简 介： 对于有五中动物和五中水果组成的FAMNIST10（Fruit-Animal MNSIT数据集合），利用LeNet进行识别进行了测试。测试了不同的卷积核尺寸，不同的学习速率以及不同的数据集合的情况。通过测试也可以看到利用AI Studio中的GPU加速环境（至尊版环境）可以在训练模型过程中大大提高训练的效率。

关键词： LeNet，FAMNIST

§01 准备数据

  在 2021年人工神经网络第四次作业-第一题：LeNet识别水果与动物 制作了基于五种水果和五中动物的数据集合，总共907张图片。这个数据集合曾经是应用 第16届全国大学生智能车竞赛 中的 AI智能视觉组 的比赛数据集合。现在这个集合可以在 AI Studio FAMNIST 上下载。

1.1 解压缩文件

  将FAMNIST.zip文件上载到AI Studio环境内，通过“加压缩文件”将famnist文件解压缩到 /data 子目录下。它包括有如下五个基本的数据集合：

├─famnist-all
├─famnist10
├─famnist10-64
├─famnist10-64-gray
└─famnist10-gray

1.1.1 解压缩程序

  可以使用zipfile 软件包将zip文件进行解压缩。可以参谋如下的程序。

import sys,os,math,time
import matplotlib.pyplot as plt
from numpy import *
import zipfilefilename = '/home/aistudio/data/famnist.zip'
outpath = '/home/aistudio/data'with zipfile.ZipFile(filename) as zfile:zfile.extractall(outpath)print("Zip file extract to dir : %s"%outpath)

§02 训练LeNet网络

2.1 处理famnist10-64-gray

2.1.1 LeNet程序

#!/usr/local/bin/python
# -*- coding: gbk -*-
#============================================================
# LENET.PY                     -- by Dr. ZhuoQing 2021-12-16
#
# Note:
#============================================================from headm import *                 # =
import cv2import paddle
import paddle.nn.functional as F
from paddle import to_tensor as TT
from paddle.nn.functional import square_error_cost as sqrc#------------------------------------------------------------
famnist = '/home/aistudio/data/famnist'
imgdir = 'famnist10-64-gray'#------------------------------------------------------------
def loadimgdata(imgdir):'''loadimgdata: Load test image data into RAMParam imgdir: Directory for storing the image picture .Return: imgdata,imglabel'''imgfile = os.listdir(imgdir)imgdata = []imglabel = []for f in imgfile:img = cv2.imread(os.path.join(imgdir, f))imgdata.append(img.T[0][newaxis,:])imglabel.append(int(f[:1]))return array(imgdata), array(imglabel)imgdata,imglabel = loadimgdata(os.path.join(famnist,imgdir))print(len(imglabel))
print(imgdata.shape)#------------------------------------------------------------
'''
printf(imgdata.shape)
printf(imglabel)
'''
#------------------------------------------------------------
class famnist(paddle.io.Dataset):def __init__(self, num_samples):super(famnist, self).__init__()self.num_samples = num_samplesdef __getitem__(self, index):data = imgdata[index]/255label = imglabel[index]return TT(data, dtype='float32'), TT(label, dtype='int64')def __len__(self):return self.num_samples_dataset = famnist(800)
train_loader = paddle.io.DataLoader(_dataset, batch_size=100, shuffle=True)#------------------------------------------------------------#data = train_loader().next()
#print(data[0].numpy())'''ROW_NUM = 3
COL_NUM = 5plt.figure(figsize=(10,6))for j in range(ROW_NUM):for i in range(COL_NUM):id = j*COL_NUM + iimg = data[0].numpy()[id][0].T#        print(img.shape)plt.subplot(ROW_NUM, COL_NUM, id+1)plt.imshow(img, cmap='gray')plt.axis('off')'''
#------------------------------------------------------------
imageSize = 64
ks = 5
L = ((imageSize-ks+1)//2-ks+1)//2class mnist(paddle.nn.Layer):def __init__(self, ):super(mnist, self).__init__()self.conv1 = paddle.nn.Conv2D(in_channels=1, out_channels=6, kernel_size=ks, stride=1, padding=0)self.conv2 = paddle.nn.Conv2D(in_channels=6, out_channels=16, kernel_size=ks, stride=1, padding=0)self.mp1    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)self.mp2    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)self.L1     = paddle.nn.Linear(in_features=16*L*L, out_features=120)self.L2     = paddle.nn.Linear(in_features=120, out_features=86)self.L3     = paddle.nn.Linear(in_features=86, out_features=10)def forward(self, x):x = self.conv1(x)x = F.relu(x)x = self.mp1(x)x = self.conv2(x)x = F.relu(x)x = self.mp2(x)x = paddle.flatten(x, start_axis=1, stop_axis=-1)x = self.L1(x)x = F.relu(x)x = self.L2(x)x = F.relu(x)x = self.L3(x)return x#------------------------------------------------------------
net = mnist()EPOCH_NUM = 100
optimizer = paddle.optimizer.Adam(learning_rate=0.001, parameters=net.parameters())for epoch in range(EPOCH_NUM):for batchid, data in enumerate(train_loader()):out = net(data[0])loss = F.cross_entropy(out, data[1])acc = paddle.metric.accuracy(out, data[1])loss.backward()optimizer.step()optimizer.clear_grad()if batchid %100==0:print("Pass:{}, Loss:{}, Acc:{}".format(epoch,loss.numpy(), acc.numpy()))#------------------------------------------------------------
paddle.save(net.state_dict(), './work/model.pdparams')#------------------------------------------------------------
#        END OF FILE : LENET.PY
#============================================================

（1）网络结构

paddle.summary(net, input_size=(100,1,64,64))

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-1      [[100, 1, 64, 64]]    [100, 6, 60, 60]         156      MaxPool2D-1    [[100, 6, 60, 60]]    [100, 6, 30, 30]          0       Conv2D-2      [[100, 6, 30, 30]]   [100, 16, 26, 26]        2,416     MaxPool2D-2   [[100, 16, 26, 26]]   [100, 16, 13, 13]          0       Linear-1        [[100, 2704]]          [100, 120]          324,600    Linear-2         [[100, 120]]          [100, 86]           10,406     Linear-3         [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 338,448
Trainable params: 338,448
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 31.08
Params size (MB): 1.29
Estimated Total Size (MB): 33.93
---------------------------------------------------------------------------{'total_params': 338448, 'trainable_params': 338448}

2.1.2 训练结果

（1）训练速度

• 训练时间为：16.285秒。

▲ 图2.1.1 训练过程中的Loss与Accuracy的变化曲线

strid = 5
tspgetdopstring(-strid)
strall = clipboard.paste().split('\r\n')accdim = []
lossdim = []
for s in strall:ss = s.split(',')acc = ss[-1].replace('Acc:[','').replace(']','')loss = ss[1].replace('Loss:[','').replace(']','')accdim.append(float(acc.strip(' ')))lossdim.append(float(loss))printf(accdim)plt.plot(accdim, label='Accuracy')
plt.plot(lossdim, label='Loss')
plt.xlabel("Step")
plt.ylabel("Accuracy")
plt.legend(loc='upper right')
plt.grid(True)
plt.title("Train Accuracy")
plt.tight_layout()
plt.show()

2.1.3 不同尺寸的网络结构

（1）卷积核边长为3

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-3      [[100, 1, 64, 64]]    [100, 6, 62, 62]         60       MaxPool2D-3    [[100, 6, 62, 62]]    [100, 6, 31, 31]          0       Conv2D-4      [[100, 6, 31, 31]]   [100, 16, 29, 29]         880      MaxPool2D-4   [[100, 16, 29, 29]]   [100, 16, 14, 14]          0       Linear-4        [[100, 3136]]          [100, 120]          376,440    Linear-5         [[100, 120]]          [100, 86]           10,406     Linear-6         [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 388,656
Trainable params: 388,656
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 34.82
Params size (MB): 1.48
Estimated Total Size (MB): 37.86
---------------------------------------------------------------------------

  kernal_size = 3
  训练时间： 10.673秒

▲ 图2.1.2 Loss与Accuracy曲线

（2）卷积核变成为4

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-11     [[100, 1, 64, 64]]    [100, 6, 61, 61]         102      MaxPool2D-11    [[100, 6, 61, 61]]    [100, 6, 30, 30]          0       Conv2D-12     [[100, 6, 30, 30]]   [100, 16, 27, 27]        1,552     MaxPool2D-12   [[100, 16, 27, 27]]   [100, 16, 13, 13]          0       Linear-16       [[100, 2704]]          [100, 120]          324,600    Linear-17        [[100, 120]]          [100, 86]           10,406     Linear-18        [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 337,530
Trainable params: 337,530
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 32.28
Params size (MB): 1.29
Estimated Total Size (MB): 35.13
---------------------------------------------------------------------------

▲ 图A2.1.3 Loss 与Accuracy曲线

（3）卷积核边长为7

  训练时间：10.529秒

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-5      [[100, 1, 64, 64]]    [100, 6, 58, 58]         300      MaxPool2D-5    [[100, 6, 58, 58]]    [100, 6, 29, 29]          0       Conv2D-6      [[100, 6, 29, 29]]   [100, 16, 23, 23]        4,720     MaxPool2D-6   [[100, 16, 23, 23]]   [100, 16, 11, 11]          0       Linear-7        [[100, 1936]]          [100, 120]          232,440    Linear-8         [[100, 120]]          [100, 86]           10,406     Linear-9         [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 248,736
Trainable params: 248,736
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 27.35
Params size (MB): 0.95
Estimated Total Size (MB): 29.86
---------------------------------------------------------------------------

▲ 图2.1.3 Loss与Accuracy曲线

（4）卷积核边长为2

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-9      [[100, 1, 64, 64]]    [100, 6, 63, 63]         30       MaxPool2D-9    [[100, 6, 63, 63]]    [100, 6, 31, 31]          0       Conv2D-10     [[100, 6, 31, 31]]   [100, 16, 30, 30]         400      MaxPool2D-10   [[100, 16, 30, 30]]   [100, 16, 15, 15]          0       Linear-13       [[100, 3600]]          [100, 120]          432,120    Linear-14        [[100, 120]]          [100, 86]           10,406     Linear-15        [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 443,826
Trainable params: 443,826
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 36.47
Params size (MB): 1.69
Estimated Total Size (MB): 39.72
---------------------------------------------------------------------------

▲ 图A2.1.4 Loss 与Accuracy曲线

（5）卷积核边长为1

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-7      [[100, 1, 64, 64]]    [100, 6, 64, 64]         12       MaxPool2D-7    [[100, 6, 64, 64]]    [100, 6, 32, 32]          0       Conv2D-8      [[100, 6, 32, 32]]   [100, 16, 32, 32]         112      MaxPool2D-8   [[100, 16, 32, 32]]   [100, 16, 16, 16]          0       Linear-10       [[100, 4096]]          [100, 120]          491,640    Linear-11        [[100, 120]]          [100, 86]           10,406     Linear-12        [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 503,040
Trainable params: 503,040
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 39.23
Params size (MB): 1.92
Estimated Total Size (MB): 42.71
---------------------------------------------------------------------------

▲ 图2.1.4 Loss与Accuracy曲线

2.1.4 不同的学习速率

  选择卷积核为5的情况下，测试不同的学习速率对于训练结果的影响。

（1）rate=0.1

▲ 图2.1.7 Loss and Accuracy

（2）0.01

▲ 图2.1.8 Loss and Accuracy

（3）0.001

▲ 图2.1.9 Loss and Accuracy

（4）rate=0.0001

▲ 图2.1.10 Loss and Accuracy

（5）0.00001

▲ 图2.1.11 Loss and Accuracy

2.1.5 细分不同学习速率下的收敛情况

plt.figure(figsize=(10, 6))
EPOCH_NUM = 100for i in linspace(1, 4, 20):lr = 1/exp((i+1)*log(10))net = mnist()net.__init__()optimizer = paddle.optimizer.Adam(learning_rate=lr, parameters=net.parameters())lossdim = []accdim = []print('Step:%d, LR:%0.6f'%(i+1, lr))for epoch in range(EPOCH_NUM):for batchid, data in enumerate(train_loader()):out = net(data[0])loss = F.cross_entropy(out, data[1])acc = paddle.metric.accuracy(out, data[1])loss.backward()optimizer.step()optimizer.clear_grad()if epoch %10==0:print("Pass:{}, Loss:{}, Acc:{}".format(epoch,loss.numpy(), acc.numpy()))lossdim.append(loss.numpy())accdim.append(acc.numpy())plt.plot(accdim, label='lr:%.6f'%lr)plt.xlabel("Step")plt.ylabel("Accuracy")plt.legend(loc='upper right')plt.grid(True)plt.tight_layout()plt.show()

▲ 图2.1.12 不同学习速率对应的收敛情况

▲ 图2.1.13 不同学习速率对应的收敛情况

▲ 图2.1.14 不同学习速率对应的收敛情况

§03 测试LeNet网络

3.1 在CPU下训练网络

训练参数：

ks=5：
lr=0.01：
训练时间：75.17s
训练样本：906

3.1.1 网络结构

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-5      [[100, 1, 64, 64]]    [100, 6, 60, 60]         156      MaxPool2D-5    [[100, 6, 60, 60]]    [100, 6, 30, 30]          0       Conv2D-6      [[100, 6, 30, 30]]   [100, 16, 26, 26]        2,416     MaxPool2D-6   [[100, 16, 26, 26]]   [100, 16, 13, 13]          0       Linear-7        [[100, 2704]]          [100, 120]          324,600    Linear-8         [[100, 120]]          [100, 86]           10,406     Linear-9         [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 338,448
Trainable params: 338,448
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.56
Forward/backward pass size (MB): 31.08
Params size (MB): 1.29
Estimated Total Size (MB): 33.93
---------------------------------------------------------------------------

3.1.2 训练收敛情况

▲ 图3.1.1 Loss and Accuracy

3.2 测试网络

data = train_loader().next()
out = net(data[0][:30])
out1 = paddle.argmax(out, axis=1).numpy()print(out1)
print(data[1][:30].numpy().T.flatten())

[0 9 9 0 5 7 2 4 4 8 0 6 2 7 4 7 4 8 2 6 4 6 3 9 4 1 4 7 3 2]
[0 9 9 0 5 7 2 4 4 8 0 6 2 7 4 7 4 8 2 6 4 6 3 9 4 1 4 7 3 2]

§04 测试FAMNIST10

4.1 测试FAMNIST10-Gray

4.1.1 网络结构

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-7      [[100, 1, 32, 32]]    [100, 6, 28, 28]         156      MaxPool2D-7    [[100, 6, 28, 28]]    [100, 6, 14, 14]          0       Conv2D-8      [[100, 6, 14, 14]]   [100, 16, 10, 10]        2,416     MaxPool2D-8   [[100, 16, 10, 10]]    [100, 16, 5, 5]           0       Linear-10        [[100, 400]]          [100, 120]          48,120     Linear-11        [[100, 120]]          [100, 86]           10,406     Linear-12        [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 61,968
Trainable params: 61,968
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.39
Forward/backward pass size (MB): 6.18
Params size (MB): 0.24
Estimated Total Size (MB): 6.80
---------------------------------------------------------------------------

4.1.2 训练结果

▲ 图4.1.1 收敛情况

  可以看到LeNet对于FAMNIST10也同样具有很好的识别特性。

4.2 测试FAMNIST10

4.2.1 网络结构

---------------------------------------------------------------------------Layer (type)       Input Shape          Output Shape         Param #
===========================================================================Conv2D-15     [[100, 3, 32, 32]]    [100, 6, 28, 28]         456      MaxPool2D-15    [[100, 6, 28, 28]]    [100, 6, 14, 14]          0       Conv2D-16     [[100, 6, 14, 14]]   [100, 16, 10, 10]        2,416     MaxPool2D-16   [[100, 16, 10, 10]]    [100, 16, 5, 5]           0       Linear-22        [[100, 400]]          [100, 120]          48,120     Linear-23        [[100, 120]]          [100, 86]           10,406     Linear-24        [[100, 86]]           [100, 10]             870
===========================================================================
Total params: 62,268
Trainable params: 62,268
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 1.17
Forward/backward pass size (MB): 6.18
Params size (MB): 0.24
Estimated Total Size (MB): 7.59
---------------------------------------------------------------------------

4.2.2 训练结果

▲ 图4.2.1 训练收敛情况

4.3 测试原始图像

原始图像参数：

尺寸：283×283
色彩：彩色图像

  在AI Studio 至尊版环境下，花费了88.68秒训练完毕。

4.3.1 网络结构

----------------------------------------------------------------------------Layer (type)        Input Shape          Output Shape         Param #
============================================================================Conv2D-1     [[100, 3, 283, 283]]   [100, 6, 279, 279]        456      MaxPool2D-1   [[100, 6, 279, 279]]   [100, 6, 139, 139]         0       Conv2D-2     [[100, 6, 139, 139]]  [100, 16, 135, 135]       2,416     MaxPool2D-2   [[100, 16, 135, 135]]  [100, 16, 67, 67]          0       Linear-1        [[100, 71824]]          [100, 120]         8,619,000   Linear-2         [[100, 120]]           [100, 86]           10,406     Linear-3          [[100, 86]]           [100, 10]             870
============================================================================
Total params: 8,633,148
Trainable params: 8,633,148
Non-trainable params: 0
----------------------------------------------------------------------------
Input size (MB): 91.65
Forward/backward pass size (MB): 722.21
Params size (MB): 32.93
Estimated Total Size (MB): 846.80
----------------------------------------------------------------------------

+--------------+---------------------+---------------------+---------+------------+
|  Layer Name  |     Input Shape     |     Output Shape    |  Params |   Flops    |
+--------------+---------------------+---------------------+---------+------------+
|   conv2d_0   |  [100, 3, 283, 283] |  [100, 6, 279, 279] |   456   | 3549549568 |
|   conv2d_1   |  [100, 6, 139, 139] | [100, 16, 135, 135] |   2416  | 4403160064 |
| max_pool2d_0 |  [100, 6, 279, 279] |  [100, 6, 139, 139] |    0    |     0      |
| max_pool2d_1 | [100, 16, 135, 135] |  [100, 16, 67, 67]  |    0    |     0      |
|   linear_0   |     [100, 71824]    |      [100, 120]     | 8619000 | 861888000  |
|   linear_1   |      [100, 120]     |      [100, 86]      |  10406  |  1032000   |
|   linear_2   |      [100, 86]      |      [100, 10]      |   870   |   86000    |
+--------------+---------------------+---------------------+---------+------------+
Total Flops: 8815715632     Total Params: 8633148

4.3.2 训练情况

▲ 图4.3.1 网络收敛情况

※ 测试总结 ※

  对于有五中动物和五中水果组成的FAMNIST10（Fruit-Animal MNSIT数据集合），利用LeNet进行识别进行了测试。测试了不同的卷积核尺寸，不同的学习速率以及不同的数据集合的情况。

  通过测试也可以看到利用AI Studio中的GPU加速环境（至尊版环境）可以在训练模型过程中大大提高训练的效率。

■ 相关文献链接:

• 404
• 第十六届全国大学智能汽车竞赛竞速比赛规则
• 第十六届智能汽车竞赛AI视觉组分赛区数据集发布
• AI Studio FAMNIST

● 相关图表链接:

• 图2.1.1 训练过程中的Loss与Accuracy的变化曲线
• 图2.1.2 Loss与Accuracy曲线
• 图A2.1.3 Loss 与Accuracy曲线
• 图2.1.3 Loss与Accuracy曲线
• 图A2.1.4 Loss 与Accuracy曲线
• 图2.1.4 Loss与Accuracy曲线
• 图2.1.7 Loss and Accuracy
• 图2.1.8 Loss and Accuracy
• 图2.1.9 Loss and Accuracy
• 图2.1.10 Loss and Accuracy
• 图2.1.11 Loss and Accuracy
• 图2.1.12 不同学习速率对应的收敛情况
• 图2.1.13 不同学习速率对应的收敛情况
• 图2.1.14 不同学习速率对应的收敛情况
• 图3.1.1 Loss and Accuracy
• 图4.1.1 收敛情况
• 图4.2.1 训练收敛情况
• 图4.3.1 网络收敛情况

一、实验程序

1、解压缩数据文件

#!/usr/local/bin/python
# -*- coding: gbk -*-
#============================================================
# UNZIP.PY                     -- by Dr. ZhuoQing 2021-12-16
#
# Note:
#============================================================from headm import *                 # =
import zipfilefilename = '/home/aistudio/data/famnist.zip'
outpath = '/home/aistudio/data'with zipfile.ZipFile(filename) as zfile:zfile.extractall(outpath)printf("Zip file extract to dir : %s"%outpath)#------------------------------------------------------------
#        END OF FILE : UNZIP.PY
#============================================================

2、绘制训练曲线

#!/usr/local/bin/python
# -*- coding: gbk -*-
#============================================================
# DRAW.PY                      -- by Dr. ZhuoQing 2021-12-16
#
# Note:
#============================================================
from headm import *
strid = 8
#tspgetdopstring(-strid)
tspsendwindowkey('BML CodeLab', "c", control=1)
strall = clipboard.paste().split('\r\n')
accdim = []
lossdim = []
for s in strall:ss = s.split(',')acc = ss[-1].replace('Acc:[','').replace(']','')loss = ss[1].replace('Loss:[','').replace(']','')accdim.append(float(acc.strip(' ')))lossdim.append(float(loss))
printf(accdim)
plt.plot(accdim, label='Accuracy')
plt.plot(lossdim, label='Loss')
plt.xlabel("Step")
plt.ylabel("Accuracy")
plt.legend(loc='upper right')
plt.grid(True)
plt.title("Train Accuracy")
plt.tight_layout()
plt.show()
#------------------------------------------------------------
#        END OF FILE : DRAW.PY
#============================================================

3、实验程序

#!/usr/local/bin/python
# -*- coding: gbk -*-
#============================================================
# LENET.PY                     -- by Dr. ZhuoQing 2021-12-16
#
# Note:
#============================================================from headm import *                 # =
import cv2import paddle
import paddle.nn.functional as F
from paddle import to_tensor as TT
from paddle.nn.functional import square_error_cost as sqrc#------------------------------------------------------------
famnist = '/home/aistudio/data/famnist'
imgdir = 'famnist-all'#------------------------------------------------------------
def loadimgdata(imgdir):'''loadimgdata: Load test image data into RAMParam imgdir: Directory for storing the image picture .Return: imgdata,imglabel'''imgfile = os.listdir(imgdir)imgdata = []imglabel = []for f in imgfile:img = cv2.imread(os.path.join(imgdir, f))imgdata.append(img.T)imglabel.append(int(f[:1]))return array(imgdata), array(imglabel)imgdata,imglabel = loadimgdata(os.path.join(famnist,imgdir))print(len(imglabel))
print(imgdata.shape)#------------------------------------------------------------
'''
printf(imgdata.shape)
printf(imglabel)
'''
#------------------------------------------------------------
class famnist(paddle.io.Dataset):def __init__(self, num_samples):super(famnist, self).__init__()self.num_samples = num_samplesdef __getitem__(self, index):data = imgdata[index]/255label = imglabel[index]return TT(data, dtype='float32'), TT(label, dtype='int64')def __len__(self):return self.num_samples_dataset = famnist(len(imglabel))
train_loader = paddle.io.DataLoader(_dataset, batch_size=100, shuffle=True)#------------------------------------------------------------#data = train_loader().next()
#print(data[0].numpy())'''ROW_NUM = 3
COL_NUM = 5plt.figure(figsize=(10,6))for j in range(ROW_NUM):for i in range(COL_NUM):id = j*COL_NUM + iimg = data[0].numpy()[id][0].T#        print(img.shape)plt.subplot(ROW_NUM, COL_NUM, id+1)plt.imshow(img, cmap='gray')plt.axis('off')'''
#------------------------------------------------------------
imageSize = 283
ks = 5
in_channel=3
L = ((imageSize-ks+1)//2-ks+1)//2class mnist(paddle.nn.Layer):def __init__(self, ):super(mnist, self).__init__()self.conv1 = paddle.nn.Conv2D(in_channels=in_channel, out_channels=6, kernel_size=ks, stride=1, padding=0)self.conv2 = paddle.nn.Conv2D(in_channels=6, out_channels=16, kernel_size=ks, stride=1, padding=0)self.mp1    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)self.mp2    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)self.L1     = paddle.nn.Linear(in_features=16*L*L, out_features=120)self.L2     = paddle.nn.Linear(in_features=120, out_features=86)self.L3     = paddle.nn.Linear(in_features=86, out_features=10)def forward(self, x):x = self.conv1(x)x = F.relu(x)x = self.mp1(x)x = self.conv2(x)x = F.relu(x)x = self.mp2(x)x = paddle.flatten(x, start_axis=1, stop_axis=-1)x = self.L1(x)x = F.relu(x)x = self.L2(x)x = F.relu(x)x = self.L3(x)return x#------------------------------------------------------------net = mnist()EPOCH_NUM = 100
optimizer = paddle.optimizer.Adam(learning_rate=0.01, parameters=net.parameters())for epoch in range(EPOCH_NUM):for batchid, data in enumerate(train_loader()):out = net(data[0])loss = F.cross_entropy(out, data[1])acc = paddle.metric.accuracy(out, data[1])loss.backward()optimizer.step()optimizer.clear_grad()if batchid %100==0:print("Pass:{}, Loss:{}, Acc:{}".format(epoch,loss.numpy(), acc.numpy()))#------------------------------------------------------------
'''
plt.figure(figsize=(10, 6))
EPOCH_NUM = 100for i in linspace(0, 2, 20):lr = 1/exp((i+1)*log(10))net = mnist()net.__init__()optimizer = paddle.optimizer.Adam(learning_rate=lr, parameters=net.parameters())lossdim = []accdim = []printf('Step:%d, LR:%0.6f'%(i+1, lr))for epoch in range(EPOCH_NUM):for batchid, data in enumerate(train_loader()):out = net(data[0])loss = F.cross_entropy(out, data[1])acc = paddle.metric.accuracy(out, data[1])loss.backward()optimizer.step()optimizer.clear_grad()if epoch %10==0:print("Pass:{}, Loss:{}, Acc:{}".format(epoch,loss.numpy(), acc.numpy()))lossdim.append(loss.numpy())accdim.append(acc.numpy())plt.plot(accdim, label='lr:%.6f'%lr)plt.xlabel("Step")plt.ylabel("Accuracy")plt.legend(loc='upper right')plt.grid(True)plt.tight_layout()plt.show()'''
#------------------------------------------------------------
paddle.save(net.state_dict(), './work/model3.pdparams')#------------------------------------------------------------
print(' ')
paddle.summary(net, input_size=(100,in_channel,imageSize,imageSize))paddle.flops(net, input_size=(100, in_channel, imageSize, imageSize), print_detail=True)#------------------------------------------------------------
'''
data = train_loader().next()
out = net(data[0][:30])
out1 = paddle.argmax(out, axis=1).numpy()printf(out1)
printf(data[1][:30].numpy().T.flatten())'''
#------------------------------------------------------------
#        END OF FILE : LENET.PY
#============================================================

对于FAMNIST中的十种动物和水果进行识别测试相关推荐

1. 利用LeNet识别十种动物和水果

   简 介: ※利用经典的LeNet深度学习网络,可以完成对于智能车竞赛中智能视觉组对搬运物品分类的要求.虽然实际比赛中还会包含有五种交通工具,那么利用LeNet仍然是可以完成小类别的识别与定位的.仅仅使 ...

2. 生活中的十种保养皮肤的食物

   生活中的十种保养皮肤的食物 1.西兰花--它含有丰富的维生素A.维生素C和胡萝卜素,能增强皮肤的抗损伤能力.有助于保持皮肤弹性. 2.胡萝卜--胡萝卜素有助于维持皮肤细胞组织的正常机能.减少皮肤皱纹, ...

3. Spring实例(IOC)——动物吃水果

   Spring实例"IOC"--动物吃水果 一.问题:在控制台打印XX(动物)吃XX(水果) 二.解决方法 2.1.常规方法(new对象) 2.2.Spring-ioc(控制反转) ...

4. 学数学建模算法对计算机的好处,数学建模中常见十种算法 (期末论文).doc

   数学建模中常见十种算法 (期末论文) 数 学 系 毕 业 论 文 论文 (设计)题目: 数学建模中常见的十种算法 姓 名 黄小芬______ 学 号 100501313 专 业 数学与应用数学 班 级 ...

5. 英文投稿过程中的十种状态

   以下是英文投稿过程中的十种状态. 1. Submitted to Journal 当上传结束后,显示的状态是Submitted to Journal,这bai个状态是自然形成的无需处理. 2. Wit ...

6. Animals-10 数据集下载，十种动物

   kaggle上的一个数据集 十种动物 狗,猫,马,大象,蝴蝶,鸡,牛,羊,蜘蛛,松鼠 每个种类一两千图片, 也有四五千的. 格式.jpeg 官方下载地址:https://www.kaggle.com/ ...

7. 【02】水果蔬菜识别系统-基于tensorflow2.3开发

   ------------------------------------------------2021年6月18日重大更新-------------------------------------- ...

8. 水果数据集(Fruit-Dataset )+水果分类识别训练代码(支持googlenet, resnet, inception_v3, mobilenet_v2)

   水果数据集(Fruit-Dataset )+水果分类识别训练代码(支持googlenet, resnet, inception_v3, mobilenet_v2) 目录 Fruit-Dataset水果 ...

9. 基于卷积神经网络CNN的水果分类预测，卷积神经网络水果等级识别

   目录 背影 卷积神经网络CNN的原理 卷积神经网络CNN的定义 卷积神经网络CNN的神经元 卷积神经网络CNN的激活函数 卷积神经网络CNN的传递函数 卷积神经网络CNN水果分类预测 基本结构 主要参 ...

最新文章

1. C# WinForm 通过URL取得服务器上的某图片文件到本地
2. JVM 监控以及内存分析
3. Spark2.11 两种流操作 + Kafka
4. .htaccess更改目录下的默认主页
5. AD16更改器件封装如管脚间距等常规设置
6. python具有可嵌入性_如何构建可嵌入Python
7. Acrel-3200远程预付费电能管理系统的功能介绍
8. 【小程序实战系列】电商平台源码及功能实现
9. yahoo也推出站点统计了
10. 斗地主牌型基本算法升级版本
11. “校园知网”5月5日冲刺计划书
12. C# 通过Http获取网页内容
13. 建筑工程测量与测绘毕业论文范文
14. html5语文答题制作,语文万能答题模板
15. Event-triggered MPC Design for Distributed Systems With Network Communications
16. python编程题12-计算并联电阻
17. 环境变量配置以mysql为例
18. 目前最值得收藏的100个各类资源站
19. 使用Python修改可执行文件-EXE的图标
20. 百度IFE2018任务--20-21天

热门文章

1. 从Visual Studio里抓取抽象语法树（AST）
2. 《程序员之禅》一一第3章 为什么要进行禅修编程
3. SQL中CONVERT转化函数的用法
4. poj 2677 tour
5. ubuntu平台下搭建PHPWind网站运行环境
6. 刚刚，蚂蚁金服荣膺“中国金融大数据领军企业”称号
7. Python--day28--set去重
8. python全栈学习笔记（一）网络基础之网络协议篇
9. 对于容器环境来说 全栈监控究竟意味着什么？
10. vs2013编译 protoBuffer编译出现的问题