HubAI conversion + deprecate blobconverter + ConvertOnTrainEnd callback that exports and archives

.github/workflows/ci.yaml

@@ -166,6 +166,7 @@ jobs:
       env:
         LUXONISML_BUCKET: luxonis-test-bucket
         SUITE: ${{ matrix.suite }}
+        HUBAI_API_KEY: ${{ secrets.HUBAI_API_KEY }}
       run: pytest -x --cov --junitxml=junit.xml -o junit_family=legacy -m "${SUITE}"
 
     - name: Upload test results to Codecov

.github/workflows/tests.yaml

@@ -29,6 +29,9 @@ on:
       CODECOV_TOKEN:
         description: 'Codecov upload token'
         required: true
+      HUBAI_API_KEY:
+        description: 'HubAI API key'
+        required: true
 
 permissions:
   pull-requests: write
@@ -92,6 +95,7 @@ jobs:
       working-directory: luxonis-train
       env:
         LUXONISML_BUCKET: luxonis-test-bucket
+        HUBAI_API_KEY: ${{ secrets.HUBAI_API_KEY }}
       run: pytest --cov --junitxml=junit.xml -o junit_family=legacy
 
     - name: Upload test results to Codecov

configs/README.md

@@ -28,7 +28,8 @@ You can create your own config or use/edit one of the examples.
   - [Trainer Tips](#trainer-tips)
 - [Exporter](#exporter)
   - [`ONNX`](#onnx)
-  - [Blob](#blob)
+  - [HubAI](#hubai)
+  - [Blob (Deprecated)](#blob-deprecated)
 - [Tuner](#tuner)
   - [Storage](#storage)
 - [ENVIRON](#environ)
@@ -355,10 +356,13 @@ trainer:
         patience: 3
         monitor: "val/loss"
         mode: "min"
-    - name: "ExportOnTrainEnd"
+    - name: "ConvertOnTrainEnd"
     - name: "TestOnTrainEnd"
 ```
 
+> [!NOTE]
+> `ConvertOnTrainEnd` is the recommended callback for model conversion. It combines export, archive, and platform-specific conversion (blobconverter/HubAI SDK) into a single step. Use this instead of separate `ExportOnTrainEnd` and `ArchiveOnTrainEnd` callbacks.
+
 ### Optimizer
 
 What optimizer to use for training.
@@ -490,14 +494,15 @@ Here you can define configuration for exporting.
 | ------------------------ | --------------------------------- | ------------- | ---------------------------------------------------------------------------------------------- |
 | `name`                   | `str \| None`                     | `None`        | Name of the exported model                                                                     |
 | `input_shape`            | `list\[int\] \| None`             | `None`        | Input shape of the model. If not provided, inferred from the dataset                           |
-| `data_type`              | `Literal["INT8", "FP16", "FP32"]` | `"FP16"`      | Data type of the exported model. Only used for conversion to BLOB                              |
+| `target_precision`       | `Literal["INT8", "FP16", "FP32"]` | `"FP16"`      | Data type of the exported model. Alias: `data_type`                                            |
 | `reverse_input_channels` | `bool`                            | `True`        | Whether to reverse the image channels in the exported model. Relevant for `BLOB` export        |
 | `scale_values`           | `list[float] \| None`             | `None`        | What scale values to use for input normalization. If not provided, inferred from augmentations |
 | `mean_values`            | `list[float] \| None`             | `None`        | What mean values to use for input normalization. If not provided, inferred from augmentations  |
 | `upload_to_run`          | `bool`                            | `True`        | Whether to upload the exported files to tracked run as artifact                                |
 | `upload_url`             | `str \| None`                     | `None`        | Exported model will be uploaded to this URL if specified                                       |
 | `onnx`                   | `dict`                            | `{}`          | Options specific for ONNX export. See [ONNX](#onnx) section for details                        |
-| `blobconverter`          | `dict`                            | `{}`          | Options for converting to BLOB format. See [Blob](#blob) section for details                   |
+| `hubai`                  | `dict`                            | `{}`          | Options for HubAI SDK conversion. See [HubAI](#hubai) section for details                      |
+| `blobconverter`          | `dict`                            | `{}`          | Options for converting to BLOB format (deprecated). See [Blob](#blob-deprecated) section       |
 
 ### `ONNX`
 
@@ -510,7 +515,37 @@ Option specific for `ONNX` export.
 | `disable_onnx_simplification` | `bool`                   | `False`       | Disable ONNX simplification after export                                        |
 | `unique_onnx_initializers`    | `bool`                   | `False`       | Re-assign names to identifiers after export to ensure they are per-block unique |
 
-### `Blob`
+### `HubAI`
+
+The [HubAI SDK](https://github.com/luxonis/hubai-sdk) provides model conversion for multiple platforms (RVC2, RVC3, RVC4, Hailo).
+This is the recommended way to convert models for deployment.
+
+> [!NOTE]
+> Requires `HUBAI_API_KEY` environment variable to be set.
+
+| Key                   | Type                                       | Default value | Description                                                       |
+| --------------------- | ------------------------------------------ | ------------- | ----------------------------------------------------------------- |
+| `active`              | `bool`                                     | `False`       | Whether to use HubAI SDK for conversion                           |
+| `platform`            | `Literal["rvc2", "rvc3", "rvc4", "hailo"]` | `None`        | Target platform for conversion. Required when `active` is `True`  |
+| `delete_remote_model` | `bool`                                     | `False`       | Clean up by deleting remote uploaded variant in HubAI             |
+| `params`              | `dict`                                     | `{}`          | Additional parameters passed to the HubAI SDK conversion function |
+
+**Example:**
+
+```yaml
+exporter:
+  target_precision: fp16
+  hubai:
+    active: true
+    platform: rvc2
+    params:
+      superblob: True
+```
+
+### `Blob` (Deprecated)
+
+> [!WARNING]
+> `blobconverter` is deprecated and only supports RVC2 legacy conversion to `.blob`.
 
 | Key       | Type                                                             | Default value | Description                              |
 | --------- | ---------------------------------------------------------------- | ------------- | ---------------------------------------- |
@@ -554,6 +589,7 @@ Here you can specify options for tuning.
 > - `UploadCheckpoint`
 > - `ExportOnTrainEnd`
 > - `ArchiveOnTrainEnd`
+> - `ConvertOnTrainEnd`
 > - `TestOnTrainEnd`
 
 ### Storage
@@ -609,6 +645,7 @@ For more info on the variables, see [Credentials](../README.md#credentials).
 | `AWS_ACCESS_KEY_ID`        | `str \| None`                                              | `None`           |
 | `AWS_SECRET_ACCESS_KEY`    | `str \| None`                                              | `None`           |
 | `AWS_S3_ENDPOINT_URL`      | `str \| None`                                              | `None`           |
+| `HUBAI_API_KEY`            | `str \| None`                                              | `None`           |
 | `MLFLOW_CLOUDFLARE_ID`     | `str \| None`                                              | `None`           |
 | `MLFLOW_CLOUDFLARE_SECRET` | `str \| None`                                              | `None`           |
 | `MLFLOW_S3_BUCKET`         | `str \| None`                                              | `None`           |

configs/complex_model.yaml

@@ -116,8 +116,7 @@ trainer:
         patience: 3
         monitor: val/loss
         mode: min
-    - name: ExportOnTrainEnd
-    - name: ArchiveOnTrainEnd
+    - name: ConvertOnTrainEnd
     - name: TestOnTrainEnd
 
   optimizer:

configs/detection_heavy_model.yaml

@@ -46,8 +46,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
 
   training_strategy:
     name: TripleLRSGDStrategy

configs/detection_light_model.yaml

@@ -46,8 +46,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
 
   training_strategy:
     name: "TripleLRSGDStrategy"

configs/fomo_heavy_model.yaml

@@ -27,5 +27,4 @@ trainer:
   gradient_clip_val: 10
 
   callbacks:
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd

configs/fomo_light_model.yaml

@@ -27,5 +27,4 @@ trainer:
   gradient_clip_val: 10
 
   callbacks:
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd

configs/instance_segmentation_heavy_model.yaml

@@ -42,8 +42,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
     - name: GradientAccumulationScheduler
       params:
         # warmup phase is 3 epochs

configs/instance_segmentation_light_model.yaml

@@ -42,8 +42,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
     - name: GradientAccumulationScheduler
       params:
         scheduling: # warmup phase is 3 epochs

configs/keypoint_bbox_heavy_model.yaml

@@ -49,8 +49,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
     # For best results, always accumulate gradients to
     # effectively use 64 batch size
     - name: GradientAccumulationScheduler

configs/keypoint_bbox_light_model.yaml

@@ -49,8 +49,7 @@ trainer:
         decay: 0.9999
         use_dynamic_decay: True
         decay_tau: 2000
-    - name: ExportOnTrainEnd
-    - name: TestOnTrainEnd
+    - name: ConvertOnTrainEnd
     # For best results, always accumulate gradients to
     # effectively use 64 batch size
     - name: GradientAccumulationScheduler

configs/ocr_recognition_light_model.yaml

@@ -28,8 +28,7 @@ trainer:
   n_log_images: 8
 
   callbacks:
-    - name: TestOnTrainEnd
-    - name: ExportOnTrainEnd
+    - name: ConvertOnTrainEnd
 
   optimizer:
     name: Adam

configs/segmentation_heavy_model.yaml

@@ -24,8 +24,7 @@ trainer:
   n_log_images: 8
 
   callbacks:
-    - name: TestOnTrainEnd
-    - name: ExportOnTrainEnd
+    - name: ConvertOnTrainEnd
 
   optimizer:
     name: SGD

configs/segmentation_light_model.yaml

@@ -25,8 +25,7 @@ trainer:
   n_log_images: 8
 
   callbacks:
-    - name: TestOnTrainEnd
-    - name: ExportOnTrainEnd
+    - name: ConvertOnTrainEnd
 
   optimizer:
     name: SGD

luxonis_train/__main__.py

@@ -52,7 +52,7 @@
     from luxonis_train.utils.dataset_metadata import DatasetMetadata
 
     if weights is not None and config is None:
         ckpt = torch.load(weights, map_location="cpu")  # nosemgre
         if "config" not in ckpt:  # pragma: no cover
             raise ValueError(
                 f"Checkpoint '{weights}' does not contain the 'config' key. "
@@ -401,6 +401,36 @@
     )
 
 
+@app.command(group=export_group, sort_key=3)
+def convert(
+    opts: list[str] | None = None,
+    /,
+    *,
+    config: str | None = None,
+    save_dir: str | None = None,
+    weights: str | None = None,
+):
+    """Export, archive, and convert the model to target platform format.
+
+    This is a unified command that combines export, archive, and
+    platform conversion (RVC2/RVC3/RVC4/Hailo) steps based on the
+    configuration.
+
+    @type config: str
+    @param config: Path to the configuration file.
+    @type save_dir: str
+    @param save_dir: Directory where all outputs will be saved. If not
+        specified, the default run save directory will be used.
+    @type weights: str
+    @param weights: Path to the model weights.
+    @type opts: list[str]
+    @param opts: A list of optional CLI overrides of the config file.
+    """
+    create_model(config, opts, weights=weights).convert(
+        weights=weights, save_dir=save_dir
+    )
+
+
 @upgrade_app.command()
 def config(
     config: Annotated[

luxonis_train/callbacks/README.md

@@ -9,6 +9,7 @@ List of all supported callbacks.
 - [`PytorchLightning` Callbacks](#pytorchlightning-callbacks)
 - [`ExportOnTrainEnd`](#exportontrainend)
 - [`ArchiveOnTrainEnd`](#archiveontrainend)
+- [`ConvertOnTrainEnd`](#convertontrainend)
 - [`MetadataLogger`](#metadatalogger)
 - [`TestOnTrainEnd`](#testontrainend)
 - [`UploadCheckpoint`](#uploadcheckpoint)
@@ -51,6 +52,25 @@ Callback to create an `NN Archive` at the end of the training.
 | ---------------------- | --------------------------- | ------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `preferred_checkpoint` | `Literal["metric", "loss"]` | `"metric"`    | Which checkpoint should the callback use. If the preferred checkpoint is not available, the other option is used. If none is available, the callback is skipped |
 
+## `ConvertOnTrainEnd`
+
+Unified callback that exports, archives, and converts the archive to the target platform at the end of training. This is the recommended callback for model conversion as it combines the functionality of `ExportOnTrainEnd` and `ArchiveOnTrainEnd`, and also runs platform-specific conversions (blobconverter or HubAI SDK) if configured.
+
+**Steps:**
+
+<ol>
+  <li>Exports the model to ONNX</li>
+  <li>Creates an NN Archive from the ONNX</li>
+  <li>Runs blobconverter if `exporter.blobconverter.active` is `true`</li>
+  <li>Runs HubAI SDK conversion if `exporter.hubai.active` is `true`</li>
+</ol>
+
+**Parameters:**
+
+| Key                    | Type                        | Default value | Description                                                                                                                                                     |
+| ---------------------- | --------------------------- | ------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `preferred_checkpoint` | `Literal["metric", "loss"]` | `"metric"`    | Which checkpoint should the callback use. If the preferred checkpoint is not available, the other option is used. If none is available, the callback is skipped |
+
 ## `MetadataLogger`
 
 Callback that logs training metadata.

luxonis_train/callbacks/__init__.py

@@ -12,6 +12,7 @@
 from luxonis_train.registry import CALLBACKS
 
 from .archive_on_train_end import ArchiveOnTrainEnd
+from .convert_on_train_end import ConvertOnTrainEnd
 from .ema import EMACallback
 from .export_on_train_end import ExportOnTrainEnd
 from .gpu_stats_monitor import GPUStatsMonitor
@@ -43,11 +44,13 @@
 CALLBACKS.register(module=TrainingManager)
 CALLBACKS.register(module=GracefulInterruptCallback)
 CALLBACKS.register(module=TrainingProgressCallback)
+CALLBACKS.register(module=ConvertOnTrainEnd)
 
 
 __all__ = [
     "ArchiveOnTrainEnd",
     "BaseLuxonisProgressBar",
+    "ConvertOnTrainEnd",
     "EMACallback",
     "ExportOnTrainEnd",
     "GPUStatsMonitor",

luxonis_train/callbacks/convert_on_train_end.py

@@ -0,0 +1,28 @@
+import lightning.pytorch as pl
+from loguru import logger
+
+import luxonis_train as lxt
+
+from .needs_checkpoint import NeedsCheckpoint
+
+
+class ConvertOnTrainEnd(NeedsCheckpoint):
+    """Callback that exports, archives, and converts the model on train
+    end."""
+
+    def on_train_end(
+        self, _: pl.Trainer, pl_module: "lxt.LuxonisLightningModule"
+    ) -> None:
+        """Converts the model on train end.
+
+        @type trainer: L{pl.Trainer}
+        @param trainer: Pytorch Lightning trainer.
+        @type pl_module: L{pl.LightningModule}
+        @param pl_module: Pytorch Lightning module.
+        """
+        checkpoint = self.get_checkpoint(pl_module)
+        if checkpoint is None:  # pragma: no cover
+            logger.warning("Skipping model conversion.")
+            return
+
+        pl_module.core.convert(weights=checkpoint)