Loading a model saved with TFGPT2LMHeadModel.save_pretrained with GPT2LMHeadModel.from_pretrained(..., from_tf=True)

[JulesGM]

Environment info

• transformers version: 4.5.0
• Platform: Linux-4.19.0-16-cloud-amd64-x86_64-with-glibc2.10
• Python version: 3.8.8
• PyTorch version (GPU?): 1.8.1+cu102 (False)
• Tensorflow version (GPU?): 2.4.1 (False)
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Who can help

@patrickvonplaten, @LysandreJik

Information

Hello,

(My problem seems related to #5588)

I fine-tuned a TFGPT2LMHeadModel and saved it with .save_pretrained, giving me a tf_model.h5 and a config.json files.
I try loading it with

model = transformers.GPT2LMHeadModel.from_pretrained(
    ".", from_tf=True, config="./config.json"
)
```.
The path is fine.
I get the following messages:

All TF 2.0 model weights were used when initializing GPT2LMHeadModel.

Some weights of GPT2LMHeadModel were not initialized from the TF 2.0 model and are newly initialized: ['transformer.h.0.attn.bias', 'transformer.h.0.attn.masked_bias', 'transformer.h.1.attn.bias', 'transformer.h.1.attn.masked_bias', 'transformer.h.2.attn.bias', 'transformer.h.2.attn.masked_bias', 'transformer.h.3.attn.bias', 'transformer.h.3.attn.masked_bias', 'transformer.h.4.attn.bias', 'transformer.h.4.attn.masked_bias', 'transformer.h.5.attn.bias', 'transformer.h.5.attn.masked_bias', 'transformer.h.6.attn.bias', 'transformer.h.6.attn.masked_bias', 'transformer.h.7.attn.bias', 'transformer.h.7.attn.masked_bias', 'transformer.h.8.attn.bias', 'transformer.h.8.attn.masked_bias', 'transformer.h.9.attn.bias', 'transformer.h.9.attn.masked_bias', 'transformer.h.10.attn.bias', 'transformer.h.10.attn.masked_bias', 'transformer.h.11.attn.bias', 'transformer.h.11.attn.masked_bias', 'transformer.h.12.attn.bias', 'transformer.h.12.attn.masked_bias', 'transformer.h.13.attn.bias', 'transformer.h.13.attn.masked_bias', 'transformer.h.14.attn.bias', 'transformer.h.14.attn.masked_bias', 'transformer.h.15.attn.bias', 'transformer.h.15.attn.masked_bias', 'transformer.h.16.attn.bias', 'transformer.h.16.attn.masked_bias', 'transformer.h.17.attn.bias', 'transformer.h.17.attn.masked_bias', 'transformer.h.18.attn.bias', 'transformer.h.18.attn.masked_bias', 'transformer.h.19.attn.bias', 'transformer.h.19.attn.masked_bias', 'transformer.h.20.attn.bias', 'transformer.h.20.attn.masked_bias', 'transformer.h.21.attn.bias', 'transformer.h.21.attn.masked_bias', 'transformer.h.22.attn.bias', 'transformer.h.22.attn.masked_bias', 'transformer.h.23.attn.bias', 'transformer.h.23.attn.masked_bias', 'transformer.h.24.attn.bias', 'transformer.h.24.attn.masked_bias', 'transformer.h.25.attn.bias', 'transformer.h.25.attn.masked_bias', 'transformer.h.26.attn.bias', 'transformer.h.26.attn.masked_bias', 'transformer.h.27.attn.bias', 'transformer.h.27.attn.masked_bias', 'transformer.h.28.attn.bias', 'transformer.h.28.attn.masked_bias', 'transformer.h.29.attn.bias', 'transformer.h.29.attn.masked_bias', 'transformer.h.30.attn.bias', 'transformer.h.30.attn.masked_bias', 'transformer.h.31.attn.bias', 'transformer.h.31.attn.masked_bias', 'transformer.h.32.attn.bias', 'transformer.h.32.attn.masked_bias', 'transformer.h.33.attn.bias', 'transformer.h.33.attn.masked_bias', 'transformer.h.34.attn.bias', 'transformer.h.34.attn.masked_bias', 'transformer.h.35.attn.bias', 'transformer.h.35.attn.masked_bias', 'transformer.h.36.attn.bias', 'transformer.h.36.attn.masked_bias', 'transformer.h.37.attn.bias', 'transformer.h.37.attn.masked_bias', 'transformer.h.38.attn.bias', 'transformer.h.38.attn.masked_bias', 'transformer.h.39.attn.bias', 'transformer.h.39.attn.masked_bias', 'transformer.h.40.attn.bias', 'transformer.h.40.attn.masked_bias', 'transformer.h.41.attn.bias', 'transformer.h.41.attn.masked_bias', 'transformer.h.42.attn.bias', 'transformer.h.42.attn.masked_bias', 'transformer.h.43.attn.bias', 'transformer.h.43.attn.masked_bias', 'transformer.h.44.attn.bias', 'transformer.h.44.attn.masked_bias', 'transformer.h.45.attn.bias', 'transformer.h.45.attn.masked_bias', 'transformer.h.46.attn.bias', 'transformer.h.46.attn.masked_bias', 'transformer.h.47.attn.bias', 'transformer.h.47.attn.masked_bias', 'lm_head.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

It means that the conversion hasn't worked, right? 
Can I just use the model for generation?  Should I change the way the model is saved ?

[LysandreJik]

Actually that's not an issue, this warning shouldn't be here. I'll open a PR to remove it shortly.

[LysandreJik]

If you try generating text with it, you should get sensible results!

[JulesGM]

Great to hear, thanks.