add length checker

(`c838b62` in https://bitbucket.org/maum-system/cvpr22-demo-gradio)

app.py

@@ -25,10 +25,11 @@ DESCRIPTION = Path("docs/description.md").read_text()
 class GradioApplication:
     def __init__(self, rest_ip, rest_port, max_seed):
         self.lang_list = {
-            'Korean': 'ko_KR',
-            'English': 'en_US',
-            'Japanese': 'ja_JP',
-            'Chinese': 'zh_CN'
+            'ko': 'ko_KR',
+            'en': 'en_US',
+            'ja': 'ja_JP',
+            'zh': 'zh_CN',
+            'zh-CN': 'zh_CN'
         }
         self.background_list = [None,
                                 "background_image/cvpr.png",
@@ -85,14 +86,14 @@ class GradioApplication:
         return background_data, is_video_background
     
     @staticmethod
-    def return_format(toxicity_prob, target_text, lang_dest, video_filename):
-        return {'Toxicity': toxicity_prob}, f"Language: {lang_dest}\nText: \n{target_text}", str(video_filename)   
+    def return_format(toxicity_prob, target_text, lang_dest, video_filename, detail=""):
+        return {'Toxicity': toxicity_prob}, f"Language: {lang_dest}\nText: {target_text}\nDetails: {detail}", str(video_filename)   
 
     def infer(self, text, lang, duration_rate, action, background_index):
         self._counter_file_seed()
         print(f"File Seed: {self._file_seed}")
         toxicity_prob = 0.0
-        target_text = "(Sorry, it seems that the input text is too toxic.)"
+        target_text = ""
         lang_dest = ""
         video_filename = "vacant.mp4"
         
@@ -103,16 +104,24 @@ class GradioApplication:
             pass
         
         if toxicity_prob > TOXICITY_THRESHOLD:
-            return self.return_format(toxicity_prob, target_text, lang_dest, video_filename)
+            detail = "Sorry, it seems that the input text is too toxic."
+            return self.return_format(toxicity_prob, target_text, lang_dest, video_filename, detail=detail)
         
         # Google Translate API
         try:
             target_text, lang_dest = self.translator.get_translation(text, lang)
-            lang_rpc_code = self.get_lang_code(lang_dest)
         except Exception as e:
-            target_text = f"Error from language translation: ({e})"
+            target_text = ""
             lang_dest = ""
-            return self.return_format(toxicity_prob, target_text, lang_dest, video_filename)
+            detail = f"Error from language translation: ({e})"
+            return self.return_format(toxicity_prob, target_text, lang_dest, video_filename, detail=detail)
+        
+        try:
+            self.translator.length_check(lang_dest, target_text)  # assertion check
+        except AssertionError as e:
+            return self.return_format(toxicity_prob, target_text, lang_dest, video_filename, detail=str(e))
+            
+        lang_rpc_code = self.get_lang_code(lang_dest)
 
         # Video Inference
         background_data, is_video_background = self.get_background_data(background_index)
@@ -125,7 +134,7 @@ class GradioApplication:
         with open(video_filename, "wb") as video_file:
             video_file.write(video_data)
         
-        return {'Toxicity': toxicity_prob}, f"Language: {lang_dest}\nText: \n{target_text}", str(video_filename)        
+        return self.return_format(toxicity_prob, target_text, lang_dest, video_filename)     
 
     def run(self, server_port=7860, share=False):
         try:

docs/description.md

@@ -5,6 +5,10 @@ If the input text language and the target language are different, the input text
 
 ### Updates
 
+(2022.06.17.) Thank you for visiting our demo!😊 This demo attracted a lot more attention than we anticipated. This, unfortunately, means that the computational burden is heavier than this demo was designed for. So, to maximize everyone's experience, we capped the length of the translated texts at:
+- 200 characters for English
+- 100 characters for Chinese, Japaense, and Korean.
+
 (2022.06.17.) We were originally planning to support any input text. However, when checking the logs recently, we found that there were a lot of inappropriate input texts. So, we decided to filter the inputs based on toxicity using [Perspective API @Google](https://developers.perspectiveapi.com/s/). Now, if you enter a possibily toxic text, the video generation will fail. We hope you understand.
 
 (2022.06.05.) Due to the latency from HuggingFace Spaces and video rendering, it takes 15 ~ 30 seconds to get a video result.

translator/module.py

@@ -1,17 +1,34 @@
 from .v3 import GoogleAuthTranslation
 from pathlib import Path
 import yaml
+import os
 
+MAX_ENG_TEXT_LENGTH = int(os.getenv('MAX_ENG_TEXT_LENGTH', 200))
+MAX_CJK_TEXT_LENGTH = int(os.getenv('MAX_CJK_TEXT_LENGTH', 100))
 
 class Translator:
     def __init__(self, yaml_path='./lang.yaml'):
         self.google_translation = GoogleAuthTranslation(project_id="cvpr-2022-demonstration")
         with open(yaml_path) as f:
             self.supporting_languages = yaml.load(f, Loader=yaml.FullLoader)
-            
+    
+    @staticmethod 
+    def length_check(lang, text):
+        if lang in ['en']:
+            if len(text) > MAX_ENG_TEXT_LENGTH:
+                raise AssertionError(f"Input text is too long. For English, the text length should be less than {MAX_ENG_TEXT_LENGTH}. | Length: {len(text)}")
+        elif lang in ['ko', 'ja', 'zh-CN', 'zh']:
+            if len(text) > MAX_CJK_TEXT_LENGTH:
+                raise AssertionError(f"Input text is too long. For CJK, the text length should be less than {MAX_CJK_TEXT_LENGTH}. | Length: {len(text)}")
+        else:
+            raise AssertionError(f"Not in ['ko', 'ja', 'zh-CN', 'zh', 'en'] ! | Language: {lang}")
+        
+        return
+        
     def _get_text_with_lang(self, text, lang):
         lang_detected = self.google_translation.detect(text)
-        print(lang_detected, lang)
+        print(f"Detected as: {lang_detected} | Destination: {lang}")
+        
         if lang is None:
             lang = lang_detected
             
@@ -19,22 +36,16 @@ class Translator:
             target_text = self.google_translation.translate(text, lang=lang)
         else:
             target_text = text
-            
+                
         return target_text, lang
     
     def _convert_lang_from_index(self, lang):
         try:
-            lang_finder = [name for name in self.supporting_languages
-                            if self.supporting_languages[name]['language'] == lang]
+            lang = [name for name in self.supporting_languages
+                    if self.supporting_languages[name]['language'] == lang][0]
         except Exception as e:
             raise RuntimeError(e)
         
-        if len(lang_finder) == 1:
-            lang = lang_finder[0]
-        else:
-            raise AssertionError("Given language index can't be understood!"
-                                 f"Only one of ['Korean', 'English', 'Japanese', 'Chinese'] can be supported. | lang: {lang}")
-        
         return lang
 
     def get_translation(self, text, lang, use_translation=True):
@@ -45,4 +56,4 @@ class Translator:
         else:
             target_text = text
 
-        return target_text, lang 
+        return target_text, lang_