InferenceModel class do not have from_pretrained()

[akira-sasaki]

I build cogview2 with colab. And run your script, I have this error.

InferenceModel class do not have from_pretrained().

!./text2image.sh --input-source input.txt

WARNING: No training data specified
using world size: 1 and model-parallel size: 1 
Traceback (most recent call last):
  File "cogview2_text2image.py", line 233, in <module>
    main(args)
  File "cogview2_text2image.py", line 48, in main
    model, args = InferenceModel.from_pretrained(args, 'coglm')
AttributeError: type object 'InferenceModel' has no attribute 'from_pretrained'

Are there any omissions in the committed sources?

[Sleepychord]

This is strange, InferenceModel should have this method by Inheriting. Could you share your version of Python and SwissArmyTransformer(you can upgrade it to 0.2.3 and try again)?

[akira-sasaki]

I used old version SwissArmyTransformer==0.1.2. I update to 0.2.3 and try again.

[akira-sasaki]

I had error, again

Traceback (most recent call last):
  File "cogview2_text2image.py", line 19, in <module>
    from SwissArmyTransformer import get_args, get_tokenizer
  File "/usr/local/lib/python3.7/dist-packages/SwissArmyTransformer/__init__.py", line 4, in <module>
    from .model import AutoModel
  File "/usr/local/lib/python3.7/dist-packages/SwissArmyTransformer/model/__init__.py", line 3, in <module>
    from .encoder_decoder_model import EncoderDecoderModel
  File "/usr/local/lib/python3.7/dist-packages/SwissArmyTransformer/model/encoder_decoder_model.py", line 87
    return encoder_outputs, decoder_outputs, *mems
                                             ^
SyntaxError: invalid syntax

I use python 3.7, which version of Python should I use?　

[Sleepychord]

according to this issue #4 , maybe you should upgrade it to python 3.8.

[metaphorz]

Can you share the Colab notebook? I can periodically try it and pray that I get an A100.

[chenxwh]

@metaphorz Here is Replicate demo that is backed with A100 if you are interested :) https://replicate.com/thudm/cogview2

[metaphorz]

@chenxwh thank you I will try

[akira-sasaki]

@metaphorz I shared colab working. https://github.com/akira-sasaki/playground/blob/main/CogView2.ipynb

The GPU memory is 16GB at ColabPro+. So, we need --only-first-stage option.

If you want to make it all work, you might need an A100.

[metaphorz]

@akira-sasaki thanks very much.