在这一项目中,你需要对 CIFAR-10 数据集 进行图像分类。
该数据集包含飞机,狗,猫等一共10个类型的图片,共50000张。你需要先对数据集进行预处理,然后用卷积神经网络对所有样本进行训练。你要先将图像归一化(normalize),对标签进行独热编码(one-hot encode the labels),然后建立卷积层、最大池化层和全连接层。最后,你将看到该模型对样本图像作出的预测结果。
-
MNIST
Model Epochs Accuracy 传统神经网络 FC1024
,FC1024
50 98% 卷积神经网络 类VGG结构 50 99.5% -
CIFAR10
Model Epochs Accuracy TensorFlow CONV16
,CONV32
,CONV64
,FC128
100 70% TensorFlow CONV16
,CONV32
,CONV64
,FC128
(Xavier, BN, L2)100 72% Keras 类VGG (CONV / CONV / POOL) * 3
,FC128
100 80% Keras 类VGG (CONV / CONV / POOL) * 3
,FC128
(With Data Augmentation)100 89%
-
Preprocessing
- Normalize
- One-hot Encoding
-
Model Construction
- Input Layer
- Convolution + Maxpooling Layer
- Flatten Layer
- Fully Connected Layer
- Dropout
- Output Layer
-
Parameter Fine Tuning
- epoch
- batch_size
- stddev
- conv_num_output
- conv_size (width x height)
- conv_stride
- pooling_size (width x height)
- pooling_stride
- full_num_output
- Activation Function (Conv Layer, Fully Connected Layer)
- Padding Style