I’m now working on a project involving image classification using Convolutional Neural Networks (CNNs), and I want some help with implementation. I’m facing challenges with fine-tuning pre-trained models for a new dataset and optimizing hyperparameters for improved performance. Convolutional Neural Network (CNN) | The Batch.
I’ve checked the relevant literature and documentation, and I’ve tried several approaches to address the issue, including experimenting with different architectures, adjusting learning rates, and implementing data augmentation techniques. Even after this i haven’t been able to achieve the desired results in terms of classification accuracy and generalization to unseen data.
If anyone has faced similar challenges in their projects, I would greatly appreciate any suggestions for overcoming this issue. I’m interested in learning about best practices for fine-tuning pre-trained models, optimizing hyperparameters, and addressing overfitting in CNNs.
Your support would be immensely valuable in helping me progress with my project.
@stevediaz have you taken any of the classes here yet on Conv nets in that regard ? And also, can you provide a more detailed description of your data / classification set so that we can get a better sense of what you are dealing with / trying to accomplish ?
Yes, as Anthony mentions, there is a lot of relevant information in DLS C4, which is all about ConvNets. There is a section in Week 2 about Transfer Learning which discusses a range of techniques. It sounds like you are probably already aware of a lot of them, but it might be worth taking a look at those lectures just in case. You can watch them at no charge by taking the course in “audit” mode. You won’t be able to see the corresponding programming assignment, but you can view the lectures and see if they give you any additional insight.
There is also this article by François Chollet about Fine Tuning in the TF/Keras documentation site. Worth a look if you haven’t seen it yet. In particular, he demonstrates how to unfreeze and do additional training on some of the layers of the pretrained network.
But if you want help beyond that, we’ll need a lot more information to offer anything more concrete and specific than the above, as Tom says. What pretrained models have you tried? How big is your dataset? What kind of classification are you trying to do? What are the best case accuracy results you are currently getting and what are your real goals for accuracy? And more along that line …