run_mlm.py : Missing key(s) in state_dict & Unexpected key(s) in state_dict

[TingNLP]

Environment info

• transformers version: 4.6.0.dev0
• Platform: Ubuntu 16.04.3 LTS
• Python version: Python 3.6.13 :: Anaconda, Inc.
• PyTorch version (GPU?): 1.8.1+cu102
• Tensorflow version (GPU?):
• Using GPU in script?: YES
• Using distributed or parallel set-up in script?: YES

Who can help

@sgugger

Information

Model I am using roberta:

The problem arises when using:

• the official example scripts: run_mlm.py

The tasks I am working on is:

• my own task or dataset: wikitext-2-raw-txt
  (https://www.salesforce.com/products/einstein/ai-research/the-wikitext-dependency-language-modeling-dataset/)

To reproduce

Steps to reproduce the behavior:

I follow the example
https://github.com/huggingface/transformers/tree/master/examples/pytorch/language-modeling

When I run

python run_mlm.py \
    --output_dir tmp/test-mlm \
    --model_name_or_path roberta-base \
    --do_train \
    --train_file wikitext-2-raw-txt/wiki.train.txt \
    --do_eval \
    --validation_file wikitext-2-raw-txt/wiki.valid.txt \
    --line_by_line

and the error occurs

2021-04-28 16:18:24.068938: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0
04/28/2021 16:18:25 - WARNING - __main__ -   Process rank: -1, device: cuda:0, n_gpu: 4distributed training: False, 16-bits training: False
04/28/2021 16:18:25 - INFO - __main__ -   Training/evaluation parameters TrainingArguments(output_dir=tmp/test-mlm, overwrite_output_dir=False, do_train=True, do_eval=True, do_predict=False, evaluation_strategy=IntervalStrategy.NO, prediction_loss_only=False, per_device_train_batch_size=8, per_device_eval_batch_size=8, gradient_accumulation_steps=1, eval_accumulation_steps=None, learning_rate=5e-05, weight_decay=0.0, adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08, max_grad_norm=1.0, num_train_epochs=3.0, max_steps=-1, lr_scheduler_type=SchedulerType.LINEAR, warmup_ratio=0.0, warmup_steps=0, logging_dir=runs/Apr28_16-18-25_Devbox4, logging_strategy=IntervalStrategy.STEPS, logging_first_step=False, logging_steps=500, save_strategy=IntervalStrategy.STEPS, save_steps=500, save_total_limit=None, no_cuda=False, seed=42, fp16=False, fp16_opt_level=O1, fp16_backend=auto, fp16_full_eval=False, local_rank=-1, tpu_num_cores=None, tpu_metrics_debug=False, debug=False, dataloader_drop_last=False, eval_steps=500, dataloader_num_workers=0, past_index=-1, run_name=tmp/test-mlm, disable_tqdm=False, remove_unused_columns=True, label_names=None, load_best_model_at_end=False, metric_for_best_model=None, greater_is_better=None, ignore_data_skip=False, sharded_ddp=[], deepspeed=None, label_smoothing_factor=0.0, adafactor=False, group_by_length=False, length_column_name=length, report_to=['tensorboard', 'wandb'], ddp_find_unused_parameters=None, dataloader_pin_memory=True, skip_memory_metrics=False, use_legacy_prediction_loop=False, push_to_hub=False, _n_gpu=4, mp_parameters=)
04/28/2021 16:18:26 - WARNING - datasets.builder -   Using custom data configuration default-b1467a68ec9fe52f
04/28/2021 16:18:27 - WARNING - datasets.builder -   Reusing dataset text (/home/A50442/.cache/huggingface/datasets/text/default-b1467a68ec9fe52f/0.0.0/e16f44aa1b321ece1f87b07977cc5d70be93d69b20486d6dacd62e12cf25c9a5)
[INFO|configuration_utils.py:498] 2021-04-28 16:18:27,029 >> loading configuration file roberta-base/config.json
[INFO|configuration_utils.py:536] 2021-04-28 16:18:27,029 >> Model config RobertaConfig {
  "architectures": [
    "RobertaForMaskedLM"
  ],
  "attention_probs_dropout_prob": 0.1,
  "bos_token_id": 0,
  "eos_token_id": 2,
  "gradient_checkpointing": false,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "initializer_range": 0.02,
  "intermediate_size": 3072,
  "layer_norm_eps": 1e-05,
  "max_position_embeddings": 514,
  "model_type": "roberta",
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "pad_token_id": 1,
  "position_embedding_type": "absolute",
  "transformers_version": "4.6.0.dev0",
  "type_vocab_size": 1,
  "use_cache": true,
  "vocab_size": 50265
}

[INFO|configuration_utils.py:498] 2021-04-28 16:18:27,030 >> loading configuration file roberta-base/config.json
[INFO|configuration_utils.py:536] 2021-04-28 16:18:27,030 >> Model config RobertaConfig {
  "architectures": [
    "RobertaForMaskedLM"
  ],
  "attention_probs_dropout_prob": 0.1,
  "bos_token_id": 0,
  "eos_token_id": 2,
  "gradient_checkpointing": false,
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "initializer_range": 0.02,
  "intermediate_size": 3072,
  "layer_norm_eps": 1e-05,
  "max_position_embeddings": 514,
  "model_type": "roberta",
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "pad_token_id": 1,
  "position_embedding_type": "absolute",
  "transformers_version": "4.6.0.dev0",
  "type_vocab_size": 1,
  "use_cache": true,
  "vocab_size": 50265
}

[INFO|tokenization_utils_base.py:1649] 2021-04-28 16:18:27,030 >> Didn't find file roberta-base/added_tokens.json. We won't load it.
[INFO|tokenization_utils_base.py:1649] 2021-04-28 16:18:27,030 >> Didn't find file roberta-base/special_tokens_map.json. We won't load it.
[INFO|tokenization_utils_base.py:1649] 2021-04-28 16:18:27,030 >> Didn't find file roberta-base/tokenizer_config.json. We won't load it.
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,030 >> loading file roberta-base/vocab.json
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,030 >> loading file roberta-base/merges.txt
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,031 >> loading file roberta-base/tokenizer.json
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,031 >> loading file None
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,031 >> loading file None
[INFO|tokenization_utils_base.py:1713] 2021-04-28 16:18:27,031 >> loading file None
[INFO|modeling_utils.py:1111] 2021-04-28 16:18:27,103 >> loading weights file roberta-base/pytorch_model.bin
[INFO|modeling_utils.py:1257] 2021-04-28 16:18:30,300 >> All model checkpoint weights were used when initializing RobertaForMaskedLM.

[INFO|modeling_utils.py:1266] 2021-04-28 16:18:30,300 >> All the weights of RobertaForMaskedLM were initialized from the model checkpoint at roberta-base.
If your task is similar to the task the model of the checkpoint was trained on, you can already use RobertaForMaskedLM for predictions without further training.
100%|██████████████████████████████████████████████████████████████████████████████████████| 37/37 [00:01<00:00, 18.82ba/s]
100%|████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 20.73ba/s]
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
[INFO|trainer.py:1027] 2021-04-28 16:18:34,809 >> Loading model from roberta-base).
Traceback (most recent call last):
  File "run_mlm.py", line 496, in <module>
    main()
  File "run_mlm.py", line 459, in main
    train_result = trainer.train(resume_from_checkpoint=checkpoint)
  File "/home/A50442/anaconda3/envs/transformer/lib/python3.6/site-packages/transformers/trainer.py", line 1046, in train
    self.model.load_state_dict(state_dict)
  File "/home/A50442/anaconda3/envs/transformer/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1224, in load_state_dict
    self.__class__.__name__, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for RobertaForMaskedLM:
	Missing key(s) in state_dict: "roberta.embeddings.position_ids", "lm_head.decoder.bias". 
	Unexpected key(s) in state_dict: "roberta.pooler.dense.weight", "roberta.pooler.dense.bias".

Expected behavior

The expected behavior is that I will get a new pretrain language model based on my dataset

[sgugger]

The command runs for me and according to your logs, the Trainer is loading a local checkpoint named roberta-base. Do you have a local folder named roberta-base? It looks like it contains a checkpoint different from the actual roberta-base model, which messes up and creates the error. Could you move that folder and try again?

[TingNLP]

@sgugger
Yes, I create a local folder named roberta-base, but the roberta-base folder contents is download from huggingface (https://huggingface.co/roberta-base/tree/main)

the language-modeling folder screenshot as shown below:

the roberta-base folder screenshot as shown below:

so i am confused...

[sgugger]

I think it's linked to the bug #11492 is fixing. Should be merged today and then you can try on a source install!