{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "2fd69392", "metadata": {}, "outputs": [], "source": [ "# Tutorial url\n", "# https://medium.com/data-and-beyond/complete-guide-to-building-bert-model-from-sratch-3e6562228891" ] }, { "cell_type": "code", "execution_count": 1, "id": "7bc1129e", "metadata": { "scrolled": true }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Requirement already satisfied: transformers in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (4.24.0)\n", "Collecting datasets\n", " Downloading datasets-2.13.1-py3-none-any.whl (486 kB)\n", " ------------------------------------- 486.2/486.2 kB 10.1 MB/s eta 0:00:00\n", "Requirement already satisfied: tokenizers in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (0.11.4)\n", "Requirement already satisfied: packaging>=20.0 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (22.0)\n", "Requirement already satisfied: tqdm>=4.27 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (4.64.1)\n", "Requirement already satisfied: numpy>=1.17 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (1.23.5)\n", "Requirement already satisfied: huggingface-hub<1.0,>=0.10.0 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (0.10.1)\n", "Requirement already satisfied: pyyaml>=5.1 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (6.0)\n", "Requirement already satisfied: requests in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (2.28.1)\n", "Requirement already satisfied: filelock in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (3.9.0)\n", "Requirement already satisfied: regex!=2019.12.17 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from transformers) (2022.7.9)\n", "Collecting huggingface-hub<1.0,>=0.10.0\n", " Downloading huggingface_hub-0.16.2-py3-none-any.whl (268 kB)\n", " ------------------------------------- 268.5/268.5 kB 16.1 MB/s eta 0:00:00\n", "Requirement already satisfied: pandas in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from datasets) (1.5.3)\n", "Collecting aiohttp\n", " Downloading aiohttp-3.8.4-cp310-cp310-win_amd64.whl (319 kB)\n", " ---------------------------------------- 319.8/319.8 kB ? eta 0:00:00\n", "Requirement already satisfied: fsspec[http]>=2021.11.1 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from datasets) (2022.11.0)\n", "Collecting xxhash\n", " Downloading xxhash-3.2.0-cp310-cp310-win_amd64.whl (30 kB)\n", "Collecting pyarrow>=8.0.0\n", " Downloading pyarrow-12.0.1-cp310-cp310-win_amd64.whl (21.5 MB)\n", " --------------------------------------- 21.5/21.5 MB 26.1 MB/s eta 0:00:00\n", "Collecting multiprocess\n", " Downloading multiprocess-0.70.14-py310-none-any.whl (134 kB)\n", " -------------------------------------- 134.3/134.3 kB 8.3 MB/s eta 0:00:00\n", "Requirement already satisfied: dill<0.3.7,>=0.3.0 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from datasets) (0.3.6)\n", "Requirement already satisfied: charset-normalizer<4.0,>=2.0 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from aiohttp->datasets) (2.0.4)\n", "Collecting yarl<2.0,>=1.0\n", " Downloading yarl-1.9.2-cp310-cp310-win_amd64.whl (61 kB)\n", " ---------------------------------------- 61.0/61.0 kB ? eta 0:00:00\n", "Collecting aiosignal>=1.1.2\n", " Downloading aiosignal-1.3.1-py3-none-any.whl (7.6 kB)\n", "Collecting frozenlist>=1.1.1\n", " Downloading frozenlist-1.3.3-cp310-cp310-win_amd64.whl (33 kB)\n", "Collecting async-timeout<5.0,>=4.0.0a3\n", " Downloading async_timeout-4.0.2-py3-none-any.whl (5.8 kB)\n", "Collecting multidict<7.0,>=4.5\n", " Downloading multidict-6.0.4-cp310-cp310-win_amd64.whl (28 kB)\n", "Requirement already satisfied: attrs>=17.3.0 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from aiohttp->datasets) (22.1.0)\n", "Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from huggingface-hub<1.0,>=0.10.0->transformers) (4.4.0)\n", "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from requests->transformers) (3.4)\n", "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from requests->transformers) (1.26.14)\n", "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from requests->transformers) (2022.12.7)\n", "Requirement already satisfied: colorama in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from tqdm>=4.27->transformers) (0.4.6)\n", "Requirement already satisfied: python-dateutil>=2.8.1 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from pandas->datasets) (2.8.2)\n", "Requirement already satisfied: pytz>=2020.1 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from pandas->datasets) (2022.7)\n", "Requirement already satisfied: six>=1.5 in c:\\users\\yozhan\\appdata\\local\\anaconda3\\lib\\site-packages (from python-dateutil>=2.8.1->pandas->datasets) (1.16.0)\n", "Installing collected packages: xxhash, pyarrow, multiprocess, multidict, frozenlist, async-timeout, yarl, huggingface-hub, aiosignal, aiohttp, datasets\n", " Attempting uninstall: huggingface-hub\n", " Found existing installation: huggingface-hub 0.10.1\n", " Uninstalling huggingface-hub-0.10.1:\n", " Successfully uninstalled huggingface-hub-0.10.1\n", "Successfully installed aiohttp-3.8.4 aiosignal-1.3.1 async-timeout-4.0.2 datasets-2.13.1 frozenlist-1.3.3 huggingface-hub-0.16.2 multidict-6.0.4 multiprocess-0.70.14 pyarrow-12.0.1 xxhash-3.2.0 yarl-1.9.2\n" ] } ], "source": [ "!pip install transformers datasets tokenizers" ] }, { "cell_type": "code", "execution_count": 2, "id": "c23bad9a", "metadata": {}, "outputs": [], "source": [ "import os\n", "os.chdir(r\"C:\\Users\\yozhan\\cryptocurrency\\tutorial\")" ] }, { "cell_type": "code", "execution_count": 4, "id": "d6d279f1", "metadata": {}, "outputs": [], "source": [ "import os\n", "from pathlib import Path\n", "import torch\n", "import re\n", "import random\n", "import transformers, datasets\n", "from tokenizers import BertWordPieceTokenizer\n", "from transformers import BertTokenizer\n", "import tqdm\n", "from torch.utils.data import Dataset, DataLoader\n", "import itertools\n", "import math\n", "import torch.nn.functional as F\n", "import numpy as np\n", "from torch.optim import Adam\n", "\n", "MAX_LEN = 64" ] }, { "cell_type": "code", "execution_count": 8, "id": "af0bf9cb", "metadata": {}, "outputs": [], "source": [ "### loading all data into memory\n", "corpus_movie_conv = r'.\\cornell_movie-dialogs_corpus\\movie_conversations.txt'\n", "corpus_movie_lines = '.\\cornell_movie-dialogs_corpus\\movie_lines.txt'\n", "with open(corpus_movie_conv, 'r', encoding='iso-8859-1') as c:\n", " conv = c.readlines()\n", "with open(corpus_movie_lines, 'r', encoding='iso-8859-1') as l:\n", " lines = l.readlines()\n", "\n", "### splitting text using special lines\n", "lines_dic = {}\n", "for line in lines:\n", " objects = line.split(\" +++$+++ \")\n", " lines_dic[objects[0]] = objects[-1]\n", "\n", "### generate question answer pairs\n", "pairs = []\n", "for con in conv:\n", " \n", " # get a list of sentence ids\n", " ids = eval(con.split(\" +++$+++ \")[-1])\n", " for i in range(len(ids)):\n", " qa_pairs = []\n", " \n", " # if this is the last id\n", " if i == len(ids) - 1:\n", " break\n", "\n", " first = lines_dic[ids[i]].strip() \n", " second = lines_dic[ids[i+1]].strip() \n", "\n", " qa_pairs.append(' '.join(first.split()[:MAX_LEN]))\n", " qa_pairs.append(' '.join(second.split()[:MAX_LEN]))\n", " pairs.append(qa_pairs)" ] }, { "cell_type": "code", "execution_count": 17, "id": "c9f5a78d", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "['Can we make this quick? Roxanne Korrine and Andrew Barrett are having an incredibly horrendous public break- up on the quad. Again.', \"Well, I thought we'd start with pronunciation, if that's okay with you.\"]\n" ] } ], "source": [ "print(pairs[0])" ] }, { "cell_type": "code", "execution_count": 18, "id": "28b4a8b9", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 221616/221616 [00:00<00:00, 1154039.87it/s]\n", "C:\\Users\\yozhan\\AppData\\Local\\anaconda3\\lib\\site-packages\\transformers\\tokenization_utils_base.py:1679: FutureWarning: Calling BertTokenizer.from_pretrained() with the path to a single file or url is deprecated and won't be possible anymore in v5. Use a model identifier or the path to a directory instead.\n", " warnings.warn(\n" ] } ], "source": [ "os.mkdir('./data')\n", "text_data = []\n", "file_count = 0\n", "\n", "for sample in tqdm.tqdm([x[0] for x in pairs]):\n", " text_data.append(sample)\n", "\n", " # once we hit the 10K mark, save to file\n", " if len(text_data) == 10000:\n", " with open(f'./data/text_{file_count}.txt', 'w', encoding='utf-8') as fp:\n", " fp.write('\\n'.join(text_data))\n", " text_data = []\n", " file_count += 1\n", "\n", "paths = [str(x) for x in Path('./data').glob('**/*.txt')]\n", "\n", "### training own tokenizer\n", "tokenizer = BertWordPieceTokenizer(\n", " clean_text=True,\n", " handle_chinese_chars=False,\n", " strip_accents=False,\n", " lowercase=True\n", ")\n", "\n", "tokenizer.train( \n", " files=paths,\n", " vocab_size=30_000, \n", " min_frequency=5,\n", " limit_alphabet=1000, \n", " wordpieces_prefix='##',\n", " special_tokens=['[PAD]', '[CLS]', '[SEP]', '[MASK]', '[UNK]']\n", " )\n", "\n", "os.mkdir('./bert-it-1')\n", "tokenizer.save_model('./bert-it-1', 'bert-it')\n", "tokenizer = BertTokenizer.from_pretrained('./bert-it-1/bert-it-vocab.txt', local_files_only=True)" ] }, { "cell_type": "code", "execution_count": 19, "id": "30e775ad", "metadata": {}, "outputs": [], "source": [ "class BERTDataset(Dataset):\n", " def __init__(self, data_pair, tokenizer, seq_len=64):\n", "\n", " self.tokenizer = tokenizer\n", " self.seq_len = seq_len\n", " self.corpus_lines = len(data_pair)\n", " self.lines = data_pair\n", "\n", " def __len__(self):\n", " return self.corpus_lines\n", "\n", " def __getitem__(self, item):\n", "\n", " # Step 1: get random sentence pair, either negative or positive (saved as is_next_label)\n", " t1, t2, is_next_label = self.get_sent(item)\n", "\n", " # Step 2: replace random words in sentence with mask / random words\n", " t1_random, t1_label = self.random_word(t1)\n", " t2_random, t2_label = self.random_word(t2)\n", "\n", " # Step 3: Adding CLS and SEP tokens to the start and end of sentences\n", " # Adding PAD token for labels\n", " t1 = [self.tokenizer.vocab['[CLS]']] + t1_random + [self.tokenizer.vocab['[SEP]']]\n", " t2 = t2_random + [self.tokenizer.vocab['[SEP]']]\n", " t1_label = [self.tokenizer.vocab['[PAD]']] + t1_label + [self.tokenizer.vocab['[PAD]']]\n", " t2_label = t2_label + [self.tokenizer.vocab['[PAD]']]\n", "\n", " # Step 4: combine sentence 1 and 2 as one input\n", " # adding PAD tokens to make the sentence same length as seq_len\n", " segment_label = ([1 for _ in range(len(t1))] + [2 for _ in range(len(t2))])[:self.seq_len]\n", " bert_input = (t1 + t2)[:self.seq_len]\n", " bert_label = (t1_label + t2_label)[:self.seq_len]\n", " padding = [self.tokenizer.vocab['[PAD]'] for _ in range(self.seq_len - len(bert_input))]\n", " bert_input.extend(padding), bert_label.extend(padding), segment_label.extend(padding)\n", "\n", " output = {\"bert_input\": bert_input,\n", " \"bert_label\": bert_label,\n", " \"segment_label\": segment_label,\n", " \"is_next\": is_next_label}\n", "\n", " return {key: torch.tensor(value) for key, value in output.items()}\n", "\n", " def random_word(self, sentence):\n", " tokens = sentence.split()\n", " output_label = []\n", " output = []\n", "\n", " # 15% of the tokens would be replaced\n", " for i, token in enumerate(tokens):\n", " prob = random.random()\n", "\n", " # remove cls and sep token\n", " token_id = self.tokenizer(token)['input_ids'][1:-1]\n", "\n", " if prob < 0.15:\n", " prob /= 0.15\n", "\n", " # 80% chance change token to mask token\n", " if prob < 0.8:\n", " for i in range(len(token_id)):\n", " output.append(self.tokenizer.vocab['[MASK]'])\n", "\n", " # 10% chance change token to random token\n", " elif prob < 0.9:\n", " for i in range(len(token_id)):\n", " output.append(random.randrange(len(self.tokenizer.vocab)))\n", "\n", " # 10% chance change token to current token\n", " else:\n", " output.append(token_id)\n", "\n", " output_label.append(token_id)\n", "\n", " else:\n", " output.append(token_id)\n", " for i in range(len(token_id)):\n", " output_label.append(0)\n", "\n", " # flattening\n", " output = list(itertools.chain(*[[x] if not isinstance(x, list) else x for x in output]))\n", " output_label = list(itertools.chain(*[[x] if not isinstance(x, list) else x for x in output_label]))\n", " assert len(output) == len(output_label)\n", " return output, output_label\n", "\n", " def get_sent(self, index):\n", " '''return random sentence pair'''\n", " t1, t2 = self.get_corpus_line(index)\n", "\n", " # negative or positive pair, for next sentence prediction\n", " if random.random() > 0.5:\n", " return t1, t2, 1\n", " else:\n", " return t1, self.get_random_line(), 0\n", "\n", " def get_corpus_line(self, item):\n", " '''return sentence pair'''\n", " return self.lines[item][0], self.lines[item][1]\n", "\n", " def get_random_line(self):\n", " '''return random single sentence'''\n", " return self.lines[random.randrange(len(self.lines))][1]" ] }, { "cell_type": "code", "execution_count": 20, "id": "b0d9f35c", "metadata": {}, "outputs": [], "source": [ "train_data = BERTDataset(\n", " pairs, seq_len=MAX_LEN, tokenizer=tokenizer)\n", "train_loader = DataLoader(\n", " train_data, batch_size=32, shuffle=True, pin_memory=True)\n", "sample_data = next(iter(train_loader))" ] }, { "cell_type": "code", "execution_count": 21, "id": "ad60cf79", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "{'bert_input': tensor([ 1, 182, 11, 58, 162, 874, 34, 2, 6, 3232, 108, 512,\n", " 17, 6, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0]), 'bert_label': tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]), 'segment_label': tensor([1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\n", " 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]), 'is_next': tensor(1)}\n" ] } ], "source": [ "print(train_data[random.randrange(len(train_data))])" ] }, { "cell_type": "code", "execution_count": 22, "id": "bd70e96e", "metadata": {}, "outputs": [], "source": [ "class PositionalEmbedding(torch.nn.Module):\n", "\n", " def __init__(self, d_model, max_len=128):\n", " super().__init__()\n", "\n", " # Compute the positional encodings once in log space.\n", " pe = torch.zeros(max_len, d_model).float()\n", " pe.require_grad = False\n", "\n", " for pos in range(max_len): \n", " # for each dimension of the each position\n", " for i in range(0, d_model, 2): \n", " pe[pos, i] = math.sin(pos / (10000 ** ((2 * i)/d_model)))\n", " pe[pos, i + 1] = math.cos(pos / (10000 ** ((2 * (i + 1))/d_model)))\n", "\n", " # include the batch size\n", " self.pe = pe.unsqueeze(0) \n", " # self.register_buffer('pe', pe)\n", "\n", " def forward(self, x):\n", " return self.pe\n", "\n", "class BERTEmbedding(torch.nn.Module):\n", " \"\"\"\n", " BERT Embedding which is consisted with under features\n", " 1. TokenEmbedding : normal embedding matrix\n", " 2. PositionalEmbedding : adding positional information using sin, cos\n", " 2. SegmentEmbedding : adding sentence segment info, (sent_A:1, sent_B:2)\n", " sum of all these features are output of BERTEmbedding\n", " \"\"\"\n", "\n", " def __init__(self, vocab_size, embed_size, seq_len=64, dropout=0.1):\n", " \"\"\"\n", " :param vocab_size: total vocab size\n", " :param embed_size: embedding size of token embedding\n", " :param dropout: dropout rate\n", " \"\"\"\n", "\n", " super().__init__()\n", " self.embed_size = embed_size\n", " # (m, seq_len) --> (m, seq_len, embed_size)\n", " # padding_idx is not updated during training, remains as fixed pad (0)\n", " self.token = torch.nn.Embedding(vocab_size, embed_size, padding_idx=0)\n", " self.segment = torch.nn.Embedding(3, embed_size, padding_idx=0)\n", " self.position = PositionalEmbedding(d_model=embed_size, max_len=seq_len)\n", " self.dropout = torch.nn.Dropout(p=dropout)\n", " \n", " def forward(self, sequence, segment_label):\n", " x = self.token(sequence) + self.position(sequence) + self.segment(segment_label)\n", " return self.dropout(x)" ] }, { "cell_type": "code", "execution_count": 23, "id": "baa5caa0", "metadata": {}, "outputs": [], "source": [ "### attention layers\n", "class MultiHeadedAttention(torch.nn.Module):\n", " \n", " def __init__(self, heads, d_model, dropout=0.1):\n", " super(MultiHeadedAttention, self).__init__()\n", " \n", " assert d_model % heads == 0\n", " self.d_k = d_model // heads\n", " self.heads = heads\n", " self.dropout = torch.nn.Dropout(dropout)\n", "\n", " self.query = torch.nn.Linear(d_model, d_model)\n", " self.key = torch.nn.Linear(d_model, d_model)\n", " self.value = torch.nn.Linear(d_model, d_model)\n", " self.output_linear = torch.nn.Linear(d_model, d_model)\n", " \n", " def forward(self, query, key, value, mask):\n", " \"\"\"\n", " query, key, value of shape: (batch_size, max_len, d_model)\n", " mask of shape: (batch_size, 1, 1, max_words)\n", " \"\"\"\n", " # (batch_size, max_len, d_model)\n", " query = self.query(query)\n", " key = self.key(key) \n", " value = self.value(value) \n", " \n", " # (batch_size, max_len, d_model) --> (batch_size, max_len, h, d_k) --> (batch_size, h, max_len, d_k)\n", " query = query.view(query.shape[0], -1, self.heads, self.d_k).permute(0, 2, 1, 3) \n", " key = key.view(key.shape[0], -1, self.heads, self.d_k).permute(0, 2, 1, 3) \n", " value = value.view(value.shape[0], -1, self.heads, self.d_k).permute(0, 2, 1, 3) \n", " \n", " # (batch_size, h, max_len, d_k) matmul (batch_size, h, d_k, max_len) --> (batch_size, h, max_len, max_len)\n", " scores = torch.matmul(query, key.permute(0, 1, 3, 2)) / math.sqrt(query.size(-1))\n", "\n", " # fill 0 mask with super small number so it wont affect the softmax weight\n", " # (batch_size, h, max_len, max_len)\n", " scores = scores.masked_fill(mask == 0, -1e9) \n", "\n", " # (batch_size, h, max_len, max_len)\n", " # softmax to put attention weight for all non-pad tokens\n", " # max_len X max_len matrix of attention\n", " weights = F.softmax(scores, dim=-1) \n", " weights = self.dropout(weights)\n", "\n", " # (batch_size, h, max_len, max_len) matmul (batch_size, h, max_len, d_k) --> (batch_size, h, max_len, d_k)\n", " context = torch.matmul(weights, value)\n", "\n", " # (batch_size, h, max_len, d_k) --> (batch_size, max_len, h, d_k) --> (batch_size, max_len, d_model)\n", " context = context.permute(0, 2, 1, 3).contiguous().view(context.shape[0], -1, self.heads * self.d_k)\n", "\n", " # (batch_size, max_len, d_model)\n", " return self.output_linear(context)\n", "\n", "class FeedForward(torch.nn.Module):\n", " \"Implements FFN equation.\"\n", "\n", " def __init__(self, d_model, middle_dim=2048, dropout=0.1):\n", " super(FeedForward, self).__init__()\n", " \n", " self.fc1 = torch.nn.Linear(d_model, middle_dim)\n", " self.fc2 = torch.nn.Linear(middle_dim, d_model)\n", " self.dropout = torch.nn.Dropout(dropout)\n", " self.activation = torch.nn.GELU()\n", "\n", " def forward(self, x):\n", " out = self.activation(self.fc1(x))\n", " out = self.fc2(self.dropout(out))\n", " return out\n", "\n", "class EncoderLayer(torch.nn.Module):\n", " def __init__(\n", " self, \n", " d_model=768,\n", " heads=12, \n", " feed_forward_hidden=768 * 4, \n", " dropout=0.1\n", " ):\n", " super(EncoderLayer, self).__init__()\n", " self.layernorm = torch.nn.LayerNorm(d_model)\n", " self.self_multihead = MultiHeadedAttention(heads, d_model)\n", " self.feed_forward = FeedForward(d_model, middle_dim=feed_forward_hidden)\n", " self.dropout = torch.nn.Dropout(dropout)\n", "\n", " def forward(self, embeddings, mask):\n", " # embeddings: (batch_size, max_len, d_model)\n", " # encoder mask: (batch_size, 1, 1, max_len)\n", " # result: (batch_size, max_len, d_model)\n", " interacted = self.dropout(self.self_multihead(embeddings, embeddings, embeddings, mask))\n", " # residual layer\n", " interacted = self.layernorm(interacted + embeddings)\n", " # bottleneck\n", " feed_forward_out = self.dropout(self.feed_forward(interacted))\n", " encoded = self.layernorm(feed_forward_out + interacted)\n", " return encoded" ] }, { "cell_type": "code", "execution_count": 27, "id": "33fe273b", "metadata": {}, "outputs": [], "source": [ "class BERT(torch.nn.Module):\n", " \"\"\"\n", " BERT model : Bidirectional Encoder Representations from Transformers.\n", " \"\"\"\n", "\n", " def __init__(self, vocab_size, d_model=768, n_layers=12, heads=12, dropout=0.1):\n", " \"\"\"\n", " :param vocab_size: vocab_size of total words\n", " :param hidden: BERT model hidden size\n", " :param n_layers: numbers of Transformer blocks(layers)\n", " :param attn_heads: number of attention heads\n", " :param dropout: dropout rate\n", " \"\"\"\n", "\n", " super().__init__()\n", " self.d_model = d_model\n", " self.n_layers = n_layers\n", " self.heads = heads\n", "\n", " # paper noted they used 4 * hidden_size for ff_network_hidden_size\n", " self.feed_forward_hidden = d_model * 4\n", "\n", " # embedding for BERT, sum of positional, segment, token embeddings\n", " self.embedding = BERTEmbedding(vocab_size=vocab_size, embed_size=d_model)\n", "\n", " # multi-layers transformer blocks, deep network\n", " self.encoder_blocks = torch.nn.ModuleList(\n", " [EncoderLayer(d_model, heads, d_model * 4, dropout) for _ in range(n_layers)])\n", "\n", " def forward(self, x, segment_info):\n", " # attention masking for padded token\n", " # (batch_size, 1, seq_len, seq_len)\n", " mask = (x > 0).unsqueeze(1).repeat(1, x.size(1), 1).unsqueeze(1)\n", "\n", " # embedding the indexed sequence to sequence of vectors\n", " x = self.embedding(x, segment_info)\n", "\n", " # running over multiple transformer blocks\n", " for encoder in self.encoder_blocks:\n", " x = encoder.forward(x, mask)\n", " return x\n", "\n", "class NextSentencePrediction(torch.nn.Module):\n", " \"\"\"\n", " 2-class classification model : is_next, is_not_next\n", " \"\"\"\n", "\n", " def __init__(self, hidden):\n", " \"\"\"\n", " :param hidden: BERT model output size\n", " \"\"\"\n", " super().__init__()\n", " self.linear = torch.nn.Linear(hidden, 2)\n", " self.softmax = torch.nn.LogSoftmax(dim=-1)\n", "\n", " def forward(self, x):\n", " # use only the first token which is the [CLS]\n", " return self.softmax(self.linear(x[:, 0]))\n", "\n", "class MaskedLanguageModel(torch.nn.Module):\n", " \"\"\"\n", " predicting origin token from masked input sequence\n", " n-class classification problem, n-class = vocab_size\n", " \"\"\"\n", "\n", " def __init__(self, hidden, vocab_size):\n", " \"\"\"\n", " :param hidden: output size of BERT model\n", " :param vocab_size: total vocab size\n", " \"\"\"\n", " super().__init__()\n", " self.linear = torch.nn.Linear(hidden, vocab_size)\n", " self.softmax = torch.nn.LogSoftmax(dim=-1)\n", "\n", " def forward(self, x):\n", " return self.softmax(self.linear(x))\n", "\n", "class BERTLM(torch.nn.Module):\n", " \"\"\"\n", " BERT Language Model\n", " Next Sentence Prediction Model + Masked Language Model\n", " \"\"\"\n", "\n", " def __init__(self, bert: BERT, vocab_size):\n", " \"\"\"\n", " :param bert: BERT model which should be trained\n", " :param vocab_size: total vocab size for masked_lm\n", " \"\"\"\n", "\n", " super().__init__()\n", " self.bert = bert\n", " self.next_sentence = NextSentencePrediction(self.bert.d_model)\n", " self.mask_lm = MaskedLanguageModel(self.bert.d_model, vocab_size)\n", "\n", " def forward(self, x, segment_label):\n", " x = self.bert(x, segment_label)\n", " return self.next_sentence(x), self.mask_lm(x)" ] }, { "cell_type": "code", "execution_count": 24, "id": "b8f7c1f8", "metadata": {}, "outputs": [], "source": [ "class ScheduledOptim():\n", " '''A simple wrapper class for learning rate scheduling'''\n", "\n", " def __init__(self, optimizer, d_model, n_warmup_steps):\n", " self._optimizer = optimizer\n", " self.n_warmup_steps = n_warmup_steps\n", " self.n_current_steps = 0\n", " self.init_lr = np.power(d_model, -0.5)\n", "\n", " def step_and_update_lr(self):\n", " \"Step with the inner optimizer\"\n", " self._update_learning_rate()\n", " self._optimizer.step()\n", "\n", " def zero_grad(self):\n", " \"Zero out the gradients by the inner optimizer\"\n", " self._optimizer.zero_grad()\n", "\n", " def _get_lr_scale(self):\n", " return np.min([\n", " np.power(self.n_current_steps, -0.5),\n", " np.power(self.n_warmup_steps, -1.5) * self.n_current_steps])\n", "\n", " def _update_learning_rate(self):\n", " ''' Learning rate scheduling per step '''\n", "\n", " self.n_current_steps += 1\n", " lr = self.init_lr * self._get_lr_scale()\n", "\n", " for param_group in self._optimizer.param_groups:\n", " param_group['lr'] = lr" ] }, { "cell_type": "code", "execution_count": 25, "id": "9dd8e50e", "metadata": {}, "outputs": [], "source": [ "class BERTTrainer:\n", " def __init__(\n", " self, \n", " model, \n", " train_dataloader, \n", " test_dataloader=None, \n", " lr= 1e-4,\n", " weight_decay=0.01,\n", " betas=(0.9, 0.999),\n", " warmup_steps=10000,\n", " log_freq=10,\n", " device='cuda'\n", " ):\n", "\n", " self.device = device\n", " self.model = model\n", " self.train_data = train_dataloader\n", " self.test_data = test_dataloader\n", "\n", " # Setting the Adam optimizer with hyper-param\n", " self.optim = Adam(self.model.parameters(), lr=lr, betas=betas, weight_decay=weight_decay)\n", " self.optim_schedule = ScheduledOptim(\n", " self.optim, self.model.bert.d_model, n_warmup_steps=warmup_steps\n", " )\n", "\n", " # Using Negative Log Likelihood Loss function for predicting the masked_token\n", " self.criterion = torch.nn.NLLLoss(ignore_index=0)\n", " self.log_freq = log_freq\n", " print(\"Total Parameters:\", sum([p.nelement() for p in self.model.parameters()]))\n", " \n", " def train(self, epoch):\n", " self.iteration(epoch, self.train_data)\n", "\n", " def test(self, epoch):\n", " self.iteration(epoch, self.test_data, train=False)\n", "\n", " def iteration(self, epoch, data_loader, train=True):\n", " \n", " avg_loss = 0.0\n", " total_correct = 0\n", " total_element = 0\n", " \n", " mode = \"train\" if train else \"test\"\n", "\n", " # progress bar\n", " data_iter = tqdm.tqdm(\n", " enumerate(data_loader),\n", " desc=\"EP_%s:%d\" % (mode, epoch),\n", " total=len(data_loader),\n", " bar_format=\"{l_bar}{r_bar}\"\n", " )\n", "\n", " for i, data in data_iter:\n", "\n", " # 0. batch_data will be sent into the device(GPU or cpu)\n", " data = {key: value.to(self.device) for key, value in data.items()}\n", "\n", " # 1. forward the next_sentence_prediction and masked_lm model\n", " next_sent_output, mask_lm_output = self.model.forward(data[\"bert_input\"], data[\"segment_label\"])\n", "\n", " # 2-1. NLL(negative log likelihood) loss of is_next classification result\n", " next_loss = self.criterion(next_sent_output, data[\"is_next\"])\n", "\n", " # 2-2. NLLLoss of predicting masked token word\n", " # transpose to (m, vocab_size, seq_len) vs (m, seq_len)\n", " # criterion(mask_lm_output.view(-1, mask_lm_output.size(-1)), data[\"bert_label\"].view(-1))\n", " mask_loss = self.criterion(mask_lm_output.transpose(1, 2), data[\"bert_label\"])\n", "\n", " # 2-3. Adding next_loss and mask_loss : 3.4 Pre-training Procedure\n", " loss = next_loss + mask_loss\n", "\n", " # 3. backward and optimization only in train\n", " if train:\n", " self.optim_schedule.zero_grad()\n", " loss.backward()\n", " self.optim_schedule.step_and_update_lr()\n", "\n", " # next sentence prediction accuracy\n", " correct = next_sent_output.argmax(dim=-1).eq(data[\"is_next\"]).sum().item()\n", " avg_loss += loss.item()\n", " total_correct += correct\n", " total_element += data[\"is_next\"].nelement()\n", "\n", " post_fix = {\n", " \"epoch\": epoch,\n", " \"iter\": i,\n", " \"avg_loss\": avg_loss / (i + 1),\n", " \"avg_acc\": total_correct / total_element * 100,\n", " \"loss\": loss.item()\n", " }\n", "\n", " if i % self.log_freq == 0:\n", " data_iter.write(str(post_fix))\n", " print(\n", " f\"EP{epoch}, {mode}: \\\n", " avg_loss={avg_loss / len(data_iter)}, \\\n", " total_acc={total_correct * 100.0 / total_element}\"\n", " ) " ] }, { "cell_type": "code", "execution_count": 28, "id": "52de2072", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Total Parameters: 46699434\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "EP_train:0: 0%|| 0/6926 [01:16 20\u001b[0m \u001b[43mbert_trainer\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mtrain\u001b[49m\u001b[43m(\u001b[49m\u001b[43mepoch\u001b[49m\u001b[43m)\u001b[49m\n", "Cell \u001b[1;32mIn[25], line 32\u001b[0m, in \u001b[0;36mBERTTrainer.train\u001b[1;34m(self, epoch)\u001b[0m\n\u001b[0;32m 31\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mtrain\u001b[39m(\u001b[38;5;28mself\u001b[39m, epoch):\n\u001b[1;32m---> 32\u001b[0m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43miteration\u001b[49m\u001b[43m(\u001b[49m\u001b[43mepoch\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mtrain_data\u001b[49m\u001b[43m)\u001b[49m\n", "Cell \u001b[1;32mIn[25], line 75\u001b[0m, in \u001b[0;36mBERTTrainer.iteration\u001b[1;34m(self, epoch, data_loader, train)\u001b[0m\n\u001b[0;32m 73\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m train:\n\u001b[0;32m 74\u001b[0m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39moptim_schedule\u001b[38;5;241m.\u001b[39mzero_grad()\n\u001b[1;32m---> 75\u001b[0m \u001b[43mloss\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mbackward\u001b[49m\u001b[43m(\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m 76\u001b[0m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39moptim_schedule\u001b[38;5;241m.\u001b[39mstep_and_update_lr()\n\u001b[0;32m 78\u001b[0m \u001b[38;5;66;03m# next sentence prediction accuracy\u001b[39;00m\n", "File \u001b[1;32m~\\AppData\\Local\\anaconda3\\lib\\site-packages\\torch\\_tensor.py:396\u001b[0m, in \u001b[0;36mTensor.backward\u001b[1;34m(self, gradient, retain_graph, create_graph, inputs)\u001b[0m\n\u001b[0;32m 387\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m has_torch_function_unary(\u001b[38;5;28mself\u001b[39m):\n\u001b[0;32m 388\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m handle_torch_function(\n\u001b[0;32m 389\u001b[0m Tensor\u001b[38;5;241m.\u001b[39mbackward,\n\u001b[0;32m 390\u001b[0m (\u001b[38;5;28mself\u001b[39m,),\n\u001b[1;32m (...)\u001b[0m\n\u001b[0;32m 394\u001b[0m create_graph\u001b[38;5;241m=\u001b[39mcreate_graph,\n\u001b[0;32m 395\u001b[0m inputs\u001b[38;5;241m=\u001b[39minputs)\n\u001b[1;32m--> 396\u001b[0m \u001b[43mtorch\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mautograd\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mbackward\u001b[49m\u001b[43m(\u001b[49m\u001b[38;5;28;43mself\u001b[39;49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mgradient\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43minputs\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43minputs\u001b[49m\u001b[43m)\u001b[49m\n", "File \u001b[1;32m~\\AppData\\Local\\anaconda3\\lib\\site-packages\\torch\\autograd\\__init__.py:173\u001b[0m, in \u001b[0;36mbackward\u001b[1;34m(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)\u001b[0m\n\u001b[0;32m 168\u001b[0m retain_graph \u001b[38;5;241m=\u001b[39m create_graph\n\u001b[0;32m 170\u001b[0m \u001b[38;5;66;03m# The reason we repeat same the comment below is that\u001b[39;00m\n\u001b[0;32m 171\u001b[0m \u001b[38;5;66;03m# some Python versions print out the first line of a multi-line function\u001b[39;00m\n\u001b[0;32m 172\u001b[0m \u001b[38;5;66;03m# calls in the traceback and some print out the last line\u001b[39;00m\n\u001b[1;32m--> 173\u001b[0m \u001b[43mVariable\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_execution_engine\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mrun_backward\u001b[49m\u001b[43m(\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;66;43;03m# Calls into the C++ engine to run the backward pass\u001b[39;49;00m\n\u001b[0;32m 174\u001b[0m \u001b[43m \u001b[49m\u001b[43mtensors\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mgrad_tensors_\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mretain_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mcreate_graph\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43minputs\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m 175\u001b[0m \u001b[43m \u001b[49m\u001b[43mallow_unreachable\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43maccumulate_grad\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m)\u001b[49m\n", "\u001b[1;31mKeyboardInterrupt\u001b[0m: " ] } ], "source": [ "train_data = BERTDataset(\n", " pairs, seq_len=MAX_LEN, tokenizer=tokenizer)\n", "\n", "train_loader = DataLoader(\n", " train_data, batch_size=32, shuffle=True, pin_memory=True)\n", "\n", "bert_model = BERT(\n", " vocab_size=len(tokenizer.vocab),\n", " d_model=768,\n", " n_layers=2,\n", " heads=12,\n", " dropout=0.1\n", ")\n", "\n", "bert_lm = BERTLM(bert_model, len(tokenizer.vocab))\n", "bert_trainer = BERTTrainer(bert_lm, train_loader, device='cpu')\n", "epochs = 20\n", "\n", "for epoch in range(epochs):\n", " bert_trainer.train(epoch)" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "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": "3.10.9" } }, "nbformat": 4, "nbformat_minor": 5 }