846, acc: 0.3125
epoch: 0, itr: 832, loss: 2.86000537872, acc: 0.21875
epoch: 0, itr: 896, loss: 2.8264093399, acc: 0.21875
epoch: 0, itr: 960, loss: 2.82296299934, acc: 0.28125
epoch: 0, itr: 1024, loss: 2.83727169037, acc: 0.125
epoch: 0, itr: 1088, loss: 2.81954884529, acc: 0.15625
epoch: 0, itr: 1152, loss: 2.82219457626, acc: 0.4375
epoch: 0, itr: 1216, loss: 2.82230329514, acc: 0.125
epoch: 0, itr: 1280, loss: 2.83293223381, acc: 0.25
epoch: 0, itr: 1344, loss: 2.81033730507, acc: 0.21875
epoch: 0, itr: 1408, loss: 2.8149960041, acc: 0.09375
epoch: 0, itr: 1472, loss: 2.82147908211, acc: 0.21875
epoch: 0, itr: 1497, loss: 2.81680178642, acc: 0.25
end
epoch: 79, itr: 118327, loss: 1.05541241169, acc: 0.5
epoch: 79, itr: 118391, loss: 1.04835057259, acc: 0.71875
epoch: 79, itr: 118455, loss: 0.992372572422, acc: 0.75
epoch: 79, itr: 118519, loss: 1.00396847725, acc: 0.5625
epoch: 79, itr: 118583, loss: 1.01601028442, acc: 0.65625
epoch: 79, itr: 118647, loss: 1.06033754349, acc: 0.71875
epoch: 79, itr: 118711, loss: 1.01991963387, acc: 0.59375
epoch: 79, itr: 118775, loss: 1.00438499451, acc: 0.5625
epoch: 79, itr: 118839, loss: 1.06097698212, acc: 0.71875
epoch: 79, itr: 118903, loss: 0.989911258221, acc: 0.5625
epoch: 79, itr: 118967, loss: 1.04876041412, acc: 0.6875
epoch: 79, itr: 119031, loss: 1.00237321854, acc: 0.6875
epoch: 79, itr: 119095, loss: 1.07973444462, acc: 0.59375
epoch: 79, itr: 119159, loss: 0.971791327, acc: 0.625
epoch: 79, itr: 119223, loss: 0.995470404625, acc: 0.625
epoch: 79, itr: 119287, loss: 1.03338265419, acc: 0.5625
epoch: 79, itr: 119351, loss: 1.0168094635, acc: 0.78125
epoch: 79, itr: 119415, loss: 1.02787959576, acc: 0.875
epoch: 79, itr: 119479, loss: 1.00060105324, acc: 0.53125
epoch: 79, itr: 119543, loss: 1.02305924892, acc: 0.53125
epoch: 79, itr: 119607, loss: 1.00871431828, acc: 0.59375
epoch: 79, itr: 119671, loss: 1.03131306171, acc: 0.65625
epoch: 79, itr: 119735, loss: 1.03671050072, acc: 0.59375
epoch: 79, itr: 119760, loss: 1.06972622871, acc: 0.625
training done
running time : 1.94760297802 hours
准确率60%
loss很高 提高学习率
觉着卷积对单维度的信息组织用途是不是不大
去掉两层卷积看看效果
Exp 12
1D数据 提高学习率 减掉2层卷积层
start
Training...
epoch: 0, itr: 64, loss: 2.85846734047, acc: 0.125
epoch: 0, itr: 128, loss: 2.83521747589, acc: 0.21875
epoch: 0, itr: 192, loss: 2.8096704483, acc: 0.15625
epoch: 0, itr: 256, loss: 2.79560756683, acc: 0.21875
epoch: 0, itr: 320, loss: 2.78388142586, acc: 0.28125
epoch: 0, itr: 384, loss: 2.76887202263, acc: 0.3125
epoch: 0, itr: 448, loss: 2.75193929672, acc: 0.21875
epoch: 0, itr: 512, loss: 2.75037741661, acc: 0.09375
epoch: 0, itr: 576, loss: 2.73654174805, acc: 0.34375
epoch: 0, itr: 640, loss: 2.72784996033, acc: 0.1875
epoch: 0, itr: 704, loss: 2.7457985878, acc: 0.1875
epoch: 0, itr: 768, loss: 2.71641659737, acc: 0.3125
epoch: 0, itr: 832, loss: 2.70596504211, acc: 0.21875
epoch: 0, itr: 896, loss: 2.69173622131, acc: 0.21875
epoch: 0, itr: 960, loss: 2.65669417381, acc: 0.3125
epoch: 0, itr: 1024, loss: 2.62912511826, acc: 0.375
epoch: 0, itr: 1088, loss: 2.50221252441, acc: 0.21875
epoch: 0, itr: 1152, loss: 2.46987724304, acc: 0.34375
epoch: 0, itr: 1216, loss: 2.39505004883, acc: 0.3125
epoch: 0, itr: 1280, loss: 2.3953461647, acc: 0.59375
epoch: 0, itr: 1344, loss: 2.35006141663, acc: 0.34375
epoch: 0, itr: 1408, loss: 2.32525873184, acc: 0.5625
epoch: 0, itr: 1472, loss: 2.31582927704, acc: 0.28125
epoch: 0, itr: 1497, loss: 2.29629206657, acc: 0.21875
end
epoch: 79, itr: 118327, loss: 0.520745158195, acc: 0.8125
epoch: 79, itr: 118391, loss: 0.503651857376, acc: 0.90625
epoch: 79, itr: 118455, loss: 0.47980183363, acc: 0.9375
epoch: 79, itr: 118519, loss: 0.509308218956, acc: 0.78125
epoch: 79, itr: 118583, loss: 0.525493919849, acc: 0.84375
epoch: 79, itr: 118647, loss: 0.518005371094, acc: 0.90625
epoch: 79, itr: 118711, loss: 0.50999045372, acc: 0.90625
epoch: 79, itr: 118775, loss: 0.508350074291, acc: 0.78125
epoch: 79, itr: 118839, loss: 0.535812020302, acc: 0.875
epoch: 79, itr: 118903, loss: 0.474978208542, acc: 0.90625
epoch: 79, itr: 118967, loss: 0.561199724674, acc: 0.875
epoch: 79, itr: 119031, loss: 0.512432396412, acc: 0.90625
epoch: 79, itr: 119095, loss: 0.538473844528, acc: 0.875
epoch: 79, itr: 119159, loss: 0.516969919205, acc: 0.71875
epoch: 79, itr: 119223, loss: 0.509480118752, acc: 0.875
epoch: 79, itr: 119287, loss: 0.513054728508, acc: 0.75
epoch: 79, itr: 119351, loss: 0.51029086113, acc: 0.875
epoch: 79, itr: 119415, loss: 0.554218769073, acc: 0.90625
epoch: 79, itr: 119479, loss: 0.540511131287, acc: 0.75
epoch: 79, itr: 119543, loss: 0.541306912899, acc: 0.90625
epoch: 79, itr: 119607, loss: 0.533096909523, acc: 0.875
epoch: 79, itr: 119671, loss: 0.537866950035, acc: 0.78125
epoch: 79, itr: 119735, loss: 0.557281136513, acc: 0.8125
epoch: 79, itr: 119760, loss: 0.601535916328, acc: 0.9375
training done
running time : 1.66067578971 hours
速度有提升 准确率不如2D loss下降难
如果取消所有的卷积网络。。。
Exp 13
1D数据 取消所有的卷积网络
Compiling theano functions...
Training...
epoch: 0, itr: 64, loss: 2.53098130226, acc: 0.125
epoch: 0, itr: 128, loss: 2.50941896439, acc: 0.21875
epoch: 0, itr: 192, loss: 2.4844994545, acc: 0.15625
epoch: 0, itr: 256, loss: 2.46341347694, acc: 0.21875
epoch: 0, itr: 320, loss: 2.44998311996, acc: 0.3125
epoch: 0, itr: 384, loss: 2.44373130798, acc: 0.40625
epoch: 0, itr: 448, loss: 2.42241668701, acc: 0.25
epoch: 0, itr: 512, loss: 2.41331577301, acc: 0.28125
epoch: 0, itr: 576, loss: 2.41891860962, acc: 0.40625
epoch: 0, itr: 640, loss: 2.38249778748, acc: 0.28125
epoch: 0, itr: 704, loss: 2.39074230194, acc: 0.40625
epoch: 0, itr: 768, loss: 2.33416771889, acc: 0.4375
epoch: 0, itr: 832, loss: 2.17741990089, acc: 0.375
epoch: 0, itr: 896, loss: 2.11201453209, acc: 0.375
epoch: 0, itr: 960, loss: 2.05201482773, acc: 0.5625
epoch: 0, itr: 1024, loss: 2.0522351265, acc: 0.59375
epoch: 0, itr: 1088, loss: 1.98330426216, acc: 0.4375
epoch: 0, itr: 1152, loss: 1.99772393703, acc: 0.3125
epoch: 0, itr: 1216, loss: 1.92867612839, acc: 0.46875
epoch: 0, itr: 1280, loss: 1.95680677891, acc: 0.6875
epoch: 0, itr: 1344, loss: 1.91538500786, acc: 0.46875
epoch: 0, itr: 1408, loss: 1.84957695007, acc: 0.59375
epoch: 0, itr: 1472, loss: 1.87200760841, acc: 0.46875
epoch: 0, itr: 1497, loss: 1.80851209164, acc: 0.34375
end
epoch: 79, itr: 118327, loss: 0.6273188591, acc: 0.78125
epoch: 79, itr: 118391, loss: 0.633217513561, acc: 0.84375
epoch: 79, itr: 118455, loss: 0.592578291893, acc: 0.90625
epoch: 79, itr: 118519, loss: 0.627317011356, acc: 0.84375
epoch: 79, itr: 118583, loss: 0.639250040054, acc: 0.96875
epoch: 79, itr: 118647, loss: 0.643329143524, acc: 0.875
epoch: 79, itr: 118711, loss: 0.629482984543, acc: 0.90625
epoch: 79, itr: 118775, loss: 0.644542336464, acc: 0.65625
epoch: 79, itr: 118839, loss: 0.668525576591, acc: 0.84375
epoch: 79, itr: 118903, loss: 0.598425626755, acc: 0.90625
epoch: 79, itr: 118967, loss: 0.654510140419, acc: 0.8125
epoch: 79, itr: 119031, loss: 0.611798763275, acc: 0.84375
epoch: 79, itr: 119095, loss: 0.626232624054, acc: 0.875
epoch: 79, itr: 119159, loss: 0.650104045868, acc: 0.71875
epoch: 79, itr: 119223, loss: 0.606007575989, acc: 0.84375
epoch: 79, itr: 119287, loss: 0.692492365837, acc: 0.8125
epoch: 79, itr: 119351, loss: 0.636899471283, acc: 0.71875
epoch: 79, itr: 119415, loss: 0.665019929409, acc: 0.8125
epoch: 79, itr: 119479, loss: 0.629163861275, acc: 0.84375
epoch: 79, itr: 119543, loss: 0.690085887909, acc: 0.96875
epoch: 79, itr: 119607, loss: 0.702319681644, acc: 0.9375
epoch: 79, itr: 119671, loss: 0.638007462025, acc: 0.84375
epoch: 79, itr: 119735, loss: 0.620165586472, acc: 0.8125
epoch: 79, itr: 119760, loss: 0.60178232193, acc: 0.84375
training done
running time : 0.328160045875 hours