Fine tunning and custom models with thousands of images

Hello, I have two questions:

  1. Would like to know how to fine tune the built in model.
    The custom model documentation say it’s possible, but how do we do it?
    Do we have to re train the same categories from scratch?
    I understand that deepstack used COCO dataset, so we would have to re train using those images and adding what we would like to include, for example our own CCTV images.

  2. I’m finishing tagging a model with three thousands of images. All the previous models that I’ve made were trained on Colab but they had a few hundred images and I had to play with some parameters so the training could finish before Colab disconnected me for passing max time connection. So now that I have much more data I’m thinking that using Colab would not be possible if I want to optimize reasonably well? I’m wrong about that? If so I will need to buy a GPU to train locally.

Thanks very much

1 Like

So I finished tagging the images and tried training on colab and it seems that is not possible because there seems to be a limit in how many files it can have on each folder.
If I navigate with the left side explorer it only shows to about 1100 files per folder.
And when I’m training it tells me:

AssertionError: Image Not Found /content/firedataset/Train/image_330.jpg
Even though the file was uploaded and unziped, it told me on the previous step:
inflating: /content/firedataset/Train/image_330.jpg
So that makes me think the file is actually there.

Would love to know if I’m doing something wrong, or if I would have to get a GPU in order to train locally.
On the documentation it says that we need a Nvidia GPU, I know it needs at least 4GB but I wonder if it needs to be Nvidia or it can be from other brand because I don’t have a big budget.

For this situation you might have outgrown Colab, or you just need to put your images in a bucket storage instead. You could also investigate AWS Sagemaker

Thanks, I just did a little research and I liked a lot… But the only problem that I see is it doesn’t support Deepstack models, so I would have to train using Tensorflow or other supported frameworks.
Finishing my AI specialization in coursera will look for a courses for TF or AWS.

By now it would be ideal if I could train using Deepstack, so I’m guessing local training would be the solution. For that I would need to clarify if I need a Nvidia GPU or it can be from other brands.

Deepstack training uses pytorch, which is available on AWS sagemaker. Re GPU, always NVIDIA is required, although google also have TPUs

1 Like

Thank you. I’m reading more about it and it seems this is the best way to proceed, instead of buying a GPU just to train this model.

1 Like