tslgithub / image_class

基于keras集成多种图像分类模型： VGG16、VGG19、InceptionV3、Xception、MobileNet、AlexNet、LeNet、ZF_Net、ResNet18、ResNet34、ResNet50、ResNet_101、ResNet_152、DenseNet

图像分类集成以下模型：ResNet18、ResNet34、ResNet50、ResNet101、ResNet152、VGG16、VGG19、InceptionV3、Xception、MobileNet、AlexNet、LeNet、ZF_Net、DenseNet、mnist_net、TSL16，在config.py里面选择使用哪种模型.

the project apply the following models:

• VGG16
• VGG19
• InceptionV3
• Xception
• MobileNet
• AlexNet
• LeNet
• ZF_Net
• ResNet18
• ResNet34
• ResNet50
• ResNet101
• ResNet152
• DenseNet(dismissed this time)
• mnist_net
• TSL16

train or test dataset

classes name contained in folder name

"training or testing dataset folder is:"

/path/classes1/cat*.jpg,

/path/classes2/dog*.jpg,

/path/classes3/people*.jpg,

/path/classes4/*.jpg,

• Attentions ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !
• classes name must be contained in folder name

environment

My environment is based on

• ubuntu16
• cuda8 (cuda9.0)
• tensorflow_gpu1.4 (tensorflow_gpu1.10 )
• keras2.0.8
• numpy
• tqdm
• opencv-python
• scikit-learn

Install packages

• pip3 install tensorflow_gpu==1.4
• pip3 install keras==2.0.8
• pip3 install numpy
• pip3 install tqdm
• pip3 install opencv-python
• pip3 install scikit-learn

1.confirm config.py

• choose model and change parameter in config.py

2.train or test dataset prepare

• python3 mk_class_idx.py

3.train your model

• Train sigle model : python3 train.py modelName
• Train All model : run " sh trainAll.sh " to train all model (in ubuntu)
• Tensorboard : take LeNet as example, run " tensorboard --logdir=./checkpoints/LeNet " to watch training with tensorboard

4.predict your model

• predict model: python3 predict.py modelName classesName

5.result of mnist dataset(number 0~9)

Any Questions???

Author email: [email protected]

image_class's Issues

loss和val_loss不更新

你好，感谢你提供的代码，我在跑VGG16,VGG19,AlexNet时都出现了结果不更新的情况，在ResNet50上结果正常，请问我该如何解决呢？
Epoch 1/10
12992/12992 [==============================] - 108s 8ms/step - loss: 1.7006 - acc: 0.8902 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00001: val_loss improved from inf to 1.67843, saving model to ./checkpoints/VGG19/VGG19.h5
Epoch 2/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00002: val_loss did not improve from 1.67843
Epoch 3/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00003: val_loss did not improve from 1.67843
Epoch 4/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00004: val_loss did not improve from 1.67843
Epoch 5/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00005: val_loss did not improve from 1.67843
Epoch 6/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00006: val_loss did not improve from 1.67843

Epoch 00006: ReduceLROnPlateau reducing learning rate to 0.0004999999888241291.
Epoch 7/10
12992/12992 [==============================] - 98s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00007: val_loss did not improve from 1.67843
Epoch 8/10
12992/12992 [==============================] - 99s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00008: val_loss did not improve from 1.67843
Epoch 9/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00009: val_loss did not improve from 1.67843
Epoch 10/10
12992/12992 [==============================] - 100s 8ms/step - loss: 1.7046 - acc: 0.8942 - val_loss: 1.6784 - val_acc: 0.8959

Epoch 00010: val_loss did not improve from 1.67843

如何适用于多标签分类

你好，你的模型如果想在一张图片上识别多个类别该如何操作？

VGG16，显存溢出问题

config文件里面，我设置了参数batch_size = 32，normal_size = 224，然后使用VGG16之后，显存直接溢出了。训练集共3400多张图片，按说不至于吧。

我尝试了使用CPU + 内存运行，我显存8G，内存32G，想看下实际模型占用需要多少内存，结果一个batch之后，内存占用竟然达到31.1G。全部加载完图片大概是10G内存占用，也就是说中间21.1G是模型的。

我想确认下，请问up你有遇到过这个情况吗？或者有比较好的解释和解决方案吗？

训练时报错：AttributeError: 'NoneType' object has no attribute 'shape'

训练时，报下面的错误，请问怎么解决？

Traceback (most recent call last):
File "train.py", line 174, in
main()
File "train.py", line 170, in main
train.start_train()
File "train.py", line 148, in start_train
X_train, X_test, y_train, y_test=self.load_data()
File "train.py", line 55, in load_data
_, w, h = img.shape[::-1]
AttributeError: 'NoneType' object has no attribute 'shape'

训练后，测试分类准确度，一类准确度为100%，其余类准确度为0%

训练后，测试分类准确度，一类准确度为100%，其余类准确度为0%，数据集类别均衡，请问是什么原因？
训练后，val_acc也只有0.25左右

请问一下为什么要取消DenseNet网络？

你好，是什么原因要取消DenseNet网络做分类啊？

why channles=1？

Channles = 1 is a gray level image. Wouldn't it be better to use color image to extract features?

请问如何在我的数据上进行试验呢

请问我的数据应该怎么放进去呢，放进去之后读取的一直是默认类名。

请问报ValueError: './dataset/train/dog' is not in list怎么解决？谢谢！

not done?

测试结果明显不对，有解决掉的吗？过拟合太厉害了

测试结果明显不对，有解决掉的吗？过拟合太厉害了

模型

这些模型从哪里下载？没有提供下载链接？

关于损失函数与评价函数的选择是否不匹配？

您好，模型的编译阶段代码：model.compile(loss="categorical_crossentropy", optimizer=adam, metrics=["accuracy"]) 。
损失函数选取了"categorical_crossentropy"，而评价函数选取了"accuracy"，个人查看了一下keras中文文档，觉得accuracy处也应该选取“categorical_accuracy''。

因为，文章中用了与"categorical_crossentropy"配合使用的“to_categorical''函数，将每个类别对应的真值label转化成了如上图所示 ：其中向量只有一个索引为1，其余索引为0，对应一个类别。而网络预测结果最后以softmax作为输出。所以我认为应该用“categorical_accuracy''，请问您是怎么考虑的关于“accuracy‘’,现在keras中文文档没有它的官方说明......也是比较神奇

请问checkpoints找不到指定路径这个问题怎么解决呢？

运行出现checkpoints找不到指定路径这个问题，不知道怎么解决，求解答

MobileNet version

what is the MobileNet version？ MobileNetV1？ MobileNetV2？ MobileNetV3?

运行predict出现ValueError: Dimension 0 in both shapes must be equal, but are 12800 and 800. Shapes are [12800,500] and [800,500]. for 'Assign_4' (op: 'Assign') with input shapes: [12800,500], [800,500].

关于predict.py的问题

我用自己的dataset做了训练也在config和predict中做了相应的路径和class修改，但是运行predict后会提示我下面这个错误
FileNotFoundError: [Errno 2] No such file or directory: './test/abnormal '
我明明这个路径下是有文件的，一开始我以为是单一文件，但是发现你提供的试验用打dataset中也是不只一张图，我想请问我该如何解决这个问题？

checkpoints文件夹

请问这个文件夹（checkpoints）下的东西要在哪里下载呢？

checkpoint问题

您好，首先很感谢您的repo对我的学习给予帮助
但是在运行过程中遇到了checkpoint的问题，以下是运行repo源程序的train.py训练中发生的错误：
……
ls: cannot access 'checkpoints/ResNet152/events*': No such file or directory
512
rm: cannot remove 'checkpoints/ResNet152/events*': No such file or directory
……
并在上一步产生错误后，用predict.py测试结果时，发生的错误：
OSError: Unable to open file (unable to open file: name = './checkpoints/VGG16/VGG16.h5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)

keras里面没有被resnet18这个网络？该怎样得到？

用两类图片训练，测试时，一类的准确率为1.0，另一类为0，请问是什么原因？

训练时，好像只训练了第一类的图片，测试两类的准确度结果如下

是数据集制作的不正确吗？

try problem

ValueError: invalid literal for int() with base 10

如何实现feature map的可视化？

想问一下我该如何修改才能实现feature map的可视化

您好，如果我在MNIST上训练请问如何操作？

Why not identify? detect.py Which is it?

Why not identify? detect.py Which is it?
为什么没有识别呢？ detect.py 是哪个？

how to load pretrained model

VGG16

At first , thank you for providing a complete code.
When i use the VGG16 model，I got this result I think it might be wrong. But why?

90/90 [======] - 34s - loss: 7.9056 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850om inf to 8.30082, saving model to ./checkpoints/VGG16/VGG16.h5
Epoch 2/300
90/90 [======] - 24s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 3/300
90/90 [======] - 23s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 4/300
90/90 [======] - 22s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 5/300
89/90 [=====>.] - ETA: 0s - loss: 7.9866 - acc: 0.5045Epoch 00004: val_loss did not improve

90/90 [======] - 22s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850
Epoch 6/300
90/90 [======] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 7/300
90/90 [======] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 8/300
89/90 [====>.] - ETA: 0s - loss: 7.9685 - acc: 0.5056Epoch 00007: val_loss did not improve

90/90 [======] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850
Epoch 9/300
90/90 [======] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 10/300
90/90 [======] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 11/300
89/90 [====>.] - ETA: 0s - loss: 7.9866 - acc: 0.5045Epoch 00010: val_loss did not improve

90/90 [=====] - 21s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850
Epoch 12/300
90/90 [=====] - 20s - loss: 7.9785 - acc: 0.5050 - val_loss: 8.3008 - val_acc: 0.4850rove
Epoch 00011: early stopping
Done