paddle作业2-4
题1：
#作业2-4
import paddle
from paddle.nn import Conv2D, Linear, MaxPool2D
import numpy as np
import paddle.nn.functional as F
class Net(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.conv1 = Conv2D(in_channels=1, out_channels=3, kernel_size=3)
self.conv2 = Conv2D(in_channels=3, out_channels=10, kernel_size=3)
self.maxpool = MaxPool2D(kernel_size=2, stride=2)
# self.fc = Linear(in_features=)
def forward(self, x):
x = self.conv1(x)
print("conv1--shape:{}".format(x.shape))
print("content:",x)
x = self.maxpool(x)
print("conv1-maxpool--shape:{}".format(x.shape))
print("content:",x)
x = self.conv2(x)
print("conv2--shape:{}".format(x.shape))
print("content:",x)
x = self.maxpool(x)
print("conv2-maxpool--shape:{}".format(x.shape))
print("content:",x)
return x
u = np.random.rand(1,1,64,64)
u = paddle.to_tensor(u).astype('float32')
net = Net()
net(u)
输出：
conv1--shape:[1, 3, 62, 62]
content: Tensor(shape=[1, 3, 62, 62], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
[[[[-1.03177869, -0.77748996, -0.37205794, ..., -1.16914952, -1.03177857, -0.58381915],
[-0.67789686, -1.58549762, -0.97597510, ..., -0.56963080, -1.82829785, -1.42108345],
[ 0.25646657, -0.19316070, -0.06493981, ..., -1.43284070, -1.58845973, -1.06161082],
...,
[-1.06657410, -0.12878723, -0.89841843, ..., 0.47988266, -0.54047906, -1.42637694],
[-0.84822047, 0.06302693, -0.20786586, ..., -0.75426561, -0.53132033, -0.93996078],
[-1.38931227, 0.21905223, -0.71452707, ..., -1.00175893, -1.16727090, -1.04180014]],
[[-0.34005663, 1.01721966, 0.46573788, ..., 0.36177707, 0.49334481, 0.93489450],
[-0.16840670, -0.19559138, -0.35780585, ..., -0.10786653, -0.10211647, 0.27730283],
[ 0.47618026, -0.57042152, 0.33241624, ..., -0.68744487, -0.44638035, 0.59348673],
...,
[ 0.43587872, 0.37292024, -0.55591142, ..., 0.35407147, -0.51093435, 0.15134723],
[ 0.77757734, -0.19600271, -0.73693144, ..., 0.06336728, -0.21787606, 0.04249654],
[ 0.45646501, 0.97877854, -0.10154613, ..., -0.20721525, 0.36874118, -0.34399495]],
[[ 0.03909174, -0.24251559, 0.32627419, ..., -0.35300124, 0.05717748, -0.15156426],
[-0.75969607, -0.72273302, -0.21693535, ..., 0.16040878, -0.60953814, -0.76194024],
[ 0.24897744, 0.20954224, -0.35446906, ..., -0.53350484, -0.65325189, -0.71377838],
...,
[-0.64520627, -0.13371752, -0.00226474, ..., 0.43722275, 0.04000597, -0.68827093],
[-0.47013941, 0.06369685, 0.07199863, ..., 0.15542001, 0.02171265, -0.53287303],
[-0.38482872, 0.12792701, 0.20928183, ..., -0.41419211, -0.34249121, -0.37504444]]]])
conv1-maxpool--shape:[1, 3, 31, 31]
content: Tensor(shape=[1, 3, 31, 31], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
[[[[-0.67789686, -0.37205794, -0.62262011, ..., 0.18379803, 0.14368679, -0.58381915],
[ 0.25646657, -0.06493981, -0.66757733, ..., 0.08953179, 0.10545183, -0.91407675],
[ 0.16089660, 0.06359247, -0.30854785, ..., 0.04017820, -0.31908986, -0.17216061],
...,
[-0.38638410, 0.43577391, -0.52118969, ..., -0.50289881, -0.37256008, -0.08074930],
[-0.12878723, -0.89841843, -0.11317723, ..., -0.04990890, 0.47988266, -0.40975910],
[ 0.21905223, -0.11454697, -0.56476015, ..., -0.02640892, -0.06468204, -0.53132033]],
[[ 1.01721966, 0.50235915, 0.48203111, ..., 1.17043245, 0.36177707, 0.93489450],
[ 0.47618026, 0.49586532, 0.79181617, ..., -0.04261628, 0.30811083, 0.59348673],
[ 0.99659878, 0.19138895, 0.81142896, ..., 0.85095757, 0.70458364, 0.81303638],
...,
[ 0.68253225, 0.43649405, 0.30807868, ..., 0.93992198, 0.08084880, 0.42516100],
[ 0.43587872, 0.00076535, 0.90764409, ..., 0.52016580, 0.75320625, 0.33648261],
[ 0.97877854, 1.03922081, 0.63778490, ..., 0.31145221, 0.45375410, 0.36874118]],
[[ 0.03909174, 0.32627419, -0.11361890, ..., 0.28261650, 0.16040878, 0.05717748],
[ 0.24897744, 0.18517545, -0.32565966, ..., -0.11045824, 0.06884519, -0.30248123],
[ 0.24063696, 0.01648869, -0.21375470, ..., 0.18693715, -0.00632723, -0.00128613],
...,
[-0.14140783, 0.34060207, -0.04902748, ..., -0.13388914, -0.24003626, -0.10710047],
[-0.13371752, -0.00226474, 0.00684804, ..., 0.11446057, 0.43722275, 0.39217085],
[ 0.12792701, 0.20928183, 0.22429550, ..., 0.40111300, 0.15542001, 0.02171265]]]])
conv2--shape:[1, 10, 29, 29]
content: Tensor(shape=[1, 10, 29, 29], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
[[[[ 0.44623196, 0.62388515, 0.17902687, ..., 0.72025532, 0.53890043, 0.24215084],
[ 1.18032026, 0.27674440, -0.19034995, ..., 0.23119810, -0.05347760, 1.21775162],
[ 0.63521343, 0.55254978, -0.01318052, ..., 0.48960438, 0.47024673, 0.33397472],
...,
[ 0.30813912, 0.09670782, 0.40232271, ..., -0.10699759, 0.89743155, -0.19544327],
[-0.10760272, 1.14544046, 0.04281932, ..., 0.48013151, 0.30368382, 0.67060882],
[ 0.52567440, 0.11980435, 0.17403756, ..., 0.73117226, -0.06727744, -0.12170704]],
[[ 1.21443903, 0.51579273, 0.85561645, ..., 0.78079736, 0.41128790, 0.91624510],
[ 0.65348673, 0.45180458, 0.52336526, ..., 0.11084333, 0.14484397, -0.34574771],
[ 0.59918010, 0.56237960, 0.84820282, ..., 0.83417332, 0.47399926, 0.71474850],
...,
[ 0.32853597, -0.23827338, 0.12160541, ..., 0.79289854, 0.47846511, 0.79586089],
[ 1.23849845, 0.18347394, 0.56944412, ..., 0.58740354, -0.08288956, 0.29207882],
[ 0.07357436, 1.00488615, 0.61807150, ..., 0.32044291, 0.12979138, 0.83888710]],
[[ 0.61238801, -0.20202814, 0.38081983, ..., 0.34322435, 0.65360999, 0.17780209],
[ 0.08400173, 0.23523483, 0.57989526, ..., 0.26658696, 0.19919813, -0.53452939],
[-0.00422187, 0.05944386, 0.65046120, ..., -0.08986238, 0.32434759, 0.11998253],
...,
[ 0.07090414, 0.42234129, 0.06542711, ..., 0.66841406, -0.26657850, 0.48105952],
[ 0.62220156, -0.12735876, 0.69410717, ..., 0.30230391, -0.06783611, -0.17921242],
[-0.10617691, 0.22948569, 0.47753474, ..., 0.14906025, 0.73869264, 0.67903733]],
...,
[[ 0.58693612, -0.18672022, 0.63746285, ..., 0.46479458, 0.50665700, -0.19183946],
[ 0.37139094, 0.53905475, 1.19847333, ..., 0.34966975, 0.17246890, 0.32180622],
[-0.29460990, -0.02060680, 0.92494738, ..., 0.27885771, 1.14765000, -0.23069966],
...,
[ 0.89485133, -0.09401222, 0.00301947, ..., 0.83435619, -0.20381044, 0.80572271],
[ 0.27982765, 0.57542384, 0.99463135, ..., -0.08496559, -0.18287888, 0.28955793],
[ 0.57256705, 0.05427426, 0.51110739, ..., 0.06475183, 0.79297543, 0.36223978]],
[[ 0.47162932, 0.20084937, 0.74264425, ..., 0.49598989, 0.40802330, 0.03926894],
[ 1.03775215, 0.64674914, 1.08719802, ..., 0.63512015, 0.14761853, 0.79650056],
[ 0.30375546, 0.38327909, 0.88709396, ..., 0.68090713, 0.49056888, 0.72294033],
...,
[ 0.78672379, 0.37813893, 0.51430082, ..., 0.42658037, 0.96949989, 0.06683272],
[ 0.53548479, 0.70059055, 0.54305643, ..., 0.12597957, -0.13162267, 1.08301067],
[ 0.75669861, 0.78659368, 0.64943862, ..., 0.39239758, 0.80140257, 1.17412639]],
[[-0.03030671, -0.23092379, -0.82495379, ..., -0.02047309, -0.13223293, -0.25102410],
[-0.98404789, -0.19093001, -0.36319470, ..., -0.16261804, -0.11625568, -0.68067783],
[-0.27763790, -0.16244254, -0.20049295, ..., -0.49187917, -0.36560684, -0.45228794],
...,
[-0.55032504, -0.06628364, -0.26941240, ..., 0.05824207, -0.91152292, -0.05734545],
[-0.33500731, -0.82734483, -0.01479590, ..., -0.38682261, 0.02892387, -0.38761553],
[-0.47419488, -0.25726256, -0.54496294, ..., -0.78106833, 0.23607337, 0.28471929]]]])
conv2-maxpool--shape:[1, 10, 14, 14]
content: Tensor(shape=[1, 10, 14, 14], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
[[[[ 1.18032026, 0.73366624, 0.78654218, ..., 0.83051729, 0.44774604, 0.72025532],
[ 0.82915610, 0.91865158, 0.95083821, ..., 1.02050555, 0.44698304, 0.80424589],
[ 1.14481282, 0.55444986, 0.57537168, ..., 0.38508829, 0.71621978, 0.53432995],
...,
[ 0.69975263, 0.92470175, 0.61134225, ..., 1.01343369, 0.78940690, 1.37072110],
[ 0.65116763, 1.21608710, 0.53525656, ..., 0.48411304, 0.85110980, 0.77779889],
[ 1.14544046, 0.91810060, 0.52337706, ..., 0.51168656, 0.81643701, 0.89743155]],
[[ 1.21443903, 0.94713080, 0.84755552, ..., 1.19978046, 0.74996358, 0.78079736],
[ 0.59918010, 0.84820282, 1.15260458, ..., 1.61056054, 0.88557887, 0.83417332],
[ 1.17461693, 0.91230959, 1.09217799, ..., 1.19446337, 0.88778430, 0.61754584],
...,
[ 1.21085870, 0.59851116, 0.84720647, ..., 1.55403900, 1.21921003, 1.58946621],
[ 0.85086536, 0.99615830, 1.03322124, ..., 0.82606924, 1.39627910, 0.81494212],
[ 1.23849845, 1.25585008, 0.69309294, ..., 0.99248862, 0.66527927, 0.79289854]],
[[ 0.61238801, 0.57989526, 0.77676761, ..., 0.21774773, 0.42530563, 0.65360999],
[ 0.41089585, 1.02843177, 0.72878689, ..., 0.23436627, 0.33608097, 0.32434759],
[ 1.09274590, 0.87833339, 0.66920429, ..., 0.35917246, 0.26506528, 0.29494682],
...,
[ 0.55407143, 0.56813610, 0.46117276, ..., -0.00291205, -0.13898167, 0.32209030],
[ 0.33941227, 0.59389615, 0.72072250, ..., 0.56878316, 0.49741602, 0.90810311],
[ 0.62220156, 0.69410717, 0.23148048, ..., 0.76910746, 0.72130215, 0.66841406]],
...,
[[ 0.58693612, 1.19847333, 0.66977620, ..., 0.89528000, 0.82926500, 0.50665700],
[ 1.01370800, 0.92494738, 0.78183734, ..., 0.97884059, 0.56055748, 1.14765000],
[ 0.95192552, 1.06201458, 0.87643880, ..., 1.43337226, 0.59050602, 0.69752371],
...,
[ 0.78550702, 0.82474697, 1.28047204, ..., 0.72682345, 1.14773560, 1.22152841],
[ 0.73417985, 1.02442372, 0.58345133, ..., 0.72091520, 0.85812181, 1.01788020],
[ 0.89485133, 0.99463135, 0.73533934, ..., 0.71784866, 1.06157923, 0.83435619]],
[[ 1.03775215, 1.16232347, 0.72688782, ..., 1.08088934, 0.96468461, 0.63512015],
[ 1.23473525, 1.06870127, 1.49160647, ..., 1.21071362, 0.75042331, 1.00574136],
[ 1.49424028, 1.16570246, 1.26114154, ..., 0.70151198, 0.76903617, 0.57344830],
...,
[ 0.81895471, 1.30834210, 1.10446966, ..., 1.15589213, 1.29913092, 1.29558706],
[ 0.79784775, 0.96286309, 1.38311434, ..., 1.26586366, 1.32836044, 1.12565398],
[ 0.78672379, 1.02574015, 1.07679272, ..., 0.91598916, 1.00484693, 0.96949989]],
[[-0.03030671, -0.18821266, 0.00659555, ..., -0.54614818, -0.26359645, -0.02047309],
[-0.16244254, -0.20049295, 0.21887742, ..., -0.30385906, -0.17079493, -0.07464117],
[-0.01447295, 0.28109685, -0.50979286, ..., -0.33690268, 0.13603975, 0.45781386],
...,
[ 0.31133324, -0.06370067, -0.25026637, ..., -0.18924497, -0.19423440, 0.27997458],
[ 0.58236080, 0.08960927, 0.00888072, ..., 0.00078234, 0.02889496, -0.00263232],
[-0.06628364, -0.01479590, -0.14334446, ..., 0.31529629, 0.36629245, 0.05824207]]]])
Tensor(shape=[1, 10, 14, 14], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
[[[[ 1.18032026, 0.73366624, 0.78654218, ..., 0.83051729, 0.44774604, 0.72025532],
[ 0.82915610, 0.91865158, 0.95083821, ..., 1.02050555, 0.44698304, 0.80424589],
[ 1.14481282, 0.55444986, 0.57537168, ..., 0.38508829, 0.71621978, 0.53432995],
...,
[ 0.69975263, 0.92470175, 0.61134225, ..., 1.01343369, 0.78940690, 1.37072110],
[ 0.65116763, 1.21608710, 0.53525656, ..., 0.48411304, 0.85110980, 0.77779889],
[ 1.14544046, 0.91810060, 0.52337706, ..., 0.51168656, 0.81643701, 0.89743155]],
[[ 1.21443903, 0.94713080, 0.84755552, ..., 1.19978046, 0.74996358, 0.78079736],
[ 0.59918010, 0.84820282, 1.15260458, ..., 1.61056054, 0.88557887, 0.83417332],
[ 1.17461693, 0.91230959, 1.09217799, ..., 1.19446337, 0.88778430, 0.61754584],
...,
[ 1.21085870, 0.59851116, 0.84720647, ..., 1.55403900, 1.21921003, 1.58946621],
[ 0.85086536, 0.99615830, 1.03322124, ..., 0.82606924, 1.39627910, 0.81494212],
[ 1.23849845, 1.25585008, 0.69309294, ..., 0.99248862, 0.66527927, 0.79289854]],
[[ 0.61238801, 0.57989526, 0.77676761, ..., 0.21774773, 0.42530563, 0.65360999],
[ 0.41089585, 1.02843177, 0.72878689, ..., 0.23436627, 0.33608097, 0.32434759],
[ 1.09274590, 0.87833339, 0.66920429, ..., 0.35917246, 0.26506528, 0.29494682],
...,
[ 0.55407143, 0.56813610, 0.46117276, ..., -0.00291205, -0.13898167, 0.32209030],
[ 0.33941227, 0.59389615, 0.72072250, ..., 0.56878316, 0.49741602, 0.90810311],
[ 0.62220156, 0.69410717, 0.23148048, ..., 0.76910746, 0.72130215, 0.66841406]],
...,
[[ 0.58693612, 1.19847333, 0.66977620, ..., 0.89528000, 0.82926500, 0.50665700],
[ 1.01370800, 0.92494738, 0.78183734, ..., 0.97884059, 0.56055748, 1.14765000],
[ 0.95192552, 1.06201458, 0.87643880, ..., 1.43337226, 0.59050602, 0.69752371],
...,
[ 0.78550702, 0.82474697, 1.28047204, ..., 0.72682345, 1.14773560, 1.22152841],
[ 0.73417985, 1.02442372, 0.58345133, ..., 0.72091520, 0.85812181, 1.01788020],
[ 0.89485133, 0.99463135, 0.73533934, ..., 0.71784866, 1.06157923, 0.83435619]],
[[ 1.03775215, 1.16232347, 0.72688782, ..., 1.08088934, 0.96468461, 0.63512015],
[ 1.23473525, 1.06870127, 1.49160647, ..., 1.21071362, 0.75042331, 1.00574136],
[ 1.49424028, 1.16570246, 1.26114154, ..., 0.70151198, 0.76903617, 0.57344830],
...,
[ 0.81895471, 1.30834210, 1.10446966, ..., 1.15589213, 1.29913092, 1.29558706],
[ 0.79784775, 0.96286309, 1.38311434, ..., 1.26586366, 1.32836044, 1.12565398],
[ 0.78672379, 1.02574015, 1.07679272, ..., 0.91598916, 1.00484693, 0.96949989]],
[[-0.03030671, -0.18821266, 0.00659555, ..., -0.54614818, -0.26359645, -0.02047309],
[-0.16244254, -0.20049295, 0.21887742, ..., -0.30385906, -0.17079493, -0.07464117],
[-0.01447295, 0.28109685, -0.50979286, ..., -0.33690268, 0.13603975, 0.45781386],
...,
[ 0.31133324, -0.06370067, -0.25026637, ..., -0.18924497, -0.19423440, 0.27997458],
[ 0.58236080, 0.08960927, 0.00888072, ..., 0.00078234, 0.02889496, -0.00263232],
[-0.06628364, -0.01479590, -0.14334446, ..., 0.31529629, 0.36629245, 0.05824207]]]])题2：
import matplotlib.pyplot as plt
import paddle.nn.functional as F
from paddle.nn import Conv2D, MaxPool2D, Linear
import paddle
import numpy as np
class MNIST(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.conv1 = Conv2D(in_channels=1, out_channels=20, kernel_size=5, stride=1, padding=2)
self.maxpool = MaxPool2D(kernel_size=2, stride=2)
self.conv2 = Conv2D(in_channels=20, out_channels=20, kernel_size=5, stride=1, padding=2)
self.fc = Linear(in_features=980, out_features=10)
def forward(self, inputs, label=None, check_shape=False, check_content=False):
outputs1 = self.conv1(inputs)
outputs2 = F.relu(outputs1)
outputs3 = self.maxpool(outputs2)
outputs4 = self.conv2(outputs3)
outputs5 = F.relu(outputs4)
outputs6 = self.maxpool(outputs5)
outputs6 = paddle.reshape(outputs6, [outputs6.shape[0], 980])
outputs7 = self.fc(outputs6)
if check_shape:
print("\n######## print networks layer's superparams #########")
print("conv1-- kernel_size:{}, padding:{}, stride:{}".format(self.conv1.weight.shape, self.conv1._padding, self.conv1._stride))
print("conv2-- kernel_size:{}, padding:{}, stride:{}".format(self.conv2.weight.shape, self.conv2._padding, self.conv2._stride))
print("fc-- weight_size:{}, bias_size_{}".format(self.fc.weight.shape, self.fc.bias.shape))
#打印每层的输出尺寸
print("\n####### print shape of features of every layer ###########")
print("inputs_shape:{}".format(inputs.shape))
print("outputs1_shape:{}".format(outputs1.shape))
print("outputs2_shape:{}".format(outputs2.shape))
print("outputs3_shape:{}".format(outputs3.shape))
print("outputs4_shape:{}".format(outputs4.shape))
print("outputs5_shape:{}".format(outputs5.shape))
print("outputs6_shape:{}".format(outputs6.shape))
print("outputs7_shape:{}".format(outputs7.shape))
if check_content:
print("\n########print convolution layer's kernel ########")
print("conv1 params -- kernel weights:", self.conv1.weight[0][0])
print("conv2 params -- kernel weights:", self.conv2.weight[0][0])
#创建随机数，随机打印某一个通道的输出值
idx1 = np.random.randint(0, outputs1.shape[1])
idx2 = np.random.randint(0, outputs4.shape[1])
#打印卷积-池化后的结果，进打印batch中第一个图像对应的特征
print("\nThe {}th channel of conv1 layer:".format(idx1), outputs1[0][idx1])
print("The {}th channel of conv2 layer:".format(idx2),outputs4[0][idx2])
print("The output of last layer:", outputs7[0], '\n')
if label is not None:
acc = paddle.metric.accuracy(input=F.softmax(outputs7), label=label)
return outputs7, acc
else:
return outputs7
transform = paddle.vision.transforms.Compose([paddle.vision.ToTensor(),
paddle.vision.Normalize((0.1307,),(0.3081,))])
train_data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
train_loader = paddle.io.DataLoader(train_data, shuffle=True, batch_size=100)
def train(model):
model.train()
EPOCH_NUM = 3
opt = paddle.optimizer.Adam(learning_rate=0.01, weight_decay=paddle.regularizer.L2Decay(), parameters=model.parameters())
iter1 = 0
iters = []
losses = []
accs = []
for epoch_id in range(EPOCH_NUM):
for batch_id, data in enumerate(train_loader()):
imgs, labels = data
imgs = paddle.to_tensor(imgs)
labels = paddle.to_tensor(labels)
predicts, acc = model(imgs, labels)
loss = F.cross_entropy(predicts, labels)
avg_loss = paddle.mean(loss)
if batch_id % 100 == 0:
print("epoch:{}, batch:{}, loss:{}, acc:{}".format(epoch_id, batch_id, avg_loss.numpy(), acc.numpy()))
iters.append(iter1)
iter1 += 100
losses.append(avg_loss.numpy())
accs.append(acc.numpy())
avg_loss.backward()
opt.step()
opt.clear_grad()
paddle.save(model.state_dict(), "mnist.pdparams")
return iters, losses, accs
model = MNIST()
iters, losses, accs = train(model)
plt.figure()
plt.title("Accuracy")
plt.xlabel("iter",fontsize=14)
plt.ylabel("accu",fontsize=14)
plt.plot(iters,accs,label='train accu')
plt.legend()
plt.show()输出：
题3：对比两个损失函数（第一个损失函数为交叉熵函数，第二个损失函数为最小二乘法函数）
import matplotlib.pyplot as plt
import paddle.nn.functional as F
from paddle.nn import Conv2D, MaxPool2D, Linear
import paddle
import numpy as np
class MNIST(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.conv1 = Conv2D(in_channels=1, out_channels=20, kernel_size=5, stride=1, padding=2)
self.maxpool = MaxPool2D(kernel_size=2, stride=2)
self.conv2 = Conv2D(in_channels=20, out_channels=20, kernel_size=5, stride=1, padding=2)
self.fc = Linear(in_features=980, out_features=10)
def forward(self, inputs, label=None, check_shape=False, check_content=False):
outputs1 = self.conv1(inputs)
outputs2 = F.relu(outputs1)
outputs3 = self.maxpool(outputs2)
outputs4 = self.conv2(outputs3)
outputs5 = F.relu(outputs4)
outputs6 = self.maxpool(outputs5)
outputs6 = paddle.reshape(outputs6, [outputs6.shape[0], 980])
outputs7 = self.fc(outputs6)
if check_shape:
print("\n######## print networks layer's superparams #########")
print("conv1-- kernel_size:{}, padding:{}, stride:{}".format(self.conv1.weight.shape, self.conv1._padding, self.conv1._stride))
print("conv2-- kernel_size:{}, padding:{}, stride:{}".format(self.conv2.weight.shape, self.conv2._padding, self.conv2._stride))
print("fc-- weight_size:{}, bias_size_{}".format(self.fc.weight.shape, self.fc.bias.shape))
#打印每层的输出尺寸
print("\n####### print shape of features of every layer ###########")
print("inputs_shape:{}".format(inputs.shape))
print("outputs1_shape:{}".format(outputs1.shape))
print("outputs2_shape:{}".format(outputs2.shape))
print("outputs3_shape:{}".format(outputs3.shape))
print("outputs4_shape:{}".format(outputs4.shape))
print("outputs5_shape:{}".format(outputs5.shape))
print("outputs6_shape:{}".format(outputs6.shape))
print("outputs7_shape:{}".format(outputs7.shape))
if check_content:
print("\n########print convolution layer's kernel ########")
print("conv1 params -- kernel weights:", self.conv1.weight[0][0])
print("conv2 params -- kernel weights:", self.conv2.weight[0][0])
#创建随机数，随机打印某一个通道的输出值
idx1 = np.random.randint(0, outputs1.shape[1])
idx2 = np.random.randint(0, outputs4.shape[1])
#打印卷积-池化后的结果，进打印batch中第一个图像对应的特征
print("\nThe {}th channel of conv1 layer:".format(idx1), outputs1[0][idx1])
print("The {}th channel of conv2 layer:".format(idx2),outputs4[0][idx2])
print("The output of last layer:", outputs7[0], '\n')
if label is not None:
acc = paddle.metric.accuracy(input=F.softmax(outputs7), label=label)
return outputs7, acc
else:
return outputs7
transform = paddle.vision.transforms.Compose([paddle.vision.ToTensor(),
paddle.vision.Normalize((0.1307,),(0.3081,))])
train_data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
train_loader = paddle.io.DataLoader(train_data, shuffle=True, batch_size=100)
def train(model):
model.train()
EPOCH_NUM = 3
opt = paddle.optimizer.Adam(learning_rate=0.01, weight_decay=paddle.regularizer.L2Decay(), parameters=model.parameters())
iter1 = 0
iters = []
losses = []
accs = []
for epoch_id in range(EPOCH_NUM):
for batch_id, data in enumerate(train_loader()):
imgs, labels = data
imgs = paddle.to_tensor(imgs)
labels = paddle.to_tensor(labels)
predicts, acc = model(imgs, labels)
loss = F.cross_entropy(predicts, labels)
avg_loss = paddle.mean(loss)
if batch_id % 100 == 0:
print("epoch:{}, batch:{}, loss:{}, acc:{}".format(epoch_id, batch_id, avg_loss.numpy(), acc.numpy()))
iters.append(iter1)
iter1 += 100
losses.append(avg_loss.numpy())
accs.append(acc.numpy())
avg_loss.backward()
opt.step()
opt.clear_grad()
paddle.save(model.state_dict(), "mnist.pdparams")
return iters, losses, accs
model = MNIST()
iters, losses, accs = train(model)输出：
epoch:0, batch:0, loss:[4.154338], acc:[0.08]
epoch:0, batch:100, loss:[0.286233], acc:[0.9]
epoch:0, batch:200, loss:[0.22489373], acc:[0.94]
epoch:0, batch:300, loss:[0.19534074], acc:[0.94]
epoch:0, batch:400, loss:[0.17166412], acc:[0.93]
epoch:0, batch:500, loss:[0.23943755], acc:[0.91]
epoch:1, batch:0, loss:[0.10949133], acc:[0.96]
epoch:1, batch:100, loss:[0.2541501], acc:[0.92]
epoch:1, batch:200, loss:[0.06903129], acc:[0.98]
epoch:1, batch:300, loss:[0.10762461], acc:[0.96]
epoch:1, batch:400, loss:[0.03312943], acc:[0.99]
epoch:1, batch:500, loss:[0.08275294], acc:[0.96]
epoch:2, batch:0, loss:[0.06648406], acc:[0.99]
epoch:2, batch:100, loss:[0.04650707], acc:[0.97]
epoch:2, batch:200, loss:[0.08721491], acc:[0.97]
epoch:2, batch:300, loss:[0.05407384], acc:[0.99]
epoch:2, batch:400, loss:[0.10466648], acc:[0.95]
epoch:2, batch:500, loss:[0.028617], acc:[0.99]plt.figure()
plt.plot(iters, accs)
plt.show()输出：
import matplotlib.pyplot as plt
import paddle.nn.functional as F
from paddle.nn import Conv2D, MaxPool2D, Linear
import paddle
import numpy as np
class MNIST(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.conv1 = Conv2D(in_channels=1, out_channels=20, kernel_size=5, stride=1, padding=2)
self.maxpool = MaxPool2D(kernel_size=2, stride=2)
self.conv2 = Conv2D(in_channels=20, out_channels=20, kernel_size=5, stride=1, padding=2)
self.fc = Linear(in_features=980, out_features=10)
def forward(self, inputs, label=None, check_shape=False, check_content=False):
outputs1 = self.conv1(inputs)
outputs2 = F.relu(outputs1)
outputs3 = self.maxpool(outputs2)
outputs4 = self.conv2(outputs3)
outputs5 = F.relu(outputs4)
outputs6 = self.maxpool(outputs5)
outputs6 = paddle.reshape(outputs6, [outputs6.shape[0], 980])
outputs7 = self.fc(outputs6)
if check_shape:
print("\n######## print networks layer's superparams #########")
print("conv1-- kernel_size:{}, padding:{}, stride:{}".format(self.conv1.weight.shape, self.conv1._padding, self.conv1._stride))
print("conv2-- kernel_size:{}, padding:{}, stride:{}".format(self.conv2.weight.shape, self.conv2._padding, self.conv2._stride))
print("fc-- weight_size:{}, bias_size_{}".format(self.fc.weight.shape, self.fc.bias.shape))
#打印每层的输出尺寸
print("\n####### print shape of features of every layer ###########")
print("inputs_shape:{}".format(inputs.shape))
print("outputs1_shape:{}".format(outputs1.shape))
print("outputs2_shape:{}".format(outputs2.shape))
print("outputs3_shape:{}".format(outputs3.shape))
print("outputs4_shape:{}".format(outputs4.shape))
print("outputs5_shape:{}".format(outputs5.shape))
print("outputs6_shape:{}".format(outputs6.shape))
print("outputs7_shape:{}".format(outputs7.shape))
if check_content:
print("\n########print convolution layer's kernel ########")
print("conv1 params -- kernel weights:", self.conv1.weight[0][0])
print("conv2 params -- kernel weights:", self.conv2.weight[0][0])
#创建随机数，随机打印某一个通道的输出值
idx1 = np.random.randint(0, outputs1.shape[1])
idx2 = np.random.randint(0, outputs4.shape[1])
#打印卷积-池化后的结果，进打印batch中第一个图像对应的特征
print("\nThe {}th channel of conv1 layer:".format(idx1), outputs1[0][idx1])
print("The {}th channel of conv2 layer:".format(idx2),outputs4[0][idx2])
print("The output of last layer:", outputs7[0], '\n')
if label is not None:
acc = paddle.metric.accuracy(input=F.softmax(outputs7), label=label)
return outputs7, acc
else:
return outputs7
transform = paddle.vision.transforms.Compose([paddle.vision.ToTensor(),
paddle.vision.Normalize((0.1307,),(0.3081,))])
train_data = paddle.vision.datasets.MNIST(mode='train', transform=transform)
train_loader = paddle.io.DataLoader(train_data, shuffle=True, batch_size=100)
def train(model):
model.train()
EPOCH_NUM = 3
opt = paddle.optimizer.Adam(learning_rate=0.01, weight_decay=paddle.regularizer.L2Decay(), parameters=model.parameters())
iter1 = 0
iters = []
losses = []
accs = []
for epoch_id in range(EPOCH_NUM):
for batch_id, data in enumerate(train_loader()):
imgs, labels = data
imgs = paddle.to_tensor(imgs)
labels = paddle.to_tensor(labels)
predicts, acc = model(imgs, labels)
predicts = predicts.astype('float32')
labels = labels.astype('float32')
loss = F.square_error_cost(predicts, labels)
avg_loss = paddle.mean(loss)
if batch_id % 100 == 0:
print("epoch:{}, batch:{}, loss:{}, acc:{}".format(epoch_id, batch_id, avg_loss.numpy(), acc.numpy()))
iters.append(iter1)
iter1 += 100
losses.append(avg_loss.numpy())
accs.append(acc.numpy())
avg_loss.backward()
opt.step()
opt.clear_grad()
paddle.save(model.state_dict(), "mnist1.pdparams")
return iters, losses, accs
model = MNIST()
iters, losses, accs = train(model)输出:
epoch:0, batch:0, loss:[41.008236], acc:[0.18]
epoch:0, batch:100, loss:[21.90275], acc:[0.13]
epoch:0, batch:200, loss:[14.258028], acc:[0.12]
epoch:0, batch:300, loss:[14.339616], acc:[0.1]
epoch:0, batch:400, loss:[10.583021], acc:[0.1]
epoch:0, batch:500, loss:[10.328342], acc:[0.05]
epoch:1, batch:0, loss:[9.5673275], acc:[0.1]
epoch:1, batch:100, loss:[9.391169], acc:[0.04]
epoch:1, batch:200, loss:[8.717621], acc:[0.11]
epoch:1, batch:300, loss:[8.4292], acc:[0.07]
epoch:1, batch:400, loss:[9.091347], acc:[0.11]
epoch:1, batch:500, loss:[8.235337], acc:[0.11]
epoch:2, batch:0, loss:[9.305361], acc:[0.13]
epoch:2, batch:100, loss:[8.223233], acc:[0.09]
epoch:2, batch:200, loss:[9.535088], acc:[0.09]
epoch:2, batch:300, loss:[7.5045133], acc:[0.1]
epoch:2, batch:400, loss:[8.797383], acc:[0.07]
epoch:2, batch:500, loss:[9.310794], acc:[0.09]plt.figure()
plt.plot(iters, accs)
plt.show()输出：
