position_idx missing in state_dict when loading from checkpoint

[gabriele-bani]

When usingExtractiveSummarizer.load_from_checkpointor ExtractiveSummarizer.load_weights to load most of the models, i find that the position_ids field is not saved in the checkpoint file, which causes an error. The only model that can be correctly loaded is distilbert-base-uncased-ext-sum

This is the error I get when running predictions_website.py and trying to use bert-base-uncased-ext-sum checkpoints.

Exception happened during processing of request from ('127.0.0.1', 34450)
Traceback (most recent call last):
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/socketserver.py", line 320, in _handle_request_noblock
    self.process_request(request, client_address)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/socketserver.py", line 351, in process_request
    self.finish_request(request, client_address)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/socketserver.py", line 364, in finish_request
    self.RequestHandlerClass(request, client_address, self)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/socketserver.py", line 724, in __init__
    self.handle()
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/http/server.py", line 418, in handle
    self.handle_one_request()
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/http/server.py", line 406, in handle_one_request
    method()
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/site-packages/gradio/networking.py", line 158, in do_POST
    prediction, durations = interface.process(raw_input)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/site-packages/gradio/interface.py", line 220, in process
    prediction = predict_fn(*processed_input)
  File "/home/myuser/intrical/repos/Intrical-Transformers/predictions_website.py", line 11, in summarize_text
    summarizer = ExtractiveSummarizer.load_from_checkpoint(model_choice)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/site-packages/pytorch_lightning/core/saving.py", line 153, in load_from_checkpoint
    model = cls._load_model_state(checkpoint, *args, strict=strict, **kwargs)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/site-packages/pytorch_lightning/core/saving.py", line 192, in _load_model_state
    model.load_state_dict(checkpoint['state_dict'], strict=strict)
  File "/home/myuser/anaconda3/envs/transformersum/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1045, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for ExtractiveSummarizer:
	Missing key(s) in state_dict: "word_embedding_model.embeddings.position_ids". 
----------------------------------------

How can I correctly load any type of model?

[HHousen]

@gabriele-bani Good catch. This is a problem with the update from 3.0.2 to 3.1.0 of huggingface/transformers. The problem is being discussed at huggingface/transformers#6882. While they work on a fix, you can install the previous version of huggingface/transformers by running pip install -U transformers==3.0.2. Let me know if this solves the problem for you.

[gabriele-bani]

@HHousen Thank you for your answer! Indeed, 3.0.2 allows to load the checkpoints correctly, so I'll be using that for now.