tensorflow cnn

[AnonymousMan]

Hii, I am new deep learning. 

I have one doubt.

Even, when my model get 90 percent accuracy on validation set during training, my model doesn't perform well on the data I feed. Why?

I try to make prediction on images that are present on the internet.

[acqmal_latief]

i think your model overfiting, becuase your model can't predict new data image

[lsolatorio · Accepted solution]

Hi @AnonymousMan, 

Thank you for joining our community.

I understand your frustration that your model isn't performing as well as expected on new data, even with high validation accuracy.

There's a number of reasons why your model is behaving like this.

• Like what @acqmal_latief mentioned, overfitting is a possible scenario, this happens when your model memorized certain patterns in your training data well enough resulting to poor performance especially on unseen data. Try increasing your training data, the more diverse, the less likely it is to overfit.
• It is also possible that the validation data might not accurately represent the new data you're feeding the model. There could be subtle differences in distribution, noise levels, or data formats. Try curating your validation set to closely resemble the data you expect the model to encounter in practice.
• Consider cleaning the new data you're feeding the model. Eliminate, if not minimize the noise or outliers that might be affecting the model's performance. If the new data distribution is significantly different, consider retraining the model with a combined dataset that better reflects the real-world scenario.

You can look into TensorFlow's Performance best practices and Best practices for implementing machine learning on Google Cloud for more information.

I hope I was able to provide you with useful insights.