14

wyk/14_pretrenowanie.ipynb

@@ -0,0 +1,338 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Pretrenowanie modeli\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "System AlphaZero uczy się grając sam ze sobą — wystarczy 24 godziny,\n",
+    "by system nauczył się grać w szachy lub go na nadludzkim poziomie.\n",
+    "\n",
+    "**Pytanie**: Dlaczego granie samemu ze sobą nie jest dobrym sposobem\n",
+    " nauczenia się grania w szachy dla człowieka, a dla maszyny jest?\n",
+    "\n",
+    "Co jest odpowiednikiem grania samemu ze sobą w świecie przetwarzania tekstu?\n",
+    "Tzn. **pretrenowanie** (*pretraining*) na dużym korpusie tekstu. (Tekst jest tani!)\n",
+    "\n",
+    "Jest kilka sposobów na pretrenowanie modelu, w każdym razie sprowadza\n",
+    "się do odgadywania następnego bądź zamaskowanego słowa.\n",
+    "W każdym razie zawsze stosujemy softmax (być może ze „sztuczkami” takimi jak\n",
+    "negatywne próbkowanie albo hierarchiczny softamx) na pewnej **representecji kontekstowej**:\n",
+    "\n",
+    "$$\\vec{p} = \\operatorname{softmax}(f(\\vec{c})).$$\n",
+    "\n",
+    "Model jest karany używając funkcji log loss:\n",
+    "\n",
+    "$$-\\log(p_j),$$\n",
+    "\n",
+    "gdzie $w_j$ jest wyrazem, który pojawił się rzeczywiście w korpusie.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Przewidywanie słowa (GPT-2)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Jeden ze sposobów pretrenowania modelu to po prostu przewidywanie\n",
+    "następnego słowa.\n",
+    "\n",
+    "Zainstalujmy najpierw bibliotekę transformers.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "! pip install transformers"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "50257\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "[('Ġon', 0.6786560416221619),\n",
+       " ('Ġupon', 0.04339785501360893),\n",
+       " ('Ġheavily', 0.02208443358540535),\n",
+       " ('Ġin', 0.021049050614237785),\n",
+       " (',', 0.020188499242067337),\n",
+       " ('Ġa', 0.01833895780146122),\n",
+       " ('Ġvery', 0.017935041338205338),\n",
+       " ('Ġentirely', 0.017528969794511795),\n",
+       " ('Ġlargely', 0.016769640147686005),\n",
+       " ('Ġto', 0.01009418722242117),\n",
+       " ('Ġgreatly', 0.010009866207838058),\n",
+       " ('Ġnot', 0.009016563184559345),\n",
+       " ('Ġmore', 0.005853226874023676),\n",
+       " ('Ġprimarily', 0.005203146021813154),\n",
+       " ('Ġstrongly', 0.0034501152113080025),\n",
+       " ('Ġpartly', 0.0033184229396283627),\n",
+       " ('Ġmuch', 0.0033095215912908316),\n",
+       " ('Ġmostly', 0.0032150144688785076),\n",
+       " ('Ġmainly', 0.0030899408739060163),\n",
+       " ('Ġfor', 0.003034428460523486),\n",
+       " ('.', 0.0028878094162791967),\n",
+       " ('Ġboth', 0.0028405177872627974),\n",
+       " ('Ġsomewhat', 0.0028194624464958906),\n",
+       " ('Ġcru', 0.002263976726680994),\n",
+       " ('Ġas', 0.00221616611815989),\n",
+       " ('Ġof', 0.0022000609897077084),\n",
+       " ('Ġalmost', 0.001968063646927476),\n",
+       " ('Ġat', 0.0018015997484326363),\n",
+       " ('Ġhighly', 0.0017461496172472835),\n",
+       " ('Ġcompletely', 0.001692073536105454)]"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import torch\n",
+    "from transformers import GPT2Tokenizer, GPT2LMHeadModel\n",
+    "tokenizer = GPT2Tokenizer.from_pretrained('gpt2-large')\n",
+    "model = GPT2LMHeadModel.from_pretrained('gpt2-large')\n",
+    "text = \"This issue depends\"\n",
+    "encoded_input = tokenizer(text, return_tensors='pt')\n",
+    "output = model(**encoded_input)\n",
+    "next_token_probs = torch.softmax(output[0][:, -1, :][0], dim=0)\n",
+    "\n",
+    "next_token_probs\n",
+    "nb_of_tokens = next_token_probs.size()[0]\n",
+    "print(nb_of_tokens)\n",
+    "\n",
+    "_, top_k_indices = torch.topk(next_token_probs, 30, sorted=True)\n",
+    "\n",
+    "words = tokenizer.convert_ids_to_tokens(top_k_indices)\n",
+    "\n",
+    "top_probs = []\n",
+    "\n",
+    "for ix in range(len(top_k_indices)):\n",
+    "     top_probs.append((words[ix], next_token_probs[top_k_indices[ix]].item()))\n",
+    "\n",
+    "top_probs"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Zalety tego podejścia:\n",
+    "\n",
+    "-   prostota,\n",
+    "-   dobra podstawa do strojenia systemów generowania tekstu zwłaszcza\n",
+    "    „otwartego” (systemy dialogowe, generowanie (fake) newsów, streszczanie tekstu),\n",
+    "    ale niekoniecznie tłumaczenia maszynowego,\n",
+    "-   zaskakująca skuteczność przy uczeniu *few-shot* i *zero-shot*.\n",
+    "\n",
+    "Wady:\n",
+    "\n",
+    "-   asymetryczność, przetwarzanie tylko z lewej do prawej, preferencja\n",
+    "    dla lewego kontekstu,\n",
+    "-   mniejsza skuteczność przy dostrajaniu do zadań klasyfikacji i innych zadań\n",
+    "    niepolegających na prostym generowaniu.\n",
+    "\n",
+    "Przykłady modeli: GPT, GPT-2, GPT-3, DialoGPT.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Maskowanie słów (BERT)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Inną metodą jest maskowanie słów (*Masked Language Modeling*, *MLM*).\n",
+    "\n",
+    "W tym podejściu losowe wybrane zastępujemy losowe słowa specjalnym\n",
+    "tokenem (`[MASK]`) i każemy modelowi odgadywać w ten sposób\n",
+    "zamaskowane słowa (z uwzględnieniem również prawego kontekstu!).\n",
+    "\n",
+    "Móciąc ściśle, w jednym z pierwszych modeli tego typu (BERT)\n",
+    "zastosowano schemat, w którym również niezamaskowane słowa są odgadywane (!):\n",
+    "\n",
+    "-   wybieramy losowe 15% wyrazów do odgadnięcia\n",
+    "-   80% z nich zastępujemy tokenem `[MASK]`,\n",
+    "-   10% zastępujemy innym losowym wyrazem,\n",
+    "-   10% pozostawiamy bez zmian.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "# Out[3]:"
+     ]
+    }
+   ],
+   "source": [
+    "from transformers import AutoModelWithLMHead, AutoTokenizer\n",
+    "import torch\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\"xlm-roberta-large\")\n",
+    "model = AutoModelWithLMHead.from_pretrained(\"xlm-roberta-large\")\n",
+    "\n",
+    "sequence = f'II wojna światowa zakończyła się w {tokenizer.mask_token} roku.'\n",
+    "\n",
+    "input_ids = tokenizer.encode(sequence, return_tensors=\"pt\")\n",
+    "mask_token_index = torch.where(input_ids == tokenizer.mask_token_id)[1]\n",
+    "\n",
+    "token_logits = model(input_ids)[0]\n",
+    "mask_token_logits = token_logits[0, mask_token_index, :]\n",
+    "mask_token_logits = torch.softmax(mask_token_logits, dim=1)\n",
+    "\n",
+    "top_10 = torch.topk(mask_token_logits, 10, dim=1)\n",
+    "top_10_tokens = zip(top_10.indices[0].tolist(), top_10.values[0].tolist())\n",
+    "\n",
+    "for token, score in top_10_tokens:\n",
+    "    print(sequence.replace(tokenizer.mask_token, tokenizer.decode([token])), f\"(score: {score})\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Przykłady: BERT, RoBERTa (również Polish RoBERTa).\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Podejście generatywne (koder-dekoder).\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "System ma wygenerować odpowiedź na różne pytania (również\n",
+    "odpowiadające zadaniu MLM), np.:\n",
+    "\n",
+    "-   \"translate English to German: That is good.\" => \"Das ist gut.\"\n",
+    "-   \"cola sentence: The course is jumping well.\" => \"not acceptable\"\n",
+    "-   \"summarize: state authorities dispatched emergency crews tuesday to survey the damage after an onslaught of severe weather in mississippi…\"\n",
+    "    => \"six people hospitalized after a storm in attala county\"\n",
+    "-   \"Thank you for <X> me to your party <Y> week.\" => <X> for inviting <Y> last <Z>\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from transformers import T5Tokenizer, T5Config, T5ForConditionalGeneration\n",
+    "\n",
+    "T5_PATH = 't5-base'\n",
+    "\n",
+    "t5_tokenizer = T5Tokenizer.from_pretrained(T5_PATH)\n",
+    "t5_config = T5Config.from_pretrained(T5_PATH)\n",
+    "t5_mlm = T5ForConditionalGeneration.from_pretrained(T5_PATH, config=t5_config)\n",
+    "\n",
+    "slot = '<extra_id_0>'\n",
+    "\n",
+    "text = f'Warsaw is the {slot} of Poland.'\n",
+    "\n",
+    "encoded = t5_tokenizer.encode_plus(text, add_special_tokens=True, return_tensors='pt')\n",
+    "input_ids = encoded['input_ids']\n",
+    "\n",
+    "outputs = t5_mlm.generate(input_ids=input_ids,\n",
+    "                          num_beams=200, num_return_sequences=5,\n",
+    "                          max_length=5)\n",
+    "\n",
+    "_0_index = text.index(slot)\n",
+    "_result_prefix = text[:_0_index]\n",
+    "_result_suffix = text[_0_index+len(slot):]\n",
+    "\n",
+    "def _filter(output, end_token='<extra_id_1>'):\n",
+    "    _txt = t5_tokenizer.decode(output[2:], skip_special_tokens=False, clean_up_tokenization_spaces=False)\n",
+    "    if end_token in _txt:\n",
+    "        _end_token_index = _txt.index(end_token)\n",
+    "        return _result_prefix + _txt[:_end_token_index] + _result_suffix\n",
+    "    else:\n",
+    "        return _result_prefix + _txt + _result_suffix\n",
+    "\n",
+    "\n",
+    "results = [_filter(out) for out in outputs]\n",
+    "results"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "(Zob. [https://arxiv.org/pdf/1910.10683.pdf](https://arxiv.org/pdf/1910.10683.pdf))\n",
+    "\n",
+    "Przykład: T5, mT5\n",
+    "\n"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.2"
+  },
+  "org": null
+ },
+ "nbformat": 4,
+ "nbformat_minor": 1
+}

wyk/14_pretrenowanie.org

@@ -132,10 +132,10 @@ zastosowano schemat, w którym również niezamaskowane słowa są odgadywane (!
 from transformers import AutoModelWithLMHead, AutoTokenizer
 import torch
 
-tokenizer = AutoTokenizer.from_pretrained("distilroberta-base")
+tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
-model = AutoModelWithLMHead.from_pretrained("distilroberta-base")
+model = AutoModelWithLMHead.from_pretrained("xlm-roberta-large")
 
-sequence = f"Hugging Face is a French company based in {tokenizer.mask_token}"
+sequence = f'II wojna światowa zakończyła się w {tokenizer.mask_token} roku.'
 
 input_ids = tokenizer.encode(sequence, return_tensors="pt")
 mask_token_index = torch.where(input_ids == tokenizer.mask_token_id)[1]
@@ -144,12 +144,69 @@ token_logits = model(input_ids)[0]
 mask_token_logits = token_logits[0, mask_token_index, :]
 mask_token_logits = torch.softmax(mask_token_logits, dim=1)
 
-top_5 = torch.topk(mask_token_logits, 5, dim=1)
+top_10 = torch.topk(mask_token_logits, 10, dim=1)
-top_5_tokens = zip(top_5.indices[0].tolist(), top_5.values[0].tolist())
+top_10_tokens = zip(top_10.indices[0].tolist(), top_10.values[0].tolist())
 
-for token, score in top_5_tokens:
+for token, score in top_10_tokens:
     print(sequence.replace(tokenizer.mask_token, tokenizer.decode([token])), f"(score: {score})")
 #+END_SRC
 
+#+RESULTS:
+:results:
+# Out[3]:
+:end:
+
 
 Przykłady: BERT, RoBERTa (również Polish RoBERTa).
+
+** Podejście generatywne (koder-dekoder).
+
+System ma wygenerować odpowiedź na różne pytania (również
+odpowiadające zadaniu MLM), np.:
+
+- "translate English to German: That is good." => "Das ist gut."
+- "cola sentence: The course is jumping well." => "not acceptable"
+- "summarize: state authorities dispatched emergency crews tuesday to survey the damage after an onslaught of severe weather in mississippi..."
+  => "six people hospitalized after a storm in attala county"
+- "Thank you for <X> me to your party <Y> week." => <X> for inviting <Y> last <Z>
+
+#+BEGIN_SRC ipython :session mysession :exports both :results raw drawer
+from transformers import T5Tokenizer, T5Config, T5ForConditionalGeneration
+
+T5_PATH = 't5-base'
+
+t5_tokenizer = T5Tokenizer.from_pretrained(T5_PATH)
+t5_config = T5Config.from_pretrained(T5_PATH)
+t5_mlm = T5ForConditionalGeneration.from_pretrained(T5_PATH, config=t5_config)
+
+slot = '<extra_id_0>'
+
+text = f'Warsaw is the {slot} of Poland.'
+
+encoded = t5_tokenizer.encode_plus(text, add_special_tokens=True, return_tensors='pt')
+input_ids = encoded['input_ids']
+
+outputs = t5_mlm.generate(input_ids=input_ids,
+                          num_beams=200, num_return_sequences=5,
+                          max_length=5)
+
+_0_index = text.index(slot)
+_result_prefix = text[:_0_index]
+_result_suffix = text[_0_index+len(slot):]
+
+def _filter(output, end_token='<extra_id_1>'):
+    _txt = t5_tokenizer.decode(output[2:], skip_special_tokens=False, clean_up_tokenization_spaces=False)
+    if end_token in _txt:
+        _end_token_index = _txt.index(end_token)
+        return _result_prefix + _txt[:_end_token_index] + _result_suffix
+    else:
+        return _result_prefix + _txt + _result_suffix
+
+
+results = [_filter(out) for out in outputs]
+results
+#+END_SRC
+
+(Zob. https://arxiv.org/pdf/1910.10683.pdf)
+
+Przykład: T5, mT5