--cache_dir argument in run_lm_finetuning.py not used at all

[mpavlovic]

🐛 Bug

Model I am using (Bert, XLNet....): GPT-2

Language I am using the model on (English, Chinese....): English

The problem arise when using:

• the official example scripts: run_lm_finetuning.py

The tasks I am working on is:

• my own task or dataset: Language model finetuning on custom dataset from human resources domain

To Reproduce

Steps to reproduce the behavior:

1. Clone the repo
2. Navigate to transformers/examples directory
3. Prepare custom train and test datasets (.txt files)
4. Create ./cache directory
5. Run the following command in terminal (with replaced custom_ arguments):

python run_lm_finetuning.py \
   --output_dir=<custom_output_dir_path> \
   --model_type=gpt2 \
   --model_name_or_path=gpt2 \
   --do_train \
   --train_data_file=<custom_train_data_file> \
   --do_eval \
   --eval_data_file=<custom_eval_data_file> \
   --per_gpu_eval_batch_size=1   \
   --per_gpu_train_batch_size=1   \
   --save_total_limit=2 \
   --num_train_epochs=1 \
   --cache_dir=./cache

Expected behavior

When the model is downloaded from S3, it is stored to default cache directory in <user_home>/.cache/transformers/ directory, instead to ./cache, as specified in --cache_dir argument. Seems like --cache_dir argument isn't used in .from_pretrained() methods in lines 472, 473 and 477 in the run_lm_finetuning.py script.

Environment

• OS: Ubuntu 18.04
• Python version: 3.6.6
• PyTorch version: 1.3
• PyTorch Transformers version (or branch): 2.1.1
• Using GPU ? Yes
• Distributed of parallel setup ? No
• Any other relevant information:

Additional context