API

model_hub.mmdetection

class model_hub.mmdetection.MMDetTrial(context: determined.pytorch._pytorch_context.PyTorchTrialContext)

This trial serves as the trainer for MMDetection models. It replaces the mmcv runner used by MMDetection.

For nearly all use cases, you can just use this trial definition and control behavior by changing the MMDetection config. If you want to customize the trial further, you can use this trial as the starting point.

Simlar to using the MMDetection library directly, the main way users customize an experiment is by modifying the MMDetection config. We detail how to configure MMDetection through the Determined experiment configuration in the readme.

Helper Functions

class model_hub.mmdetection.GCSBackend(*args: Any, **kwargs: Any)

To use a Google Storage bucket as the storage backend, set data.file_client_args field of the experiment config as follows:

data:
  file_client_args:
    backend: gcs
    bucket_name: <FILL IN>

class model_hub.mmdetection.S3Backend(*args: Any, **kwargs: Any)

To use a S3 bucket as the storage backend, set data.file_client_args field of the experiment config as follows:

data:
  file_client_args:
    backend: s3
    bucket_name: <FILL IN>

model_hub.mmdetection.get_pretrained_ckpt_path(download_directory: str, config_file: str) → Tuple[Any, Any]

If the config_file has an associated pretrained checkpoint, return path to downloaded checkpoint and preloaded checkpoint

Parameters

• download_directory – path to download checkpoints to

• config_file – mmdet config file path for which to find and load pretrained weights

Returns

checkpoint path, loaded checkpoint