Fix mmcls bug not wrapping model in DataParallel on CPUs (#2601)

Wrap multi-label and h-label classification models by MMDataParallel in case of CPU training.
---------
Signed-off-by: Songki Choi <[email protected]>

Author: Songki Choi

CHANGELOG.md

@@ -8,7 +8,11 @@ All notable changes to this project will be documented in this file.
 
 - Update ModelAPI configuration(<https://github.com/openvinotoolkit/training_extensions/pull/2564>)
 - Add Anomaly modelAPI changes (<https://github.com/openvinotoolkit/training_extensions/pull/2563>)
+
+### Bug fixes
+
 - Fix IBLoss enablement with DeiT-Tiny when class incremental training (<https://github.com/openvinotoolkit/training_extensions/pull/2595>)
+- Fix mmcls bug not wrapping model in DataParallel on CPUs (<https://github.com/openvinotoolkit/training_extensions/pull/2601>)
 
 ## \[v1.4.3\]
 

src/otx/algorithms/classification/adapters/mmcls/task.py

@@ -381,6 +381,10 @@ def _train_model(
 
         # Model
         model = self.build_model(cfg, fp16=cfg.get("fp16", False))
+        if not torch.cuda.is_available():
+            # NOTE: mmcls does not wrap models w/ DP for CPU training not like mmdet
+            # Raw DataContainer "img_metas" is exposed, which results in errors
+            model = build_data_parallel(model, cfg, distributed=False)
         model.train()
 
         if cfg.distributed: