# Bypassing tokenizers

**URL:** <https://discuss.huggingface.co/t/bypassing-tokenizers/2162>\
**Category:** 🤗Tokenizers\
**Created:** [November 22, 2020, 11:08am UTC](https://discuss.huggingface.co/t/bypassing-tokenizers/2162 "2020-11-22T11:08:23Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![thenoone](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/thenoone/32/626_2.png) [@thenoone](https://discuss.huggingface.co/u/thenoone)\
**Post date:** [November 22, 2020, 11:08am UTC](https://discuss.huggingface.co/t/bypassing-tokenizers/2162/1 "2020-11-22T11:08:23Z")

</div>

Hi everyone,  
Is it possible to bypass the tokenizer and directly provide the input embeddings to train the BERT model?  
Thanks!

---

<div class="post-metadata">

**Author:** ![BramVanroy](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/bramvanroy/32/47360_2.png) [@BramVanroy](https://discuss.huggingface.co/u/BramVanroy)\
**Post date:** [November 22, 2020, 4:51pm UTC](https://discuss.huggingface.co/t/bypassing-tokenizers/2162/2 "2020-11-22T16:51:35Z")

</div>

You can just create a subclass of the model that you want and modify its forward pass.

---

<div class="post-metadata">

**Author:** ![sgugger](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/sgugger/32/2291_2.png) [@sgugger](https://discuss.huggingface.co/u/sgugger)\
**Post date:** [November 23, 2020, 1:49pm UTC](https://discuss.huggingface.co/t/bypassing-tokenizers/2162/3 "2020-11-23T13:49:30Z")

</div>

You can also feed `input_embeds` instead of `input_ids` to your mdel.