Patch clip model for ONNX compatibility

Changes to use INT32 for tokenization, since ONNX doesn't yet support ArgMax(INT64)
Use explicit dimension for norm

clip/clip.py

@@ -192,7 +192,7 @@ def patch_float(module):
     return model, _transform(model.input_resolution.item())
 
 
-def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> torch.LongTensor:
+def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: bool = False) -> torch.IntTensor:
     """
     Returns the tokenized representation of given input string(s)
 
@@ -217,7 +217,7 @@ def tokenize(texts: Union[str, List[str]], context_length: int = 77, truncate: b
     sot_token = _tokenizer.encoder["<|startoftext|>"]
     eot_token = _tokenizer.encoder["<|endoftext|>"]
     all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]
-    result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+    result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
 
     for i, tokens in enumerate(all_tokens):
         if len(tokens) > context_length:

clip/model.py

@@ -356,8 +356,8 @@ def forward(self, image, text):
         text_features = self.encode_text(text)
 
         # normalized features
-        image_features = image_features / image_features.norm(dim=-1, keepdim=True)
-        text_features = text_features / text_features.norm(dim=-1, keepdim=True)
+        image_features = image_features / image_features.norm(dim=1, keepdim=True)
+        text_features = text_features / text_features.norm(dim=1, keepdim=True)
 
         # cosine similarity as logits
         logit_scale = self.logit_scale.exp()