Add rtmw-m-256x192 RTMW/RTMDet HF port

README.md

@@ -38,23 +38,12 @@ Detector: human AP = 56.4 on COCO val2017.
 ## Usage
 
 ```python
-from transformers import AutoImageProcessor
+from transformers import AutoConfig, AutoModel, AutoImageProcessor
 from PIL import Image
 import torch
 
-# Load model (requires trust_remote_code=True for custom architecture)
-from huggingface_hub import hf_hub_download
-
-# Or directly:
-import sys, json
-from pathlib import Path
-
-# Using the custom modules:
-from rtmw_modules.configuration_rtmw import RTMWConfig
-from rtmw_modules.modeling_rtmw import RTMWModel
-
-config = RTMWConfig.from_pretrained("akore/rtmw-m-256x192", trust_remote_code=True)
-model = RTMWModel.from_pretrained("akore/rtmw-m-256x192", trust_remote_code=True)
+config = AutoConfig.from_pretrained("akore/rtmw-m-256x192", trust_remote_code=True)
+model = AutoModel.from_pretrained("akore/rtmw-m-256x192", trust_remote_code=True)
 model.eval()
 
 processor = AutoImageProcessor.from_pretrained("akore/rtmw-m-256x192")

config.json

@@ -44,6 +44,6 @@
   "transformers_version": "5.2.0",
   "auto_map": {
     "AutoConfig": "configuration_rtmw.RTMWConfig",
-    "AutoModelForImageProcessing": "modeling_rtmw.RTMWModel"
+    "AutoModel": "modeling_rtmw.RTMWModel"
   }
 }

modeling_rtmw.py

@@ -1258,6 +1258,10 @@ class RTMWModel(PreTrainedModel):
     This model consists of a backbone, neck, and pose head for keypoint detection.
     All implementations use PyTorch only with no NumPy or OpenCV dependencies.
     """
+
+    config_class = RTMWConfig
+    base_model_prefix = "rtmw"
+    main_input_name = "pixel_values"
     
     def __init__(self, config: RTMWConfig):
         super().__init__(config)