Update app.py

app.py

@@ -37,8 +37,7 @@ emo_dict = {
 class_options = {
     "racism": ["racism", "hate speech", "bigotry", "racially targeted", "racial slur", "ethnic slur", "ethnic hate", "pro-white nationalism"],
     "LGBTQ+ hate": ["gay slur", "trans slur", "homophobic slur", "transphobia", "anti-LBGTQ+", "hate speech"],
-    "sexually explicit": ["sexually explicit", "sexually coercive", "sexual exploitation", "vulgar", "raunchy", "sexist", "sexually demeaning", "sexual violence", "victim blaming"],
-    "misophonia": ["chewing", "breathing", "mouthsounds", "popping", "sneezing", "yawning", "smacking", "sniffling", "panting"]
+    "sexually explicit": ["sexually explicit", "sexually coercive", "sexual exploitation", "vulgar", "raunchy", "sexist", "sexually demeaning", "sexual violence", "victim blaming"]
 }
 
 pipe = pipeline("automatic-speech-recognition", model="openai/whisper-large")
@@ -72,99 +71,59 @@ def classify_toxicity(audio_file, text_input, classify_anxiety, emo_class, expli
         transcribed_text = pipe(audio_file)["text"]
     else:
         transcribed_text = text_input
-    if classify_anxiety != "misophonia":
-        print("emo_class ", emo_class, "explitive select", explitive_selection)
 
-        ## SLIDER ##
-        threshold = slider_logic(slider)
-        
-        #------- explitive call ---------------
-        
-        if replace_explitives != None and emo_class == None:
-            transcribed_text = replace_explitives.sub_explitives(transcribed_text, explitive_selection)
-        
-        #### Toxicity Classifier ####
-            
-        toxicity_module = evaluate.load("toxicity",  "facebook/roberta-hate-speech-dynabench-r4-target")
-        #toxicity_module = evaluate.load("toxicity", 'DaNLP/da-electra-hatespeech-detection', module_type="measurement")
-    
-        toxicity_results = toxicity_module.compute(predictions=[transcribed_text])
-     
-        toxicity_score = toxicity_results["toxicity"][0]
-        print(toxicity_score)
-        # emo call
-        if emo_class != None:
-            classify_emotion(audio_file)
-
-        #### Text classification #####
+    print("emo_class ", emo_class, "explitive select", explitive_selection)
+
+    ## SLIDER ##
+    threshold = slider_logic(slider)
     
-        device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    #------- explitive call ---------------
     
-        text_classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
+    if replace_explitives != None and emo_class == None:
+        transcribed_text = replace_explitives.sub_explitives(transcribed_text, explitive_selection)
     
-        sequence_to_classify = transcribed_text
-        print(classify_anxiety, class_options)
-        candidate_labels = class_options.get(classify_anxiety, [])
-        # classification_output = classifier(sequence_to_classify, candidate_labels, multi_label=False)
-        classification_output = text_classifier(sequence_to_classify, candidate_labels, multi_label=True)
-        print("class output ", type(classification_output))
-        # classification_df = pd.DataFrame.from_dict(classification_output)
-        print("keys ", classification_output.keys())
+    #### Toxicity Classifier ####
         
-        # formatted_classification_output = "\n".join([f"{key}: {value}" for key, value in classification_output.items()])
-        # label_score_pairs = [(label, score) for label, score in zip(classification_output['labels'], classification_output['scores'])]
-        label_score_dict = {label: score for label, score in zip(classification_output['labels'], classification_output['scores'])}
-
-
+    toxicity_module = evaluate.load("toxicity",  "facebook/roberta-hate-speech-dynabench-r4-target")
+    #toxicity_module = evaluate.load("toxicity", 'DaNLP/da-electra-hatespeech-detection', module_type="measurement")
 
-
-
-        # plot.update(x=classification_df["labels"], y=classification_df["scores"])
-        if toxicity_score > threshold:
-            print("threshold exceeded!! Launch intervention")
-            affirm = positive_affirmations()
-        else:
-            affirm = ""
+    toxicity_results = toxicity_module.compute(predictions=[transcribed_text])
  
-        return toxicity_score, label_score_dict, transcribed_text, affirm
-        # return f"Toxicity Score ({available_models[selected_model]}): {toxicity_score:.4f}"
-    else: 
-        threshold = slider_logic(slider)
-        model = whisper.load_model("large")
-        # model = model_cache[model_name]
-        # class_names = classify_anxiety.split(",")
-        class_names_list = class_options.get(classify_anxiety, [])
-        class_str = ""
-        for elm in class_names_list:
-            class_str += elm + ","
-        #class_names = class_names_temp.split(",")
-        class_names = class_str.split(",")
-        print("class names ", class_names, "classify_anxiety ", classify_anxiety)
-        
-        tokenizer = get_tokenizer("large")
-        # tokenizer= WhisperTokenizer.from_pretrained("openai/whisper-large")
+    toxicity_score = toxicity_results["toxicity"][0]
+    print(toxicity_score)
+    # emo call
+    if emo_class != None:
+        classify_emotion(audio_file)
+
+    #### Text classification #####
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+    text_classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
+
+    sequence_to_classify = transcribed_text
+    print(classify_anxiety, class_options)
+    candidate_labels = class_options.get(classify_anxiety, [])
+    # classification_output = classifier(sequence_to_classify, candidate_labels, multi_label=False)
+    classification_output = text_classifier(sequence_to_classify, candidate_labels, multi_label=True)
+    print("class output ", type(classification_output))
+    # classification_df = pd.DataFrame.from_dict(classification_output)
+    print("keys ", classification_output.keys())
+    
+    # formatted_classification_output = "\n".join([f"{key}: {value}" for key, value in classification_output.items()])
+    # label_score_pairs = [(label, score) for label, score in zip(classification_output['labels'], classification_output['scores'])]
+    label_score_dict = {label: score for label, score in zip(classification_output['labels'], classification_output['scores'])}
+
+    # plot.update(x=classification_df["labels"], y=classification_df["scores"])
+    if toxicity_score > threshold:
+        print("threshold exceeded!! Launch intervention")
+        affirm = positive_affirmations()
+    else:
+        affirm = ""
+
+    return toxicity_score, label_score_dict, transcribed_text, affirm
+    # return f"Toxicity Score ({available_models[selected_model]}): {toxicity_score:.4f}"
     
-        internal_lm_average_logprobs = classify.calculate_internal_lm_average_logprobs(
-            model=model,
-            class_names=class_names,
-            # class_names=classify_anxiety,
-            tokenizer=tokenizer,
-        )
-        audio_features = classify.calculate_audio_features(audio_file, model)
-        average_logprobs = classify.calculate_average_logprobs(
-            model=model,
-            audio_features=audio_features,
-            class_names=class_names,
-            tokenizer=tokenizer,
-        )
-        average_logprobs -= internal_lm_average_logprobs
-        scores = average_logprobs.softmax(-1).tolist()
-        holder1 = {class_name: score for class_name, score in zip(class_names, scores)}
-        # miso_label_dict = {label: score for label, score in classify_anxiety[0].items()}
-        holder2 = ""
-        holder3= " "
-        return holder1, holder1, holder2, holder3
-        
 def positive_affirmations():
     affirmations = [
         "I have survived my anxiety before and I will survive again now",
@@ -178,7 +137,7 @@ def positive_affirmations():
 with gr.Blocks() as iface:
     show_state = gr.State([])
     with gr.Column():
-        anxiety_class = gr.Radio(["racism", "LGBTQ+ hate", "sexually explicit", "misophonia"])
+        anxiety_class = gr.Radio(["racism", "LGBTQ+ hate", "sexually explicit"])
         explit_preference = gr.Radio(choices=["N-Word", "B-Word", "All Explitives"], label="Words to omit from general anxiety classes", info="certain words may be acceptible within certain contects for given groups of people, and some people may be unbothered by explitives broadly speaking.")
         emo_class = gr.Radio(choices=["negaitve emotionality"], label="label", info="Select if you would like explitives to be considered anxiety-indiucing in the case of anger/ negative emotionality.")
         sense_slider = gr.Slider(minimum=1, maximum=5, step=1.0, label="How readily do you want the tool to intervene? 1 = in extreme cases and 5 = at every opportunity")