Proper installation and use

[vgaraujov]

Hello @adarob

I want to train a T5 model from scratch in a new language, but I am new to this framework. I went through the documentation, issues, and discussions; however, I have some doubts about the proper installation and use of the library.

1. I got a TPU with tpu_vm_base version, and then I followed these steps: https://github.com/google-research/t5x/tree/main#installation. Am I missing some dependencies to install (e.g., do I have to install also t5 library)? Is there a requirements.txt there to install the proper library versions?

2. I read in this issue Pretraining from scratch #24 that seqio Tasks are used to set up the data pipelines. I understood that I have to create a new TaskRegistry and MixtureRegistry for my datasets. However, it is not clear to me where I have to put this code. Does it go in the gin file?

3. I already have my own SentencePiece tokenizer. Is it enough to change the name or path of the tokenizer in this gin file https://github.com/google-research/t5x/blob/main/t5x/examples/t5/t5_1_1/base.gin?

I would really appreciate your help in this matter. Thanks for your contribution to this library to the community.