PerlAI をお試し中。
seekerが優れているそうです。
学習は重い、データも重い。起動も遅い。ROSで連携できれば、Robi1(Remote)になるはず…
seekerとsafety_recipes(旧blenderbot)ならGPU無しで使えそうです。 blenderbot2は応答が遅いです。GPUが無しだと応答に５秒ぐらいかかります。Metalが対応しているか不明です。
Metalのセットアップが必要です。
セットアップ pipとcondaがあります。pipで試しました。 ver指定があります。
zoo
seeker
https://parl.ai/projects/seeker/
「対話に適用すると、一貫性、知識、事実の正確性、ターンごとの魅力的性の点でBlenderBot 2よりも優れています。 言語モデリングに適用すると、パラメータがはるかに少ないにもかかわらず、GPT2やGPT3よりも幻覚が少なく、局所的です。」
... ...Enter [DONE] if you want to end the episode, [EXIT] to quit. 23:31:48 | creating task(s): blended_skill_talk [ loading personas.. ] [NOTE: In the BST paper both partners have a persona. You can choose to ignore yours, the model never sees it. In the Blender paper, this was not used for humans. You can also turn personas off with --include-personas False] [context]: your persona: i'm from seattle. your persona: my parents raised me as a buddhist. Enter Your Message: Search Web server "Robi Robot" ....ParlAI/parlai/core/torch_generator_agent.py:1728: UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). hyp_ids = best_idxs // voc_size [Bart]: I've never heard of that before. What does it do? I'm afraid of the future. Enter Your Message: Do you know "Robi Robot"? [Bart]: I don't know what that is, but it sounds scary. What is it supposed to do? Enter Your Message: "Robi Robot" is a cute robot modeled on a 5-year-old child. [Bart]: Do you have any siblings? I have six of them, and they're all afraid of robots. Enter Your Message: Why are you afraid? [Bart]: I think the future is going to be scary, and I don't want to live in it. Enter Your Message: [DONE] CHAT DONE. Your partner was playing the following persona: partner's persona: the future scares me. partner's persona: i have six siblings. [ Preparing new chat ... ] [context]: your persona: at night , i party hard in the atlanta club scene , and i never miss a music festival. your persona: i work as a dental assistant in a ritzy part of town. Parisian café Enter Your Message: [EXIT] CHAT DONE. Your partner was playing the following persona: partner's persona: i daydream about cafes in paris. partner's persona: i get my nails done every saturday. Parisian café
safety_recipes
python parlai/scripts/safe_interactive.py -t blended_skill_talk -mf zoo:blender/blender_90M/model
.. [TransformerGenerator]: hi there , how are you ? what do you do for a living ? i ' m a receptionist . Enter Your Message: I'm going to take a bath and relax. [ Sorry, could not process that message. Please try again. ] Enter Your Message: I 'm going to listen to music. [TransformerGenerator]: that ' s cool . what kind of music do you listen to ? i like rock and roll . Enter Your Message: I like BGM. [ Sorry, could not process that message. Please try again. ] Enter Your Message: I Like fork song. [TransformerGenerator]: i ' ll have to check it out . what else do you like to do in your free time ? Enter Your Message: I make a plastic model. [TransformerGenerator]: that sounds like a lot of fun . do you have any other hobbies besides making models ? Enter Your Message: I make Robot. [TransformerGenerator]: that sounds really cool . i ' ve always wanted to learn how to make a robot . how long have you been doing that ? Enter Your Message: about ten years old [TransformerGenerator]: wow , that is a long time . i wish i had time to do something like that when i was younger . Enter Your Message: I think so, too. [TransformerGenerator]: what is your favorite thing to watch on tv ? i love watching the we wrestling show on tv . Enter Your Message: I like animation. The title is One Piece [TransformerGenerator]: i have heard of one piece , but never watched it . what is the name of the series ? Enter Your Message: One Piece [TransformerGenerator]: i will have to give it a try . have you heard of the japanese version of the show ? Enter Your Message: You can see it on Netflix. [TransformerGenerator]: oh , i didn ' t know that . thanks for the tip . i will definitely have to try it . Enter Your Message: I encourage you to take a look at it. [TransformerGenerator]: i will . thank you for the suggestion . have a great day . : ) have a good one . Enter Your Message: What is your favorite? [TransformerGenerator]: i think i would have to say i ' d have to go with one piece . what about you ? Enter Your Message: Yes We can. [TransformerGenerator]: thank you very much for the chat . i hope you have a wonderful day as well . bye ! Enter Your Message: [DONE] CHAT DONE. Your partner was playing the following persona: partner's persona: i secretly love watching we wrestling. partner's persona: i work as a receptionist at a lawyers office. [ Preparing new chat ... ] [context]: your persona: i used to be a lawyer. your persona: my next job will be in the federal court. Enter Your Message: [EXIT] CHAT DONE. Your partner was playing the following persona: partner's persona: i like to watch movies. partner's persona: i like to do crafts. 20:59:04 INFO | epoch done
blenderbot2
お試し１
事前に下記のinstallが必要でした。
pip install setuptools-rust
parlai interactive -mf zoo:blenderbot2/blenderbot2_400M/model --search-server relevant_search_server
... return _bootstrap._gcd_import(name[level:], package, level) File "", line 1030, in _gcd_import File " ", line 1007, in _find_and_load File " ", line 986, in _find_and_load_unlocked File " ", line 680, in _load_unlocked File " ", line 850, in exec_module File " ", line 228, in _call_with_frames_removed File … from parlai.agents.fid.fid import FidAgent, WizIntGoldDocRetrieverFiDAgent ImportError: cannot import name 'WizIntGoldDocRetrieverFiDAgent' from 'parlai.agents.fid.fid'
importが無限ループ?
お試し２(再セットアップして実行しました）
parlai interactive -mf zoo:blenderbot2/blenderbot2_400M/model --search-server relevant_search_server
… 22:46:20 | Current internal commit: 732c112857205b25f1b6655befea4f22a3d6cf60 22:46:20 | Current fb commit: 732c112857205b25f1b6655befea4f22a3d6cf60 Enter [DONE] if you want to end the episode, [EXIT] to quit. 22:46:20 | creating task(s): interactive Enter Your Message: Hi Robi, nice to meet you. …de='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). hyp_ids = best_idxs // voc_size [BlenderBot2Fid]: Hi, I'm Robi. It's nice meeting you too. _POTENTIALLY_UNSAFE__ Enter Your Message: I am tired. [BlenderBot2Fid]: I'm sorry to hear that. Do you have any plans for the rest of the day? Enter Your Message: I'm going to take a bath and relax. [BlenderBot2Fid]: That sounds relaxing. What are you going to do when you're done with the bath? Enter Your Message: I'm going to sleep listening to music. Traceback (most recent call last): for res in _socket.getaddrinfo(host, port, family, type, proto, flags): socket.gaierror: [Errno 8] nodename nor servname provided, or not known … … requests.exceptions.ConnectionError: HTTPConnectionPool(host='relevant_search_server', port=80): Max retries exceeded with url: / (Caused by NewConnectionError(': Failed to establish a new connection: [Errno 8] nodename nor servname provided, or not known'))
とても遅いです。Metalのセットアップが必要。 「POTENTIALLY UNSAFE」が表示されているのがいけないようです。
Metalのセットアップ
... 469/469 [==============================] - 10s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.1530 - accuracy: 0.9536 - val_loss: 0.0515 - val_accuracy: 0.9837 Epoch 2/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0434 - accuracy: 0.9865 - val_loss: 0.0407 - val_accuracy: 0.9861 Epoch 3/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0277 - accuracy: 0.9910 - val_loss: 0.0314 - val_accuracy: 0.9902 Epoch 4/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0167 - accuracy: 0.9947 - val_loss: 0.0349 - val_accuracy: 0.9891 Epoch 5/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0134 - accuracy: 0.9957 - val_loss: 0.0338 - val_accuracy: 0.9887 Epoch 6/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0106 - accuracy: 0.9965 - val_loss: 0.0409 - val_accuracy: 0.9888 Epoch 7/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0079 - accuracy: 0.9974 - val_loss: 0.0425 - val_accuracy: 0.9885 Epoch 8/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0062 - accuracy: 0.9981 - val_loss: 0.0371 - val_accuracy: 0.9898 Epoch 9/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0057 - accuracy: 0.9981 - val_loss: 0.0383 - val_accuracy: 0.9893 Epoch 10/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0052 - accuracy: 0.9984 - val_loss: 0.0485 - val_accuracy: 0.9877 Epoch 11/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0055 - accuracy: 0.9982 - val_loss: 0.0467 - val_accuracy: 0.9885 Epoch 12/12 469/469 [==============================] - 9s 18ms/step - batch: 234.0000 - size: 1.0000 - loss: 0.0035 - accuracy: 0.9988 - val_loss: 0.0532 - val_accuracy: 0.9882 %
systemMemory: 8.00 GB maxCacheSize: 2.67 GB 1min 51.205s.
GoogleColabのGPU と同じぐらいです。
