standalone demo script with cpu

demo.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+import os, sys, importlib.util, re
+
+# ——— Monkey-patch mmdet to remove its mmcv-version assertion ———
+spec = importlib.util.find_spec('mmdet')
+if spec and spec.origin:
+    src = open(spec.origin, encoding='utf-8').read()
+    # strip out the mmcv_minimum_version…assert… block up to __all__
+    patched = re.sub(r'(?ms)^[ \t]*mmcv_minimum_version.*?^__all__', '__all__', src)
+    m = importlib.util.module_from_spec(spec)
+    m.__loader__ = spec.loader
+    m.__file__ = spec.origin
+    m.__path__ = spec.submodule_search_locations
+    sys.modules['mmdet'] = m
+    exec(compile(patched, spec.origin, 'exec'), m.__dict__)
+
+# ——— Configuration: set your image & output folder here ———
+IMAGE_PATH = "000000197388.jpg"
+VIS_OUT_DIR = "vis_results"
+POSE2D = "rtmo"          # hard-code the RTMO 2D model alias
+DEVICE = None            # e.g. "cuda:0" or None to auto-select
+
+# ——— Inference ———
+from mmpose.apis.inferencers import MMPoseInferencer
+
+# create output folder
+os.makedirs(VIS_OUT_DIR, exist_ok=True)
+
+# instantiate the inferencer
+inferencer = MMPoseInferencer(
+    pose2d=POSE2D,
+    scope="mmpose",
+    device=DEVICE,
+    det_cat_ids=[0],
+    # you can override any other init args here…
+)
+
+# run on our single image, with the RTMO-specific defaults,
+# iterating the generator so that visualization actually happens:
+for result in inferencer(
+    inputs=IMAGE_PATH,
+    bbox_thr=0.1,
+    nms_thr=0.65,
+    pose_based_nms=True,
+    show=False,
+    vis_out_dir=VIS_OUT_DIR,
+):
+    # result is a dict with keys "visualization" and "predictions"
+    # you can inspect it here if you like, e.g.:
+    # print(result['predictions'])
+    pass
+
+print(f"Visualized results saved to {VIS_OUT_DIR}/")

requirements.txt

@@ -0,0 +1,9 @@
+--find-links https://download.openmmlab.com/mmcv/dist/cpu/torch2.4/index.html
+mmcv==2.2.0 
+torch==2.4.0
+opencv-python
+numpy
+mmengine
+mmdet
+mmpose
+gradio