Python 3.5 + Torch 1.0 does not work

[yuhui-zh15]

When running run_lm_finetuning.py to fine-tune language model with default settings (see command below), sometimes I could run successfully, but sometimes I received different errors like RuntimeError: The size of tensor a must match the size of tensor b at non-singleton dimension 1, RuntimeError: Creating MTGP constants failed. at /pytorch/aten/src/THC/THCTensorRandom.cu:35 or RuntimeError: Dimension out of range (expected to be in range of [-1, 0], but got 1). This problem can be solved when updating python3.5 to python3.6.

python run_lm_finetuning.py \
  --bert_model ~/bert/models/bert-base-uncased/ \
  --do_train  \
  --train_file ~/bert/codes/samples/sample_text.txt  \
  --output_dir ~/bert/exp/lm \
  --num_train_epochs 5.0  \
  --learning_rate 3e-5  \
  --train_batch_size 32  \
  --max_seq_length 128 \
  --on_memory

[MuruganR96]

Thank you @yuhui-zh15 sir. i will check.

[thomwolf]

This should be fixed on master now (thanks to #191 )