bump

clip_endpoint.py

@@ -1,31 +1,28 @@
 import open_clip
 import torch
+import functools
+import onnxruntime as ort
 
-do_load = True
-if do_load:
-
-    #model_name = 'hf-hub:timm/ViT-L-16-SigLIP2-512'
-#    model, preprocess = open_clip.create_model_from_pretrained('hf-hub:timm/ViT-L-16-SigLIP2-512')
-
-
-    model_name = 'ViT-SO400M-14-SigLIP-384'
-    pretrained_name = 'webli'
-    model, _, preprocess = open_clip.create_model_and_transforms(model_name, pretrained=pretrained_name)
-    device = 'cpu'
-    model.eval()
-    tokenizer = open_clip.get_tokenizer(model_name)
+model_name = 'ViT-SO400M-16-SigLIP2-512'
+tokenizer = open_clip.get_tokenizer(model_name)
+#onnx_path = "/home/thebears/Source/ml/clip_extract/text/ViT-SO400M-16-SigLIP2-512_text"
+onnx_path = "/home/thebears/Source/ml/clip_extract/text/ViT-SO400M-16-SigLIP2-512_text_optimized"
+ort_sess = ort.InferenceSession(onnx_path)
 
 from bottle import route, run, template, request, debug
 @route('/encode')
 def get_matches():
-    query = request.query.get('query','A large bird eating corn')
+    query = request.query.get('query',None)
+    if query is None:
+        return None
+    return execute_model(query)
+
+@functools.cache
+def execute_model(query):
     with torch.no_grad():
         text_tokenized = tokenizer(query)
-        vec = model.encode_text(text_tokenized).detach().cpu().tolist()
-
+        vec = ort_sess.run(None, {'tokenized': text_tokenized.numpy() })[0].tolist()
     return {'vector':vec}
 
-
-
 debug(True)
 run(host='0.0.0.0', port=53004, server='bjoern')