# Retraining model on new images

**URL:** <https://discuss.ai.google.dev/t/retraining-model-on-new-images/12678>\
**Category:** General Discussion\
**Tags:** education, training\
**Created:** [February 15, 2022, 6:25am UTC](https://discuss.ai.google.dev/t/retraining-model-on-new-images/12678 "2022-02-15T06:25:51Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![TensorOverflow](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/tensoroverflow/32/7323_2.png) [@TensorOverflow](https://discuss.ai.google.dev/u/TensorOverflow)\
**Post date:** [February 15, 2022, 6:25am UTC](https://discuss.ai.google.dev/t/retraining-model-on-new-images/12678/1 "2022-02-15T06:25:51Z")

</div>

I have trained a model on about 2000 images (10000 steps). At this point I have created 1500 more images that I would like to train the model on.  
Should I include these new images with the previous 2000 images and then train the model on the total of 3500 images. Or, should I train the model only on the new images? What is the best practice here in order to improve accuracy?

Thanks for any guidance!

---

<div class="post-metadata">

**Author:** ![JMG](https://avatars.discourse-cdn.com/v4/letter/j/e68b1a/32.png) [@JMG](https://discuss.ai.google.dev/u/JMG)\
**Post date:** [February 15, 2022, 1:25pm UTC](https://discuss.ai.google.dev/t/retraining-model-on-new-images/12678/3 "2022-02-15T13:25:27Z")

</div>

I believe it depends on how much different (if at all) these new images are. ML seems like a lot of try and see what works. Maybe try a mix of options and just see what works best. If you only train on the new images perhaps you might want to reduce the learning rate by some factor ( say original\_lr/10 ) so your model doesn’t forget what it learned from the first images it trained on. You should be able to test different ideas by loading the current model’s weights into a test model.

edit: fix typos

---

<div class="post-metadata">

**Author:** ![Bhack](https://d1dlmcr85iqnpo.cloudfront.net/user_avatar/discuss.ai.google.dev/bhack/32/6388_2.png) [@Bhack](https://discuss.ai.google.dev/u/Bhack)\
**Post date:** [February 15, 2022, 1:38pm UTC](https://discuss.ai.google.dev/t/retraining-model-on-new-images/12678/4 "2022-02-15T13:38:52Z")

</div>

When you are fine-tuning a network you always need to take care about how much your are changing your original weights (e.g. lr, how much layers you have unfreezed etc…) cause without any extra protection/technique catastrophic forgetting could be always behind the corner:

> **[Catastrophic interference](https://en.wikipedia.org/wiki/Catastrophic_interference)**
>
> Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to abruptly and drastically forget previously learned information upon learning new information.
> Neural networks are an important part of the connectionist approach to cognitive science. The issue of catastrophic interference when modeling human memory with connectionist models was originally brought to the attention of the scientific community by research from McCloskey and Cohen (1...

Another solution is to explore some Active learning techniques:

> **[A Survey of Deep Active Learning](https://arxiv.org/abs/2009.00236)**
>
> Active learning (AL) attempts to maximize the performance gain of the model by marking the fewest samples. Deep learning (DL) is greedy for data and requires a large amount of data supply to optimize massive parameters, so that the model learns how...