在百度AI框架PaddlePaddle下，利用卷积神经网络CNN实现数字0到9手势的分类识别。
- 数据集下载 https://aistudio.baidu.com/aistudio/datasetdetail/29044
- 数据集说明
数据集内容： 图片中的手势表示0-9这10个数字，其中数据集解压后文件夹的名称就是该文件夹中的图片中的手势表示的数字；
数据集来源： 百度AI Studio data数据集； 图片说明： 彩色照片，大小100*100像素，背景统一为白色； - 各类别数量
# 解压缩数据集
!cd /home/aistudio/data/data29044 && unzip -q -o GestureC.zip
# 引入依赖的各种包
import os
# 统计各种类别图片的个数
def get_sample_count(path):
digital_count_dict = {}
for files in os.listdir(path):
count = 0
for filename in os.listdir('%s/%s' %(path, files)):
count += 1
digital_count_dict[files] = count
return digital_count_dict
# 各个类别图片的个数
get_sample_count(r'data/data29044/GestureC')
{'8': 208, '0': 205, '9': 204,'7': 206, '1': 206, '5': 207, '3': 206,'6': 207, '4': 207, '2': 206}- 将数据集中图片的路径变为PaddlePaddle中的.list形式
# 将数据集中的图片的路径写入该框架下的.list格式的文件, 每一行的前半部分是图片路径，一个空格后面是该图片手势表示的数字
def make_path_list(path):
if(os.path.exists('./alldata.list')):
os.remove('./alldata.list')
with open('./alldata.list', 'a') as alldata:
for files in os.listdir(path):
for img_path in os.listdir('%s/%s' % (path, files)):
alldata.write(os.path.join(path, img_path) + '\t' + files + '\n')
return print('列表已生成')
make_path_list(r'/home/aistudio/data/data29044/GestureC')- 按照7:2:1的比例将数据集变为训练数据集、测试数据集、验证数据集
# 引入依赖的包
import numpy as np
digital_count_dict = get_sample_count(r'/home/aistudio/data/data29044/GestureC')
def split_dataset(listpath, test_p=0.2, eval_p=0.1, count_dict=digital_count_dict):
# 每次运行更新
if(os.path.exists('./train_data.list')):
os.remove('./train_data.list')
if(os.path.exists('./test_data.list')):
os.remove('./test_data.list')
if(os.path.exists('./eval_data.list')):
os.remove('./eval_data.list')
start_sign = 10
# 开始分割数据集合
with open(r'%s/alldata.list' % listpath, 'r') as f:
lines = f.readlines()
for line in lines:
img, label = line.replace('\n', '').split('\t')
# 说明是新的手势
if label != start_sign:
count_sign = 0
# 得到该手势的总数
total_count = count_dict[label]
# 用于测试的
test_count = int(total_count * test_p)
# 用于验证的
eval_count = int(total_count * eval_p)
# 该手势的图片编号
total_sign = np.arange(total_count)
# 随机打乱
np.random.shuffle(total_sign)
# 测试的编号
test_sign = total_sign[: test_count]
# 验证的编号
eval_sign = total_sign[test_count: (test_count + eval_count)]
if count_sign in test_sign:
with open('./test_data.list', 'a') as test_data:
test_data.write(line)
elif count_sign in eval_sign:
with open('./eval_data.list', 'a') as eval_data:
eval_data.write(line)
else:
with open('./train_data.list', 'a') as train_data:
train_data.write(line)
count_sign += 1
start_sign = label
return print('数据集分割完成')
split_dataset(r'/home/aistudio/data/data29044')# 看一下各个数据集的样本总数以及各个类别的数
def view_dataset(listpath):
count_dict = {}
for ij in os.listdir(listpath):
if '.list' in ij:
type_dict = {}
with open(r'%s/%s' % (listpath, ij), 'r') as f:
lines = f.readlines()
count = 0
for line in lines:
img, label = line.replace('\n', '').split('\t')
if label in type_dict:
type_dict[label] += 1
else:
type_dict[label] = 1
count += 1
count_dict[ij] = [count, type_dict]
return count_dict
view_dataset(r'/home/aistudio/data/data29044')
{'eval_data.list':
[200,{'8': 20,'0': 20,'9': 20,'7': 20,'1': 20,'5': 20,'3': 20,'6': 20,'4': 20,'2': 20}],
'test_data.list':
[409, {'8': 41,'0': 41,'9': 40,'7': 41, '1': 41,'5': 41,'3': 41,'6': 41,'4': 41,'2': 41}],
'train_data.list':
[1453,{'8': 147,'0': 144, '9': 144,'7': 145,'1': 145,'5': 146,'3': 145,'6': 146,'4': 146,'2': 145}],
'alldata.list': [2062, {'8': 208, '0': 205,'9': 204,'7': 206,'1': 206,'5': 207,'3': 206,'6': 207,'4': 207,'2': 206}]}- 将数据集合变为PaddlePaddle框架中的数据流的形式
# 定义数据集的reader
def data_mapper(sample):
img, label = sample
img = Image.open(img)
# 利用平滑采样的方式将图片变为100*100的
img = img.resize((100, 100), Image.ANTIALIAS)
img = np.array(img).astype('float32')
# 归一化
img = img/255.0
return img.shape, label
def data_reader(data_list_path):
def reader():
with open(data_list_path, 'r') as f:
lines = f.readlines()
for line in lines:
img, label = line.split('\t')
return paddle.reader.xmap_readers(data_mapper, reader, cpu_count(), 512)
# 用于训练的数据提供器
train_reader = paddle.batch(reader=paddle.reader.shuffle(reader=data_reader('./train_data.list'), buf_size=256), batch_size=32)
# 用于测试的数据提供器
test_reader = paddle.batch(reader=data_reader('./test_data.list'), batch_size=32)
# 用于验证的数据提供器
eval_reader = paddle.batch(reader=data_reader('./eval_data.list'), batch_size=32) # 引入需要的包
import paddle.fluid as fluid
import paddle.fluid.layers as layers
from multiprocessing import cpu_count
from paddle.fluid.dygraph import Pool2D,Conv2D
from paddle.fluid.dygraph import Linear# 配置卷积神经网络
bp_config = {'learningrate': 0.002, 'iters': 500}
# 定义网络
class CNNBP(fluid.dygraph.Layer):
def __init__(self):
super(CNNBP, self).__init__()
# 卷积层
self.conv1 = Conv2D(num_channels=3, num_filters=16, filter_size=5, stride=1, act='relu')
# 池化层
self.pool1 = Pool2D(pool_size=2, pool_stride=2, pool_type='max', pool_padding=0)
# 卷积层
self.conv2 = Conv2D(num_channels=16, num_filters=32, filter_size=3, stride=1, act='relu')
# 池化层
self.pool2 = Pool2D(pool_size=2, pool_stride=2, pool_type='max', pool_padding=0)
# 卷积层
self.conv3 = Conv2D(num_channels=32, num_filters=16, filter_size=3, stride=1, act='relu')
# 全连接层
self.hidden1 = Linear(16*21*21, 512, act='relu')
# 输出层
self.hidden2 = Linear(512, 10, act='softmax')
def forward(self, input):
x = self.conv1(input)
x = self.pool1(x)
x = self.conv2(x)
x = self.pool2(x)
x = self.conv3(x)
# 全连接层之前需要将数据的维度变为(1, N)
x = layers.flatten(x)
# 防止过拟合
x = fluid.layers.dropout(x, dropout_prob=0.5)
x = self.hidden1(x)
# 防止过拟合
x = fluid.layers.dropout(x, dropout_prob=0.5)
y = self.hidden2(x)
return y每迭代完一次，进行测试数据集的误差和精确度计算，如果满足设定的条件则保存模型，最后选择在测试数据集上表现最后的模型用于验证。
# 输出验证数据集、测试数据集的误差以及精确度
train_loss_plot = []
train_acc_plot = []
test_loss_plot = []
test_acc_plot = []
#用动态图进行训练
with fluid.dygraph.guard(fluid.CUDAPlace(0)): # 其中fluid.CUDAPlace(0)表示用GPU训练，否则为空。
model=CNNBP() #模型实例化
model.train() #训练模式
# 优化求解器
opt=fluid.optimizer.SGDOptimizer(learning_rate=bp_config['learningrate'], parameter_list=model.parameters())
for pass_num in range(bp_config['iters']):
print('迭代次数：', pass_num)
# 记录训练数据的误差和精确度
batch_train_loss = []
batch_train_acc = []
for batch_id, data in enumerate(train_reader()):
images=np.array([x[0].reshape(3,100,100) for x in data],np.float32)
labels = np.array([x[1] for x in data]).astype('int64')
labels = labels[:, np.newaxis]
image=fluid.dygraph.to_variable(images)
label=fluid.dygraph.to_variable(labels)
predict=model(image)#预测
loss=fluid.layers.cross_entropy(predict,label)
avg_loss=fluid.layers.mean(loss)#获取loss值
acc=fluid.layers.accuracy(predict,label)#计算精度
# 存储训练数据集的
batch_train_loss.append(avg_loss.numpy())
batch_train_acc.append(acc.numpy())
avg_loss.backward()
opt.minimize(avg_loss)
model.clear_gradients()
# 计算测试数据的精确度以及loss
taccs = []
for test_batch_id, test_data in enumerate(test_reader()):#测试集
test_images=np.array([x[0].reshape(3,100,100) for x in test_data],np.float32)
test_labels = np.array([x[1] for x in test_data]).astype('int64')
test_labels = test_labels[:, np.newaxis]
timage=fluid.dygraph.to_variable(test_images)
tlabel=fluid.dygraph.to_variable(test_labels)
tpredict=model(timage)
tacc=fluid.layers.accuracy(tpredict, tlabel)
tloss=fluid.layers.cross_entropy(tpredict, tlabel)
taccs.append(tacc.numpy()[0])
tavg_acc = np.mean(taccs)
# 存储测试数据结果
test_loss_plot.append(np.mean(tloss.numpy()))
test_acc_plot.append(tavg_acc)
# 存储训练数据集结果
train_loss_result = sum(batch_train_loss) / len(batch_train_loss)
train_acc_result = sum(batch_train_acc) / len(batch_train_acc)
train_loss_plot.append(train_loss_result)
train_acc_plot.append(train_acc_result)
print('训练数据集：误差', train_loss_result, '正确率', train_acc_result)
print('测试数据集：误差', np.mean(tloss.numpy()), '正确率', tavg_acc)
# 保存的模型
if train_acc_result > 0.95 and tavg_acc > 0.88: # 训练数据达到0.95以后就保存模型
fluid.save_dygraph(model.state_dict(),'./para/CNNBP%s' % pass_num)迭代次数： 495
训练数据集：误差 [0.05399397] 正确率 [0.98145604]
测试数据集：误差 0.3810585 正确率 0.9151923
迭代次数： 496
训练数据集：误差 [0.03725699] 正确率 [0.98832417]
测试数据集：误差 1.1552259 正确率 0.8939423
迭代次数： 497
训练数据集：误差 [0.04244164] 正确率 [0.98832417]
测试数据集：误差 0.55333513 正确率 0.89528847
迭代次数： 498
训练数据集：误差 [0.04658408] 正确率 [0.9855769]
测试数据集：误差 0.6396506 正确率 0.87125003
迭代次数： 499
训练数据集：误差 [0.03801583] 正确率 [0.9869506]
测试数据集：误差 0.07386498 正确率 0.89355767绘制出训练和测试数据集上的误差、精确度的对比曲线。
# 绘制图片
import matplotlib.pyplot as plt
from pylab import mpl
mpl.rcParams['font.sans-serif'] = ['FangSong'] # 显示中文
mpl.rcParams['axes.unicode_minus'] = False # 显示负号
# 绘制训练数据集和测试数据集的对比曲线
def draw_loss(traindata, testdata, yl='loss'):
plt.title('train VS test', fontsize=24)
plt.xlabel("iters", fontsize=20)
plt.ylabel(yl, fontsize=20)
plt.plot(list(range(len(traindata))), traindata, color='red',label='Train')
plt.plot(list(range(len(testdata))), testdata, color='green',label='Test')
plt.legend()
plt.grid()
plt.show()
draw_loss(train_loss_plot, test_loss_plot)
draw_loss(train_acc_plot, test_acc_plot, 'Acc')
# 根据测试数据集合的精确率选择效果最后的模型
judge_best = [[i, test_acc_plot[i], test_loss_plot[i]] for i in range(len(test_acc_plot))]
best_model_sign = sorted(judge_best, key= lambda x: (-x[1], x[2]))[0][0]
# 根据得到的最好的模型针对验证数据集进行精确率的计算
with fluid.dygraph.guard(fluid.CUDAPlace(0)):
accs = []
model_dict, _ = fluid.load_dygraph('./para/CNNBP%s' % best_model_sign)
model = CNNBP()
model.load_dict(model_dict) #加载模型参数
model.eval() #训练模式
for batch_id,data in enumerate(eval_reader()):#测试集
images=np.array([x[0].reshape(3,100,100) for x in data],np.float32)
labels = np.array([x[1] for x in data]).astype('int64')
labels = labels[:, np.newaxis]
image=fluid.dygraph.to_variable(images)
label=fluid.dygraph.to_variable(labels)
predict=model(image)
acc=fluid.layers.accuracy(predict,label)
accs.append(acc.numpy()[0])
avg_acc = np.mean(accs)
print(avg_acc)
0.96428573根据自己上传的图片，利用该模型进行预测。
# 判断自己上传的照片的准确率,
def predict_picture(path):
# path 存储自己图片的文件夹.其中图片名称的第一个字符必须为图片中手势表示的数字
for p in os.listdir(path):
d_name = p[0]
pname = p.split('.')
if pname[-1] in ['jpg', 'JPG', 'png', 'PNG']:
# 图片的路径
filepath = os.path.join(path, p)
img = Image.open(filepath)
# 将分辨率调制一样
img = img.resize((100, 100), Image.ANTIALIAS)
display(img)
# 显示图片
img = np.array(img).astype('float32')
# 转换维度
img = img.transpose((2, 0, 1))
#构建预测动态图过程
with fluid.dygraph.guard(fluid.CUDAPlace(0)):
model=CNNBP()#模型实例化
model_dict,_=fluid.load_dygraph('./para/CNNBP%s' % best_model_sign)
model.load_dict(model_dict)#加载模型参数
model.eval()#评估模式
infer_img=np.array(img).astype('float32')
infer_img=infer_img[np.newaxis,:, : ,:]
infer_img = fluid.dygraph.to_variable(infer_img)
# 预测得到的softmax结果
result=model(infer_img)
print('真实数字为:', d_name, '预测的数字为', np.argmax(result.numpy()))
prob_list = list(result.numpy()[0])
# 输出各个数字的概率
prob_c_dict = {j: prob_list[j] for j in range(len(prob_list))}
sp = sorted(prob_c_dict.items(), key=lambda x: -x[1])
print('各个数字的概率', sp)
predict_picture(path='/home/aistudio/data/data29044/MyPicture')
点击获得更多项目源码。欢迎Follow，感谢Star!!! 扫描关注微信公众号pythonfan，获取更多。
