text simplification

app.py

@@ -2,6 +2,7 @@ import gradio as gr
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import torch
 from collections import Counter
+from scipy.special import softmax
 
 article_string = "Author: <a href=\"https://huggingface.co/ruanchaves\">Ruan Chaves Rodrigues</a>. Read more about our <a href=\"https://github.com/ruanchaves/eplm\">research on the evaluation of Portuguese language models</a>."
 
@@ -56,6 +57,10 @@ user_friendly_name = {
     "ruanchaves/bert-base-portuguese-cased-porsimplessent": "BERTimbau base (PorSimplesSent)",
 }
 
+reverse_user_friendly_name = { v:k for k,v in user_friendly_name.items() }
+
+user_friendly_name_list = list(user_friendly_name.values())
+
 model_array = []
 
 for model_name in model_list:
@@ -69,36 +74,43 @@ def most_frequent(array):
     occurence_count = Counter(array)
     return occurence_count.most_common(1)[0][0]
 
-def predict(s1, s2):
+def predict(s1, s2, chosen_model):
+    if not chosen_model:
+      chosen_model = user_friendly_name_list[0]
     scores = {}
+    full_chosen_model_name = reverse_user_friendly_name[chosen_model]
     for row in model_array:
-        name = user_friendly_name[row["name"]]
-        tokenizer = row["tokenizer"]
-        model = row["model"]
-        model_input = tokenizer(*([s1], [s2]), padding=True, return_tensors="pt")
-        with torch.no_grad():
-            output = model(**model_input)
-            score = output[0][0].argmax().item()
-            scores[name] = score
-    average_score = most_frequent(list(scores.values()))
-    description = score_descriptions[average_score]
-    description_pt = score_descriptions_pt[average_score]
-    final_description = description + "\n \n" + description_pt
-
-    for key, value in scores.items():
-      scores[key] = score_descriptions[value]
-
-    return final_description, scores
+        name = row["name"]
+        if name != full_chosen_model_name:
+          continue
+        else:
+          tokenizer = row["tokenizer"]
+          model = row["model"]
+          model_input = tokenizer(*([s1], [s2]), padding=True, return_tensors="pt")
+          with torch.no_grad():
+              output = model(**model_input)
+              logits = output[0][0].detach().numpy()
+              logits = softmax(logits).tolist()
+              break
+    def get_description(idx):
+      description = score_descriptions[idx]
+      description_pt = score_descriptions_pt[idx]
+      final_description = description + "\n \n" + description_pt
+      return final_description
+    
+    scores = { get_description(k):v for k,v in enumerate(logits) }
+
+    return scores
 
 
 inputs = [
-    gr.inputs.Textbox(label="Sentence A"),
-    gr.inputs.Textbox(label="Sentence B")
+    gr.inputs.Textbox(label="Question"),
+    gr.inputs.Textbox(label="Answer"),
+    gr.Dropdown(label="Model", choices=user_friendly_name_list, default=user_friendly_name_list[0])
 ]
 
 outputs = [
- gr.Textbox(label="Evaluation", value=output_textbox_component_description),
- gr.JSON(label="Results by model", value=output_json_component_description)
+ gr.Label(label="Result")
 ]
 
 

requirements.txt

@@ -1,3 +1,4 @@
 torch
 gradio
-transformers
+transformers
+scipy