Fix classification model download logic to resolve zip issue (openvin…

…otoolkit#1967)

* Replace model download logic by pytorchcv to resolve zip issue

* Update change log

* Fix pre-commit

* Configure checkpoint before model

* Add pytorchcv to cls/seg requirement

---------
Signed-off-by: Songki Choi <songki.choi@intel.com>

CHANGELOG.md

@@ -28,7 +28,8 @@ All notable changes to this project will be documented in this file.
 
 - Fix exception -> warning for anomaly dump_feature option
 - Remove `dataset.with_empty_annotations()` to keep original input structure (<https://github.com/openvinotoolkit/training_extensions/pull/1964>)
-- Fix OV batch inference (saliency map generation)  (<https://github.com/openvinotoolkit/training_extensions/pull/1965>)
+- Fix OV batch inference (saliency map generation) (<https://github.com/openvinotoolkit/training_extensions/pull/1965>)
+- Replace EfficentNetB0 model download logic by pytorchcv to resolve zip issue (<https://github.com/openvinotoolkit/training_extensions/pull/1967>)
 
 ## \[v1.1.1\]
 

otx/algorithms/classification/adapters/mmcls/tasks/stage.py

@@ -30,8 +30,8 @@ def configure(self, model_cfg, model_ckpt, data_cfg, training=True, **kwargs):
 
         # Recipe + model
         cfg = self.cfg
-        self.configure_model(cfg, model_cfg, **kwargs)
         self.configure_ckpt(cfg, model_ckpt, kwargs.get("pretrained", None))
+        self.configure_model(cfg, model_cfg, **kwargs)
         self.configure_data(cfg, data_cfg, training)
         self.configure_task(cfg, training)
         return cfg

otx/algorithms/common/adapters/mmcv/models/backbones/efficientnet.py

@@ -18,6 +18,7 @@
 from mmcv.cnn import build_activation_layer
 from mmcv.cnn.bricks import ConvModule
 from mmcv.runner import load_checkpoint
+from pytorchcv.models.model_store import download_model
 from torch import nn
 from torch.nn import init
 
@@ -1281,5 +1282,5 @@ def init_weights(self, pretrained=None):
             load_checkpoint(self, pretrained)
             logger.info(f"init weight - {pretrained}")
         elif pretrained is not None:
-            load_checkpoint(self, pretrained_urls[self.model_name])
+            download_model(net=self, model_name=self.model_name)
             logger.info(f"init weight - {pretrained_urls[self.model_name]}")

requirements/classification.txt

@@ -4,3 +4,4 @@ mmcv-full==1.7.0
 mmcls==0.25.0
 timm
 otxdeploy==0.12.1  # FIXME: mmdeploy==? when PyPI packages are available
+pytorchcv

requirements/segmentation.txt

@@ -5,3 +5,4 @@ mmsegmentation==0.30.0
 scikit-image
 otxdeploy==0.12.1  # FIXME: mmdeploy==? when PyPI packages are available
 timm
+pytorchcv