Beginners tensorflow, online download a lot of pictures, five categories, each category of training set has 3000 images, test set of 100, a total of 15000 training set and test set of 500,
Image preprocessing is the resize directly into 224 * 224,
Network (CNN) with five convolution layer, 3 full connection layer, the structure of the basic reference VGG simplifies,
Adjust vector several times, and so on super parameters, curves are the basic,
This is the last time curve, vector 0.01, batch of 20,
Speculation is data sets have what problem, but check the label not wrong, don't know what did not consider that
Ah, we had three days, the graphics card has been full, heartache

CodePudding user response:

Loss curve is good, but your curve fitting after a serious

CodePudding user response:

May be the characteristics of the training set can not cover the characteristics of the test set

CodePudding user response:

Data can be enhanced

CodePudding user response:

Validation set how much, or by running test set of the

Accuracy, belong to the category of usually loss curve, but can see from the curve model has not been trained to fit area, vector and falling too fast, may be too big cause model
Recommend
To increase the number of validation set 1
2 strengthen the data
3 adjustment vector to find the most suitable for decline curve
4 adjust the depth of model 15000 pictures too small batch of 20
5 don't know how to set your order to order and batch as far as possible consistent

Sorry to say a lot of hope to have help