add scoring

app.py

@@ -2,7 +2,7 @@ import gradio as gr
 import random
 import os
 import copy
-
+import torch
 from huggingface_hub import login
 from transformers import pipeline
 from transformers import GPT2Tokenizer, GPT2LMHeadModel,set_seed
@@ -54,18 +54,40 @@ def create_response(input_str,
   if seed != -1:
     set_seed(seed)    
   if model_name == "original_model":      
-    output_ids = original_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences )                                
-  elif model_name == "untethered_model":      
-    output_ids = untethered_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences )                                      
+    output = original_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
+    transition_scores = original_model.compute_transition_scores(output.sequences, output.scores,
+                                                                 normalize_logits=False)
+
+  elif model_name == "untethered_model":
+    output = untethered_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
+    transition_scores = untethered_model.compute_transition_scores(output.sequences, output.scores,
+                                                               normalize_logits=False)
   elif model_name == "untethered_paraphrased_model":
-    output_ids = untethered_paraphrased_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences )                                          
+    output = untethered_paraphrased_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
+    transition_scores = untethered_paraphrased_model.compute_transition_scores(output.sequences, output.scores,
+                                                               normalize_logits=False)
   else:
-    output_ids = original_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences )
-      
+    output = original_model.generate(input_ids,pad_token_id=tokenizer.eos_token_id,do_sample=do_sample, attention_mask=attention_mask, max_length=100, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty,num_return_sequences=num_return_sequences,return_dict_in_generate=True, output_scores=True )
+    transition_scores = original_model.compute_transition_scores(output.sequences, output.scores,
+                                                                 normalize_logits=False)
+  score_list = []
+  for scores in transition_scores:
+      # exp_tensor_list = torch.exp(scores)
+      # print(scores)
+      # print(exp_tensor_list)
+      sum_exp_tensor_list = torch.sum(scores)
+      len_exp_tensor_list = torch.numel(scores)
+      average_exp_tensor_list = sum_exp_tensor_list / len_exp_tensor_list
+      print(average_exp_tensor_list)
+      score_list.append(average_exp_tensor_list.item())
   outputs = model_name+" generated <br>"
-  for output_id in output_ids:
-    output = tokenizer.decode(output_id, skip_special_tokens=True)
-    outputs=outputs+output+"<br/>"
+  for index, output_id in enumerate(output.sequences):
+      sentence = tokenizer.decode(output_id, skip_special_tokens=True)
+      print(sentence + " score:" + str(score_list[index]))
+
+  # for output_id in output_ids:
+  #   output = tokenizer.decode(output_id, skip_special_tokens=True)
+  #   outputs=outputs+output+"<br/>"
   return outputs