Merge pull request #6 from Lednik7/dev

add load example, notes

README.md

@@ -62,7 +62,32 @@ probs = logits_per_image.softmax(dim=-1).detach().cpu().numpy()
 
 print("Label probs:", probs)  # prints: [[0.41456965 0.29270944 0.29272085]]
 ```
-Enjoy the speed
+#### Notes:
+If you use:
+```python3
+clip.model.ResidualAttentionBlock.attention = attention
+```
+The results may not match. To avoid this, you must manually change the **logit_scale** parameter.\
+For the example above:
+```python3
+onnx_model.logit_scale = 1600 # -> Label probs: [[0.99242663 0.00378558 0.00378777]]
+```
+**Enjoy the speed**
+
+## Load saved model
+Example for ViT-B/32 from Model Zoo
+```python3
+!wget https://clip-as-service.s3.us-east-2.amazonaws.com/models/onnx/ViT-B-32/visual.onnx
+!wget https://clip-as-service.s3.us-east-2.amazonaws.com/models/onnx/ViT-B-32/textual.onnx
+```
+```python3
+onnx_model = clip_onnx(None)
+onnx_model.load_onnx(visual_path="visual.onnx",
+                     textual_path="textual.onnx",
+                     logit_scale=100.0000) # model.logit_scale.exp()
+onnx_model.start_sessions(providers=["CPUExecutionProvider"])
+```
+
 ## Model Zoo
 Models of the original CLIP can be found on this [page](https://github.com/jina-ai/clip-as-service/blob/main/server/clip_server/model/clip_onnx.py).\
 They are not part of this library but should work correctly.