[Error: PyTorch to tf]convert_pytorch_checkpoint_to_tf2: AttributeError: bert.pooler.dense.weight not found in PyTorch model

[singhsidhukuldeep]

Environment info

• transformers version: 3.5.1
• Platform: Linux-5.4.0-1029-gcp-x86_64-with-glibc2.10
• Python version: 3.8.5
• PyTorch version (GPU?): 1.7.0 (False)
• Tensorflow version (GPU?): 2.3.1 (False)
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help

I think: @patrickvonplaten @LysandreJik @VictorSanh

Anyone is welcome!

Information

Trying to convert my pytorch checkpoint to tf using the below code:

from transformers import convert_pytorch_checkpoint_to_tf2
convert_pytorch_checkpoint_to_tf2.convert_pt_checkpoint_to_tf(
    model_type = "bert", 
    pytorch_checkpoint_path="model/pytorch_model.bin", 
    config_file="model/config.json", 
    tf_dump_path="TFmodel", 
    compare_with_pt_model=False, 
    use_cached_models=False
)

my model folder has a tiny bert trained using HuggingFace: contents of model folder are: checkpoint-500 special_tokens_map.json config.json tokenizer_config.json eval_results_mlm_wwm.txt training_args.bin pytorch_model.bin vocab.txt

To reproduce

Error:

Loading PyTorch weights from /home/3551351/bert-mlm/model/pytorch_model.bin
PyTorch checkpoint contains 8,354,548 parameters
Traceback (most recent call last):
  File "pt2tf.py", line 7, in <module>
    convert_pytorch_checkpoint_to_tf2.convert_pt_checkpoint_to_tf(
  File "/home/3551351/.conda/envs/kuldeepVenv/lib/python3.8/site-packages/transformers/convert_pytorch_checkpoint_to_tf2.py", line 283, in convert_pt_checkpoint_to_tf
    tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
  File "/home/3551351/.conda/envs/kuldeepVenv/lib/python3.8/site-packages/transformers/modeling_tf_pytorch_utils.py", line 96, in load_pytorch_checkpoint_in_tf2_model
    return load_pytorch_weights_in_tf2_model(
  File "/home/3551351/.conda/envs/kuldeepVenv/lib/python3.8/site-packages/transformers/modeling_tf_pytorch_utils.py", line 172, in load_pytorch_weights_in_tf2_model
    raise AttributeError("{} not found in PyTorch model".format(name))
AttributeError: bert.pooler.dense.weight not found in PyTorch model

Expected behavior

[patrickvonplaten]

Hey @singhsidhukuldeep,

Could you maybe upload your weights to a colab so that I can reproduce the error or upload your weights to the model hub and give me a path to it?

This way, I can reproduce the error and solve it :-)

Thanks a lot!

[singhsidhukuldeep]

Hi @patrickvonplaten ,

I tried this

pt_model = TFBertForPreTraining.from_pretrained(model_output_location, from_pt=True)

print("\n\n>>> Saving HuggingFace to tensorflow(pb)")
tf.saved_model.save(pt_model,TF_model_output_location)

and it worked, but I am not able to understand the limitations here!

[patrickvonplaten]

If you want to convert from TF to PT PT to TF this is exactly how you should to it...

[singhsidhukuldeep]

@patrickvonplaten I am looking to convert PyTorch to TF!

[patrickvonplaten]

Sorry I meant PT to TF -> your approach is correct here.

[singhsidhukuldeep]

Got it! Thanks for the help.

One last thing, this gives a *.h5 file (weights only) Is there a way to get *.pb file with structure and weights?

[pranavpawar3]

Hi @singhsidhukuldeep, I tried this, it did save the model in assets, variables, and saved_model.pb format, but couldn't get any .h5 file that I need, am I missing something?

P.S. I am trying to convert a standard config BERT MaskedLM model

Hi @patrickvonplaten ,

I tried this

pt_model = TFBertForPreTraining.from_pretrained(model_output_location, from_pt=True)

print("\n\n>>> Saving HuggingFace to tensorflow(pb)")
tf.saved_model.save(pt_model,TF_model_output_location)

and it worked, but I am not able to understand the limitations here!

[github-actions[bot]]

This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.

If you think this still needs to be addressed please comment on this thread.