vid2seq installation error: No matching distribution found for t5x

[uzl]

I was trying to install vid2seq.

I created a fresh conda virtual environment with Python=3.9. Then, according to documentation when I am running this following command

pip install -r scenic/projects/vid2seq/requirements.txt

It throws this error

Collecting dmvr@ git+https://github.com/deepmind/dmvr.git
  Cloning https://github.com/deepmind/dmvr.git to /tmp/pip-install-g5bw7bnk/dmvr_2884f2053dd747c6b11a633bad9e9ef8
  Running command git clone --filter=blob:none --quiet https://github.com/deepmind/dmvr.git /tmp/pip-install-g5bw7bnk/dmvr_2884f2053dd747c6b11a633bad9e9ef8
  Resolved https://github.com/deepmind/dmvr.git to commit 77ccedaa084d29239eaeafddb0b2e83843b613a1
  Preparing metadata (setup.py) ... done
Collecting gin-config
  Using cached gin_config-0.5.0-py3-none-any.whl (61 kB)
Collecting t5
  Using cached t5-0.9.3-py3-none-any.whl (153 kB)
ERROR: Could not find a version that satisfies the requirement t5x (from versions: 0.0.0)
ERROR: No matching distribution found for t5x

Am I missing something, or is the requirements.txt incorrect?

[antoyang]

It seems the t5x installation fails. You may try to follow the t5x installation instructions from https://github.com/google-research/t5x.

[fake-warrior8]

It seems the t5x installation fails. You may try to follow the t5x installation instructions from https://github.com/google-research/t5x.

The t5x gpu is so hard to use. After pip install -r pile_requirements.txt, there is still no t5x in the conda environment. Is there anything wrong?

One more thing, Could you give the T5.1.1 Base model link directly? I cannot found the link in the t5x Readme.

[jena-shreyas]

The t5x gpu is so hard to use. After pip install -r pile_requirements.txt, there is still no t5x in the conda environment. Is there anything wrong?

I was able to install t5x using the following steps :

git clone https://github.com/google-research/t5x.git  # Clone repo if not done already
cd t5x
python3 -m pip install -e .[gpu]

[mekasone]

The t5x gpu is so hard to use. After pip install -r pile_requirements.txt, there is still no t5x in the conda environment. Is there anything wrong?

I was able to install t5x using the following steps :

git clone https://github.com/google-research/t5x.git  # Clone repo if not done already
cd t5x
python3 -m pip install -e .[gpu]

Above worked for me on Mac. I still ran 'pip install t5x' (no errors now). thx

[toshohirasawa]

@antoyang Could you let me know which revision you used to run the vid2seq model?

[antoyang]

@toshohirasawa I used Google's internal environment which I cannot access now as this was an internship work, so I am not sure which t5x version was used. If you are struggling with the installation, you may have a look at the updated PyTorch implementation (with a few differences explained in the readme) here: https://github.com/antoyang/VidChapters.