The problem is not related just for TF packages, but seems we have a plenty of them, so I put it here. If needed, we can move it to more common repo.
While working on common preprocessing scripts for benchmarking programs (#66) I've realized that we have a lack of some general knowledge about image recognition packages. I have to use vars like CK_ENV_TENSORFLOW_MODEL_IMAGE_WIDTH and it makes these scripts not so common as they should be.
Suggestion is each image recognition package, no matter TF or Caffe or something else, should provide set for common variables, at least:
- input image size
- number of recognizable classes (it can differ even for ImageNet, e.g. 1001 for Mobilenet)
- weights file name and path (not
CK_ENV_TENSORFLOW_MODEL_WEIGHTS but something more library independent)
The problem is not related just for TF packages, but seems we have a plenty of them, so I put it here. If needed, we can move it to more common repo.
While working on common preprocessing scripts for benchmarking programs (#66) I've realized that we have a lack of some general knowledge about image recognition packages. I have to use vars like
CK_ENV_TENSORFLOW_MODEL_IMAGE_WIDTHand it makes these scripts not so common as they should be.Suggestion is each image recognition package, no matter TF or Caffe or something else, should provide set for common variables, at least:
CK_ENV_TENSORFLOW_MODEL_WEIGHTSbut something more library independent)