[OTX] consider multi GPU when setting drop last (#1520)

* when set drop_last as true, consider mutli gpu

* bugfix and revert TC batch size

* change num_gpus to num_worlds

otx/algorithms/classification/tasks/inference.py

@@ -8,6 +8,7 @@
 from typing import Optional
 
 import numpy as np
+import torch.distributed as dist
 from mmcv.utils import ConfigDict
 
 from otx.algorithms.classification.configs import ClassificationConfig
@@ -454,7 +455,11 @@ def patch_color_conversion(pipeline):
             # In train dataset, when sample size is smaller than batch size
             if subset == "train" and self._data_cfg:
                 train_data_cfg = Stage.get_data_cfg(self._data_cfg, "train")
-                if len(train_data_cfg.get("otx_dataset", [])) < self._recipe_cfg.data.get("samples_per_gpu", 2):
+                num_worlds = dist.get_world_size() if dist.is_initialized() else 1
+                if (
+                    len(train_data_cfg.get("otx_dataset", []))
+                    < self._recipe_cfg.data.get("samples_per_gpu", 2) * num_worlds
+                ):
                     cfg.drop_last = False
 
             cfg.domain = domain

otx/mpa/cls/trainer.py

@@ -8,6 +8,7 @@
 import warnings
 
 import mmcv
+import torch.distributed as dist
 from mmcls import __version__
 from mmcls.core import DistOptimizerHook
 from mmcls.datasets import build_dataloader, build_dataset
@@ -105,7 +106,11 @@ def run(self, model_cfg, model_ckpt, data_cfg, **kwargs):
         drop_last = False
         dataset_len = len(otx_dataset) if otx_dataset else 0
         # if task == h-label & dataset size is bigger than batch size
-        if train_data_cfg.get("hierarchical_info", None) and dataset_len > cfg.data.get("samples_per_gpu", 2):
+        num_worlds = dist.get_world_size() if self.distributed else 1
+        if (
+            train_data_cfg.get("hierarchical_info", None)
+            and dataset_len > cfg.data.get("samples_per_gpu", 2) * num_worlds
+        ):
             drop_last = True
         # updated to adapt list of dataset for the 'train'
         data_loaders = [

tests/integration/cli/classification/test_classification.py

@@ -464,7 +464,7 @@ def test_otx_multi_gpu_train(self, template, tmp_dir_path):
         "--learning_parameters.num_iters",
         "2",
         "--learning_parameters.batch_size",
-        "2",
+        "4",
     ],
 }