Change input format for multimodal search

misinformation/multimodal_search.py

@@ -84,12 +84,12 @@ class MultimodalSearch(AnalysisMethod):
                 enabled=(MultimodalSearch.multimodal_device != torch.device("cpu"))
             ):
                 features_image = [
-                    model.extract_features({"image": ten}, mode="image")
+                    model.extract_features({"image": ten, "text_input": ""}, mode="image")
                     for ten in images_tensors
                 ]
         else:
             features_image = [
-                model.extract_features({"image": ten}, mode="image")
+                model.extract_features({"image": ten, "text_input": ""}, mode="image")
                 for ten in images_tensors
             ]
 
@@ -114,6 +114,7 @@ class MultimodalSearch(AnalysisMethod):
 
         return features_text
     
+    
     def parsing_images(self, model_type):
         image_keys = sorted(self.keys())
         image_names = [self[k]["filename"] for k in image_keys]
@@ -132,31 +133,40 @@ class MultimodalSearch(AnalysisMethod):
         )
         MultimodalSearch.save_tensors(features_image_stacked)
 
-        return image_keys, image_names, features_image_stacked
+        return model, vis_processors, txt_processors, image_keys, image_names, features_image_stacked
 
     def multimodal_search(
-        self, model_type, image_keys, features_image_stacked, search_query
+        self, model, vis_processors, txt_processors, model_type, image_keys, features_image_stacked, search_query
     ):
         features_image_stacked.to(MultimodalSearch.multimodal_device)
-        (
-            model,
-            vis_processors,
-            txt_processors,
-        ) = MultimodalSearch.load_feature_extractor_model(
-            MultimodalSearch.multimodal_device, model_type
-        )
-        multi_text_input = [txt_processors["eval"](query) for query in search_query]
-        multi_sample = [{"text_input": [query]} for query in multi_text_input]
-        multi_features_text = [
-            model.extract_features(sample, mode="text") for sample in multi_sample
-        ]
-        multi_features_text_stacked = torch.stack(
-            [
-                features.text_embeds_proj[:, 0, :].squeeze(0)
-                for features in multi_features_text
-            ]
-        ).to(MultimodalSearch.multimodal_device)
-        similarity = features_image_stacked @ multi_features_text_stacked.t()
+        
+        for query in search_query:
+            if (len(query)!=1):
+               raise SyntaxError('Each querry must contain either an "image" or a "text_input"')
+
+        multi_sample = []
+        for query in search_query:
+            if "text_input" in query.keys():
+                text_processing = txt_processors["eval"](query["text_input"])
+                image_processing = ""
+            elif "image" in query.keys():
+                _, image_processing = MultimodalSearch.read_and_process_images([query["image"]], vis_processors)
+                text_processing = ""
+            multi_sample.append({"image": image_processing, "text_input": text_processing})
+        
+        multi_features_query = []
+        for query in multi_sample:
+            if query["image"] == "":
+                features = model.extract_features(query, mode="text")
+                features_squeeze = features.text_embeds_proj[:, 0, :].squeeze(0).to(MultimodalSearch.multimodal_device)
+                multi_features_query.append(features_squeeze)
+            if query["text_input"] == "":
+                multi_features_query.append( MultimodalSearch.extract_image_features(
+            model, query["image"], model_type))
+        
+        multi_features_stacked = torch.stack([query.squeeze(0) for query in multi_features_query]).to(MultimodalSearch.multimodal_device)
+
+        similarity = features_image_stacked @ multi_features_stacked.t()
         sorted_lists = [
             sorted(range(len(similarity)), key=lambda k: similarity[k, i], reverse=True)
             for i in range(len(similarity[0]))
@@ -165,13 +175,19 @@ class MultimodalSearch(AnalysisMethod):
 
         for q in range(len(search_query)):
             for i, key in zip(range(len(image_keys)), image_keys):
-                self[key]["rank " + search_query[q]] = places[q][i]
-                self[key][search_query[q]] = similarity[i][q].item()
+                self[key]["rank " + list(search_query[q].values())[0]] = places[q][i]
+                self[key][list(search_query[q].values())[0]] = similarity[i][q].item()
 
-        return self
+        return similarity
 
     def show_results(self, query):
-        for s in sorted(self.items(), key=lambda t: t[1][query], reverse=True):
+        if "image" in query.keys():
+            pic = Image.open(query["image"]).convert("RGB")
+            pic.thumbnail((400, 400))
+            display("Your search query: ", pic,"--------------------------------------------------", "Results:")
+        elif "text_input" in query.keys():
+            display("Your search query: " + query["text_input"], "--------------------------------------------------", "Results:")
+        for s in sorted(self.items(), key=lambda t: t[1][list(query.values())[0]], reverse=True):
             p1 = Image.open(s[1]["filename"]).convert("RGB")
             p1.thumbnail((400, 400))
-            display(p1, s[1][query])
+            display(p1, "Rank: " + str(s[1]["rank " + list(query.values())[0]]) + " Val: " + str(s[1][list(query.values())[0]]))