Update app.py

app.py

@@ -1,6 +1,6 @@
 import gradio as gr
 import random
-
+import torch
 from transformers import AutoConfig
 from transformers import GPT2Tokenizer, GPT2LMHeadModel
 
@@ -21,38 +21,38 @@ SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"]
 
 #See document for experiment https://www.linkedin.com/pulse/ai-goes-job-interview-g%C3%B6rkem-g%C3%B6knar/
 
-def get_chat_response(name, input_txt = "Hello , what is your name?"):
-  #I trained my dataset with "My name is " as prefix, this will let me dump personalities 
-  #and also gives character sense if of his/her/its name!
+
+ 
+def get_chat_response(name,history=[], input_txt = "Hello , what is your name?"):
+  
+  history.append(input_txt)
+  history_e = [tokenizer.encode(e) for e in history]
+
   personality = "My name is " + name
 
-  if input_txt[:-1] != ".":
-    #add a dot after sentence to make model understand it more clearly
-    input_txt = input_txt + "."
-    
-  ##can respond well to history as well but for this quick demo not implemented
-  ##see metayazar.com/chatbot for a min 2 history
+  bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:-1])
+  sequence = [[bos]  + tokenizer.encode(personality)] + history_e  
+  sequence = [sequence[0]] + [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence[1:])]
+  sequence = list(chain(*sequence))
+
   
-  ##this is a multi-speaker model, currently no history, so ending with "<speaker2>" to get response. depends on who starts conversation it can be speaker1
-  ##if there is a history depends on who started conversation it should end with <speaker1>
-  #historical implementation not implemented in this demo
-    bot_input_ids =  tokenizer.encode(tokenizer.bos_token +  personality + "<speaker1>" + input_txt + tokenizer.eos_token   , return_tensors='pt')
+  #bot_input_ids = tokenizer.encode(personality + tokenizer.eos_token + input_txt + tokenizer.eos_token  , return_tensors='pt')
+  sequence_len = len(sequence)
 
   #optimum response and speed
-  #50 token max length, temperature = 1.3 makes it creative
   chat_history_ids = model.generate(
-          bot_input_ids,min_length =1, max_length=50,
+          torch.tensor(sequence).unsqueeze(0), max_length=50,
           pad_token_id=tokenizer.eos_token_id,  
           no_repeat_ngram_size=3,       
           do_sample=True, 
-          top_k=50, 
-          top_p=0.9,
-          temperature = temperature 
+          top_k=60, 
+          top_p=0.8,
+          temperature = 1.3
       )
-
-  out_str = tokenizer.decode(chat_history_ids[:, bot_input_ids.shape[-1]:][0], skip_special_tokens=True)
+  out_str = tokenizer.decode(chat_history_ids[0][sequence_len:],  skip_special_tokens=True)
+  #out_str = tokenizer.decode(chat_history_ids[:, sequence.shape[-1]:][0], skip_special_tokens=False)
   return out_str
-
+  
 ##you can use anyone from below
 '''
 | Macleod | Moran | Brenda | Ramirez | Peter Parker | Quentin Beck | Andy