openvino: new model, use huggingface as model source

plugins/openvino/package-lock.json

@@ -1,12 +1,12 @@
 {
    "name": "@scrypted/openvino",
-   "version": "0.1.192",
+   "version": "0.1.193",
    "lockfileVersion": 2,
    "requires": true,
    "packages": {
       "": {
          "name": "@scrypted/openvino",
-         "version": "0.1.192",
+         "version": "0.1.193",
          "devDependencies": {
             "@scrypted/sdk": "file:../../sdk"
          }

plugins/openvino/package.json

@@ -50,5 +50,5 @@
    "devDependencies": {
       "@scrypted/sdk": "file:../../sdk"
    },
-   "version": "0.1.192"
+   "version": "0.1.193"
 }

plugins/openvino/src/ov/__init__.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 import asyncio
 import concurrent.futures
 import json
@@ -168,15 +169,8 @@ class OpenVINOPlugin(
 
         ovmodel = "best-converted"
 
-        model_version = "v7"
-        xmlFile = self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.xml",
-            f"{model_version}/{model}/{ovmodel}.xml",
-        )
-        self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.bin",
-            f"{model_version}/{model}/{ovmodel}.bin",
-        )
+        model_path = self.downloadHuggingFaceModelLocalFallback(model)
+        xmlFile = os.path.join(model_path, f"{ovmodel}.xml")
 
         try:
             self.compiled_model = self.core.compile_model(xmlFile, mode)

plugins/openvino/src/ov/face_recognition.py

@@ -1,11 +1,12 @@
 from __future__ import annotations
 
 import asyncio
+import os
 
 import numpy as np
-import openvino as ov
 from PIL import Image
 
+import openvino as ov
 from ov import async_infer
 from predict.face_recognize import FaceRecognizeDetection
 
@@ -20,18 +21,12 @@ class OpenVINOFaceRecognition(FaceRecognizeDetection):
         super().__init__(plugin=plugin, nativeId=nativeId)
 
     def downloadModel(self, model: str):
-        scrypted_yolov9 = "scrypted_yolov9" in model
         inception = "inception" in model
-        ovmodel = "best-converted" if scrypted_yolov9 else "best"
-        model_version = "v8"
-        xmlFile = self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.xml",
-            f"{model_version}/{model}/{ovmodel}.xml",
-        )
-        self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.bin",
-            f"{model_version}/{model}/{ovmodel}.bin",
-        )
+        ovmodel = "best-converted" if not inception else "best"
+        if not inception:
+            model = model + "_int8"
+        model_path = self.downloadHuggingFaceModelLocalFallback(model)
+        xmlFile = os.path.join(model_path, f"{ovmodel}.xml")
         if inception:
             model = self.plugin.core.read_model(xmlFile)
             model.reshape([1, 3, 160, 160])

plugins/openvino/src/ov/text_recognition.py

@@ -1,10 +1,11 @@
 from __future__ import annotations
 
 import asyncio
+import os
 
 import numpy as np
-import openvino as ov
 
+import openvino as ov
 from ov import async_infer
 from predict.text_recognize import TextRecognition
 
@@ -17,15 +18,8 @@ textRecognizePrepare, textRecognizePredict = async_infer.create_executors(
 class OpenVINOTextRecognition(TextRecognition):
     def downloadModel(self, model: str):
         ovmodel = "best"
-        model_version = "v6"
-        xmlFile = self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.xml",
-            f"{model_version}/{model}/{ovmodel}.xml",
-        )
-        self.downloadFile(
-            f"https://huggingface.co/scrypted/plugin-models/resolve/main/openvino/{model}/{ovmodel}.bin",
-            f"{model_version}/{model}/{ovmodel}.bin",
-        )
+        model_path = self.downloadHuggingFaceModelLocalFallback(model)
+        xmlFile = os.path.join(model_path, f"{ovmodel}.xml")
         if "vgg" in model:
             model = self.plugin.core.read_model(xmlFile)
             # this reshape causes a crash on GPU but causes a crash if NOT used with NPU...

plugins/openvino/src/predict/__init__.py

@@ -20,6 +20,10 @@ import common.colors
 from detect import DetectPlugin
 from predict.rectangle import Rectangle
 
+cache_dir = os.path.join(os.environ["SCRYPTED_PLUGIN_VOLUME"], "files", "hf")
+os.makedirs(cache_dir, exist_ok=True)
+os.environ['HF_HUB_CACHE'] = cache_dir
+
 original_getaddrinfo = socket.getaddrinfo
 
 # Sort the results to put IPv4 addresses first
@@ -84,6 +88,32 @@ class PredictPlugin(DetectPlugin, scrypted_sdk.ClusterForkInterface, scrypted_sd
         if not self.plugin and not self.forked:
             asyncio.ensure_future(self.startCluster(), loop=self.loop)
 
+    def downloadHuggingFaceModel(self, model: str, local_files_only: bool = False) -> str:
+        from huggingface_hub import snapshot_download
+        plugin_suffix = self.pluginId.split('/')[1]
+        local_path = snapshot_download(
+            repo_id="scrypted/plugin-models",
+            allow_patterns=f"{plugin_suffix}/{model}/*",
+            local_files_only=local_files_only,
+        )
+        local_path = os.path.join(local_path, plugin_suffix, model)
+        return local_path
+
+    def downloadHuggingFaceModelLocalFallback(self, model: str) -> str:
+        try:
+            local_path = self.downloadHuggingFaceModel(model)
+            print("Downloaded/refreshed model:", model)
+            return local_path
+        except Exception:
+            traceback.print_exc()
+
+            print("Unable to download model:", model)
+            print('This may be due to network or firewall issues.')
+
+        print("Trying model from Hugging Face Hub (offline):", model)
+        local_path = self.downloadHuggingFaceModel(model, local_files_only=True)
+        return local_path
+
     def downloadFile(self, url: str, filename: str):
         try:
             filesPath = os.path.join(os.environ["SCRYPTED_PLUGIN_VOLUME"], "files")

plugins/openvino/src/predict/clip.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import asyncio
 import base64
-import os
 from typing import Tuple
 
 import scrypted_sdk
@@ -25,30 +24,26 @@ class ClipEmbedding(PredictPlugin, scrypted_sdk.TextEmbedding, scrypted_sdk.Imag
         self.minThreshold = 0.5
 
         self.model = self.initModel()
-        cache_dir = os.path.join(os.environ["SCRYPTED_PLUGIN_VOLUME"], "files", "hf")
-        os.makedirs(cache_dir, exist_ok=True)
 
         self.processor = None
         print("Loading CLIP processor from local cache.")
         try:
             self.processor = CLIPProcessor.from_pretrained(
                 hf_id,
-                cache_dir=cache_dir,
                 local_files_only=True,
             )
             print("Loaded CLIP processor from local cache.")
         except Exception:
             print("CLIP processor not available in local cache yet.")
 
-        asyncio.ensure_future(self.refreshClipProcessor(hf_id, cache_dir), loop=self.loop)
+        asyncio.ensure_future(self.refreshClipProcessor(hf_id), loop=self.loop)
 
-    async def refreshClipProcessor(self, hf_id: str, cache_dir: str):
+    async def refreshClipProcessor(self, hf_id: str):
         try:
             print("Refreshing CLIP processor cache (online).")
             processor = await asyncio.to_thread(
                 CLIPProcessor.from_pretrained,
                 hf_id,
-                cache_dir=cache_dir,
             )
             self.processor = processor
             print("Refreshed CLIP processor cache.")

plugins/openvino/src/requirements.txt

@@ -13,4 +13,7 @@ openvino==2024.5.0
 Pillow==10.3.0
 opencv-python-headless==4.10.0.84
 
+# clip processor
 transformers==4.52.4
+# model downloads
+huggingface-hub