# Custom, without any pretraining, training with PyTorch

**URL:** <https://discuss.huggingface.co/t/custom-without-any-pretraining-training-with-pytorch/30752>\
**Category:** Beginners\
**Created:** [January 30, 2023, 1:45pm UTC](https://discuss.huggingface.co/t/custom-without-any-pretraining-training-with-pytorch/30752 "2023-01-30T13:45:41Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Ofek](https://avatars.discourse-cdn.com/v4/letter/o/e0b2c6/32.png) [@Ofek](https://discuss.huggingface.co/u/Ofek)\
**Post date:** [January 30, 2023, 1:45pm UTC](https://discuss.huggingface.co/t/custom-without-any-pretraining-training-with-pytorch/30752/1 "2023-01-30T13:45:41Z")

</div>

Hey!  
The tokenizer options in HuggingFace are extremely useful and easy to work with.  
I’d like to build a custom pipeline that:

1. inputs custom data ,not something in Datasets (I know how to do this in HF).
2. tokenizes it with a HF tokenizer pipeline (this I know how to do, too).
3. uses the above to feed into a custom (no pretraining, not a specific architecture) PyTorch model,
4. optimizes with any viable PyTorch loss function.

So basically, I want to work within the HF framework for the text’s preprocessing and tokenizing, while still doing whatever I wish in PyTorch afterwards…

Is this feasible?

Thanks!