🐛 update defaults

Signed-off-by: peter szemraj <[email protected]>

app.py

@@ -44,11 +44,15 @@ import transformers
 
 transformers.logging.set_verbosity_error()
 cwd = Path.cwd()
-my_cwd = str(cwd.resolve())  # string so it can be passed to os.path() objects
+_cwd_str = str(cwd.resolve())  # string so it can be passed to os.path() objects
 
 
 def chat(
-    prompt_message, temperature: float = 0.5, top_p: float = 0.95, top_k: int = 20, constrained_generation: str = "False"
+    prompt_message,
+    temperature: float = 0.5,
+    top_p: float = 0.95,
+    top_k: int = 20,
+    constrained_generation: str = "False",
 ) -> str:
     """
     chat - the main function for the chatbot. This is the function that is called when the user
@@ -84,7 +88,7 @@ def ask_gpt(
     chat_pipe,
     speaker="person alpha",
     responder="person beta",
-    min_length=4,
+    min_length=12,
     max_length=48,
     top_p=0.95,
     top_k=25,
@@ -99,7 +103,7 @@ def ask_gpt(
     :param chat_pipe: the pipeline object for the model, created by the pipeline() function
     :param str speaker: the name of the speaker, defaults to "person alpha"
     :param str responder: the name of the responder, defaults to "person beta"
-    :param int min_length: the minimum length of the response, defaults to 4
+    :param int min_length: the minimum length of the response, defaults to 12
     :param int max_length: the maximum length of the response, defaults to 64
     :param float top_p: the top_p value for the model, defaults to 0.95
     :param int top_k: the top_k value for the model, defaults to 25
@@ -128,22 +132,20 @@ def ask_gpt(
         temperature=temperature,
         max_length=max_length,
         min_length=min_length,
-        constrained_beam_search  = constrained_generation,
+        constrained_beam_search=constrained_generation,
     )
     gpt_et = time.perf_counter()
     gpt_rt = round(gpt_et - st, 2)
     rawtxt = resp["out_text"]
     # check for proper nouns
     if basic_sc:
-        cln_resp = symspeller(rawtxt, sym_checker=schnellspell)
+        cln_resp = symspeller(rawtxt, sym_checker=basic_spell)
     else:
         cln_resp = synthesize_grammar(corrector=grammarbot, message=rawtxt)
     bot_resp_a = corr(remove_repeated_words(cln_resp))
     bot_resp = fix_punct_spacing(bot_resp_a)
     corr_rt = round(time.perf_counter() - gpt_et, 4)
-    print(
-        f"{gpt_rt + corr_rt} to respond, {gpt_rt} GPT, {corr_rt} for correction\n"
-    )
+    print(f"{gpt_rt + corr_rt} to respond, {gpt_rt} GPT, {corr_rt} for correction\n")
     return remove_trailing_punctuation(bot_resp)
 
 
@@ -163,6 +165,7 @@ def get_parser():
         help="the model to use for the chatbot on https://huggingface.co/models OR a path to a local model",
     )
     parser.add_argument(
+        "-gm",
         "--gram-model",
         required=False,
         type=str,
@@ -173,9 +176,9 @@ def get_parser():
     parser.add_argument(
         "--basic-sc",
         required=False,
-        default=False,  # TODO: change this back to False once Neuspell issues are resolved.
+        default=False,
         action="store_true",
-        help="turn on symspell (baseline) correction instead of the more advanced neural net models",
+        help="use symspell (statistical spelling correction) instead of neural spell correction",
     )
 
     parser.add_argument(
@@ -188,7 +191,7 @@ def get_parser():
         "--test",
         action="store_true",
         default=False,
-        help="load the smallest model for simple testing",
+        help="load the smallest model for simple testing (ethzanalytics/distilgpt2-tiny-conversational)",
     )
 
     return parser
@@ -207,7 +210,7 @@ if __name__ == "__main__":
     gram_model = str(args.gram_model)
     device = 0 if torch.cuda.is_available() else -1
 
-    print(f"CUDA avail is {torch.cuda.is_available()}")
+    logging.info(f"CUDA avail is {torch.cuda.is_available()}")
 
     my_chatbot = (
         pipeline("text-generation", model=model_loc.resolve(), device=device)
@@ -218,12 +221,12 @@ if __name__ == "__main__":
 
     if basic_sc:
         print("Using the baseline spellchecker")
-        schnellspell = build_symspell_obj()
+        basic_spell = build_symspell_obj()
     else:
         print("using neural spell checker")
         grammarbot = pipeline("text2text-generation", gram_model, device=device)
 
-    print(f"using model stored here: \n {model_loc} \n")
+    logging.info(f"using model stored here: \n {model_loc} \n")
     iface = gr.Interface(
         chat,
         inputs=[
@@ -238,7 +241,11 @@ if __name__ == "__main__":
             ),
             Slider(minimum=0.0, maximum=1.0, step=0.01, default=0.95, label="top_p"),
             Slider(minimum=0, maximum=100, step=5, default=20, label="top_k"),
-            Radio(choices=["True", "False"], default="False", label="constrained_generation"),
+            Radio(
+                choices=["True", "False"],
+                default="False",
+                label="constrained_generation",
+            ),
         ],
         outputs="html",
         examples_per_page=8,

converse.py

@@ -23,7 +23,7 @@ def discussion(
     responder: str,
     pipeline,
     timeout=45,
-    min_length=4,
+    min_length=8,
     max_length=64,
     top_p=0.95,
     top_k=50,
@@ -60,6 +60,8 @@ def discussion(
         str, the generated text
     """
 
+    logging.debug(f"input args: {locals()}")
+
     p_list = []  # track conversation
     p_list.append(speaker.lower() + ":" + "\n")
     p_list.append(prompt_text.lower() + "\n")
@@ -75,6 +77,8 @@ def discussion(
         response = constrained_generation(
             prompt=this_prompt,
             pipeline=pipeline,
+            min_generated_tokens=min_length,
+            max_generated_tokens=max_length,
             no_repeat_ngram_size=no_repeat_ngram_size,
             length_penalty=length_penalty,
             repetition_penalty=1.0,
@@ -101,6 +105,7 @@ def discussion(
             speaker,
             responder,
             timeout=timeout,
+            min_length=min_length,
             max_length=max_length,
             top_p=top_p,
             top_k=top_k,
@@ -112,6 +117,7 @@ def discussion(
             device=device,
             verbose=verbose,
         )
+    logging.debug(f"generation done. bot_dialogue: {bot_dialogue}")
     if isinstance(bot_dialogue, list) and len(bot_dialogue) > 1:
         bot_resp = ", ".join(bot_dialogue)
     elif isinstance(bot_dialogue, list) and len(bot_dialogue) == 1:
@@ -123,12 +129,12 @@ def discussion(
     # remove the last ',' '.' chars
     bot_resp = remove_trailing_punctuation(bot_resp)
     if verbose:
+        print("\nfinished!")
         print("\n... bot response:\n")
         pp.pprint(bot_resp)
     p_list.append(bot_resp + "\n")
     p_list.append("\n")
 
-    print("\nfinished!")
     logging.info(f"finished generating response:\n\t{bot_resp}")
     # return the bot response and the full conversation