ConfigurationΒΆ

TensorVision comes with reasonable defaults. You only need to read this if you want tweak it to your needs.

TensorVision is configured with environment variables. It is quite easy to set them yourself (see multiple ways). The supported variables are:

  • TV_DIR_DATA: The default directory where to look for data.
  • TV_DIR_RUNS: The default directory where to look for the model.
  • TV_IS_DEV: Either 0 or 1 - set if you want to see debug messages.
  • TV_PLUGIN_DIR: Directory with Python scripts which will be loaded by utils.py
  • TV_SAVE: Whether to keep all runs on default. By default runs will be written to TV_DIR_RUNS/debug and overwritten by newer runs, unless tv-train --save is called.
  • TV_USE_GPUS: Controll which gpus to use. Default all GPUs are used, GPUs can be specified using --gpus. Setting TV_USE_GPUS='force' makes the flag --gpus compulsory, this is useful in cluster environoments. Use TV_USE_GPUS='0,3' to run Tensorflow an the GPUs with ids 0 and 3.
  • TV_STEP_SHOW: After how many epochs of training should the TV_STEP_STR be printed?
  • TV_STEP_EVAL: After how many epochs of training evaluation is done.
  • TV_STEP_WRITE: After how many epochs of training checkpoints are written to disk.
  • TV_MAX_KEEP: How many checkpoints to keep.
  • TV_STEP_STR: Set what you want to see each 100th step of the training. The default is
Step {step}/{total_steps}: loss = {loss_value:.2f}
( {sec_per_batch:.3f} sec (per Batch);
{examples_per_sec:.1f} examples/sec;)