Update app.py

app.py

@@ -1,8 +1,10 @@
 import gradio as gr
+import numpy as np
 import openai
 
 from sentence_transformers import SentenceTransformer
 from langchain.prompts import PromptTemplate
+from collections import Counter
 
 
 def process(api, caption, category, asr, ocr):
@@ -10,19 +12,69 @@ def process(api, caption, category, asr, ocr):
     preference = "兴趣标签"
     example = "例如，给定一个视频，它的\"标题\"为\"长安系最便宜的轿车，4W起很多人都看不上它，但我知道车只是代步工具，又需要什么面子呢" \
               "！\"，\"类别\"为\"汽车\"，\"ocr\"为\"长安系最便宜的一款轿车\"，\"asr\"为\"我不否认现在的国产和合资还有一定的差距，" \
-              "但确实是他们让我们5万开了MP V8万开上了轿车，10万开张了ICV15万开张了大七座。\"，{}生成机器人推断出合理的\"{}\"为\"长" \
-              "安轿车报价、最便宜的长安轿车、新款长安轿车\"。".format(preference, preference)
+              "但确实是他们让我们5万开了MP V8万开上了轿车，10万开张了ICV15万开张了大七座。\"，\"{}\"生成机器人推断出合理的\"{}\"为\"" \
+              "长安轿车报价、最便宜的长安轿车、新款长安轿车\"。".format(preference, preference)
 
     prompt = PromptTemplate(
         input_variables=["preference", "caption", "ocr", "asr", "category", "example"],
-        template="你是一个视频的{preference}生成机器人，根据输入的视频标题、类别、ocr、asr推理出合理的\"{preference}\"，以多个多"
+        template="你是一个视频的\"{preference}\"生成机器人，根据输入的视频标题、类别、ocr、asr推理出合理的\"{preference}\"，以多个多"
                  "于两字的标签形式进行表达，以顿号隔开。{example}那么，给定一个新的视频，它的\"标题\"为\"{caption}\"，\"类别\"为"
                  "\"{category}\"，\"ocr\"为\"{ocr}\"，\"asr\"为\"{asr}\"，请推断出该视频的\"{preference}\"："
     )
 
     text = prompt.format(preference=preference, caption=caption, category=category, ocr=ocr, asr=asr, example=example)
 
-    return text
+    try:
+        completion = openai.ChatCompletion.create(
+            model="gpt-3.5-turbo",
+            messages=[{"role": "user", "content": text}],
+            temperature=1.5,
+            n=5
+        )
+
+        res = []
+        for j in range(5):
+            ans = completion.choices[j].message["content"].strip()
+            ans = ans.replace("\n", "")
+            ans = ans.replace("。", "")
+            ans = ans.replace("，", "、")
+            res += ans.split('、')
+
+        tag_count = Counter(res)
+        tag_count = sorted(tag_count.items(), key=lambda x: x[1], reverse=True)[:10]
+
+        tags_embed = np.load('./tag_data/tags_embed.npy')
+        tags_dis = np.load('./tag_data/tags_dis.npy')
+
+        candidate_tags = [_[0] for _ in tag_count]
+        encoder = SentenceTransformer("hfl/chinese-roberta-wwm-ext-large")
+        candidate_tags_embed = encoder.encode(candidate_tags)
+        candidate_tags_dis = [np.sqrt(np.dot(_, _.T)) for _ in candidate_tags_embed]
+
+        scores = np.dot(candidate_tags_embed, tags_embed.T)
+        f = open('./tag_data/tags.txt', 'r')
+        all_tags = []
+        for line in f.readlines():
+            all_tags.append(line.strip())
+        f.close()
+
+        final_ans = []
+        for i in range(scores.shape[0]):
+            for j in range(scores.shape[1]):
+                score = scores[i][j] / (candidate_tags_dis[i] * tags_dis[j])
+                if score > 0.8:
+                    final_ans.append(all_tags[j])
+
+        print(final_ans)
+
+        final_ans = Counter(final_ans)
+        final_ans = sorted(final_ans.items(), key=lambda x: x[1], reverse=True)[:5]
+        final_ans = [_[0] for _ in final_ans]
+
+        return "、".join(final_ans)
+
+    except:
+        return 'api error'
 
 
 with gr.Blocks() as demo:
@@ -39,4 +91,4 @@ with gr.Blocks() as demo:
 
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(share=True)