Inference

After training a model, you can use the direct predict command to perform inference.

To perform inference on a single machine run the following code block in your linux machine:

$ direct predict <output_directory> --cfg <cfg_path_or_url> --checkpoint <checkpoint_path_or_url> \
        --num-gpus <num_gpus> [--data-root <data_root>] [ --cfg <cfg_filename>.yaml --other-flags <other_flags>]

To predict using multiple machines run the following code (one command on each machine):

(machine0)$ direct predict <output_directory> --cfg <cfg_path_or_url> --checkpoint <checkpoint_path_or_url> \
        --machine-rank 0 --num-machines 2 --dist-url <URL> [--data-root <data_root>] [--other-flags]
(machine1)$ direct predict <output_directory> --cfg <cfg_path_or_url> --checkpoint <checkpoint_path_or_url> \
        --machine-rank 1 --num-machines 2 --dist-url <URL> [--data-root <data_root>] [--other-flags]

The cfg_path_or_url should point to a configuration file that includes all the model parameters used for the trained model checkpoint checkpoint_path_or_url and should also include an inference configuration as follows:

  inference:
    dataset:
      name: InferenceDataset
      <lists: ...>
      transforms:
        masking:
          ...
        ...
      text_description: <inference_description>
      ...
    batch_size: <batch_size>
...

Notes

Note that if the dataset you want to use requires downloaded path, you can pass the –data-root argument which points to the data directory.