Resnet 网络训练结果异常

使用MATLAB训练神经网络，会在最后一轮结束的时候准确率突降。学习率、训练次数都改过，问题只在Resnet上出现，同一组数据在其它网络都正常

Answer 1

[Shivansh]

Hi Jiangtao,

I understand that you have trained a ResNet model in MATLAB and you are getting worse results with it when compared to other models. Every model has their strengths and weaknesses. ResNet can give outstanding results for some cases and can perform worse sometimes. It depends on the end goal, problem statement and data. I can still list out some possible reasons for not getting optimal results using ResNet.

1. Modification in Architecture: ResNet is a deep neural network architecture that uses residual connections. It is possible that the specific architecture of ResNet is not well-suited for your dataset or task. You may want to consider trying different ResNet architecture like ResNet18, ResNet34, or WideResNet to better fit your problem.
2. Hyperparameters: Hyperparameters setting can have a huge impact on the training of a deep learning network like ResNet. You might try experimenting with hyperparameters such as learning rate, batch size, weight decay, and optimizer choice tuning to get optimal results.
3. Overfitting: ResNet is a deep neural network model and can certainly overfit and adapt to the training data instead of generalisation. Select the right amount of depth for the model with respect to the data. Check if the sudden drop in accuracy coincides with an increase in the model's validation loss. It could indicate overfitting. Regularization techniques like dropout or weight decay can help mitigate overfitting.
4. Training convergence: It is possible that the ResNet model requires longer training time or more iterations to converge compared to other networks. Ensure that the training process is carried out for an adequate number of epochs.

These can be possible reasons to explain the behaviour exhibited by the ResNet model. To analyse the issue further, it would be helpful to analyse the specific training logs, monitor the loss and accuracy curves, and compare the behaviour of the ResNet network with other networks on the same dataset.

I hope it resolves the issue!

For more information, you can refer to the documentation page for resnet 50 here https://in.mathworks.com/help/deeplearning/ref/resnet50.html.