Can't train without --use-cpu

Hi, when launching my training on a GPU, I get:

[ERROR:sockeye.train] Mismatch in arguments for training continuation.
[ERROR:sockeye.train] Differing arguments: use_cpu.

The command is

python3 -m sockeye.train \
  --batch-size 80 \
  --batch-type sentence \
  --checkpoint-frequency 30000 \
  --decode-and-evaluate 100 \
  --decoder rnn \
  --embed-dropout 0.2 \
  --encoder rnn \
  --initial-learning-rate 0.0001 \
  --keep-last-params 4 \
  --learning-rate-reduce-factor 0.7 \
  --learning-rate-reduce-num-not-improved 2 \
  --max-num-checkpoint-not-improved 4 \
  --max-seq-len 100 \
  --num-embed 512:512 \
  --num-layers 8:8 \
  --optimized-metric bleu \
  --optimizer adam \
  --rnn-attention-in-upper-layers \
  --rnn-attention-type bilinear \
  --rnn-decoder-hidden-dropout 0.2 \
  --rnn-num-hidden 1024 \
  --use-cpu \
  --weight-init xavier \
  --weight-init-scale 3.0 \
  --weight-init-xavier-factor-type avg \
  -d train_data \
  -o sockeye-commit-suggester \
  -vs test.3000.bpe.diff \
  -vt test.3000.bpe.msg

What am I doing wrong? Many thanks to everybody

[tdomhan]

True. Maybe we could additionally extend the Mismatch in arguments for training continuation. message though to tell the user that the existing directory should be deleted if training continuation is not intended.

[tdomhan]

When you run Sockeye multiple times with the same arguments then it will try to resume training. Do you think a message that just prints something like:

Trying to resume training from existing directory %s.

would make this clearer?