فصل مربوط به شبکههای بازگشتی (RNN) این هفته به پایان رسید. یک هفته استراحت خواهیم داشت و پس از آن وارد فصل سوم دوره (شبکههای کانولوشنی) میشویم.
در فصل دوم، یکی از پروژه هایی که با استفاده از RNNها انجام دادیم تولید متن بود. این شبکه ها از اولین مدلهایی بودند که برای اینکار استفاده میشدند. با این حال، RNNها محدودیتهایی داشتند و در تولید متن عملکرد چندان دقیق و کارآمدی نداشتند که بعدها با معرفی مکانیزم توجه و سپس مدلهای مبتنی بر ترنسفورمرها، یک تحول بزرگ در تولید محتوا رخ داد و نتیجه آن مدلهای قدرتمندی مانند ChatGPTاست.
✅ در این پروژه، یک شبکهی سبک LSTM (با تنها یک لایه بازگشتی و تعداد نورون محدود) رو روی یک دیتاست بسیار کوچک آموزش دادیم. سپس با استفاده از مدل آموزش دیده، چند نمونه متن به ازای مقادیر مختلف temperature تولید کردیم که خروجی جالبی داشتند.
پارامتر temperature که از مکانیک آماری قرض گرفته شده، در مدلهای زبانی برای کنترل میزان خلاقیت، ریسک پذیری و تنوع در انتخاب کلمات استفاده میشه. مقادیر پایینتر، خروجی را محافظه کارانه، تکراری و قابل پیش بینی میکنند. و در عوض مقادیر بالاتر، باعث تولید متنهای خلاقانه تر و غیرقابل پیش بینی تر میشوند
در ادامه چند نمونه از متنهای تولید شده را با شما به اشتراک میگذارم:
Prime text: i like to say
Temperature: 0.99
i like to say my word its time. how i should have to tell you. but they are married who were in any fix they mostly went to badger the dangerous look at the beginning of the stone own under a tree not long to be a leader. he left we thus with sid words that no star is a continual feast. little twoeyes. here we thank you lady mary so did she get a little ashamed of the questions who lived in that way. she said. he came down and carried it quickly about her. asked beth anxiously. so she was an odd merchant seeing that his son might choose. then they went and stood before the door and the one kept a wonderful one. then the kings daughter came with the cattle and went on his way …
Prime text: i like to say
Temperature: 0.1
i like to say my word its being rather dirty.
but i cant tell you how we can manage.
i am glad you think that i was a ghost or a man.
i want to be able to find it.
he hesitated.
he forgot all about the ball.
i want to lift my whole in the world to fill such a thing or two.
my brother say a name i have killed many times.
ye have called me a silly ass.
i was dumb.
i hadnt a bit of it out of the window…
Prime text: the girl
Temperature: 0.3
the girl was afraid of the beautiful character. the poor woman declared with a little milk at once with a ball of dry kind. she began to beat her with a great storm as a sacrifice. the tin woodman followed with the swallow and fell asleep. when she saw him turning she threw her into the basin. she was so beautiful that he hoped it would be wise to him. but he never told anybody except that they should not come. so long after wealth and darling lived only a few hours earlier to …
Prime text: i love to
Temperature: 0.3
i love to hunt my husband.
it is a pity because i am a capital fellow.
i dont know how to read it.
i have never seen him before.
i dont know how to give him to mine.
i have been married behind.
i am a traveler please excuse me.
i will not tell you how to behave.
but you must not talk about that.
it is a pity because i am a smart toad and i dont know.
i was sweeping my room my lady my lady i bought my mates in...
💡در فصل چهارم مدلهای کارآمدتری بررسی خواهیم کرد😊
🏢 آکادمی آنلاین مهندسی پزشکی و هوش مصنوعی
@Onlinebme