Skip to content

Pretrained Model

ZCls provides many pretrained model file in remote, and you can also use pths in local. There are three config items to load pretraining model:

  1. cfg.MODEL.RECOGNIZER.PRETRAINED_REMOTE
  2. cfg.MODEL.RECOGNIZER.PRETRAINED_LOCAL
  3. cfg.MODEL.RECOGNIZER.PRELOADED

PRETRAINED_REMOTE

Open the PRETRAINED_REMOTE item in config file, if zcls provides a pretrained model in remote, then will download and load it automatically

MODEL:
  ...
  ...
  RECOGNIZER:
    PRETRAINED_REMOTE: True
    PRETRAINED_NUM_CLASSES: 1000

Set the number of output categories of the pre-training model correctly. If the current number of training categories is inconsistent with the pre-training categories, the pre-training parameters of the last classification layer will not be loaded.

PRETRAINED_LOCAL

Configure the local pretraining model path in the configuration file like this:

MODEL:
  ...
  ...
  RECOGNIZER:
    PRETRAINED_LOCAL: "/path/to/pretrained"
    PRETRAINED_NUM_CLASSES: 1000

If both PRETRAINED_LOCAL and PRETRAINED_REMOTE are set, then PRETRAINED_LOCAL takes precedence.

PRELOADED

If the pre-training model contains the training-time module (use cfg.MODEL.CONV.ADD_BLOCKS) provided by zcls, the path of the pre-training model can be set in PRELOADED. 

MODEL:
  ...
  ...
  RECOGNIZER:
    PRELOADED: "/path/to/pretrained"