Jakub Adamski dc373e55a8 GPT2 custom & dataset metrics

2023-02-18 10:25:07 +01:00

289 KiB

Raw Permalink Blame History

Instalacja pakietów

!pip install transformers datasets torch

Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Collecting transformers
  Downloading transformers-4.26.1-py3-none-any.whl (6.3 MB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m6.3/6.3 MB[0m [31m51.6 MB/s[0m eta [36m0:00:00[0m
[?25hCollecting datasets
  Downloading datasets-2.9.0-py3-none-any.whl (462 kB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m462.8/462.8 KB[0m [31m24.0 MB/s[0m eta [36m0:00:00[0m
[?25hRequirement already satisfied: torch in /usr/local/lib/python3.8/dist-packages (1.13.1+cu116)
Requirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.8/dist-packages (from transformers) (2022.6.2)
Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.8/dist-packages (from transformers) (1.21.6)
Requirement already satisfied: filelock in /usr/local/lib/python3.8/dist-packages (from transformers) (3.9.0)
Collecting tokenizers!=0.11.3,<0.14,>=0.11.1
  Downloading tokenizers-0.13.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (7.6 MB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m7.6/7.6 MB[0m [31m30.9 MB/s[0m eta [36m0:00:00[0m
[?25hRequirement already satisfied: packaging>=20.0 in /usr/local/lib/python3.8/dist-packages (from transformers) (23.0)
Requirement already satisfied: pyyaml>=5.1 in /usr/local/lib/python3.8/dist-packages (from transformers) (6.0)
Requirement already satisfied: requests in /usr/local/lib/python3.8/dist-packages (from transformers) (2.25.1)
Requirement already satisfied: tqdm>=4.27 in /usr/local/lib/python3.8/dist-packages (from transformers) (4.64.1)
Collecting huggingface-hub<1.0,>=0.11.0
  Downloading huggingface_hub-0.12.1-py3-none-any.whl (190 kB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m190.3/190.3 KB[0m [31m18.0 MB/s[0m eta [36m0:00:00[0m
[?25hCollecting multiprocess
  Downloading multiprocess-0.70.14-py38-none-any.whl (132 kB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m132.0/132.0 KB[0m [31m10.9 MB/s[0m eta [36m0:00:00[0m
[?25hCollecting xxhash
  Downloading xxhash-3.2.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (213 kB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m213.0/213.0 KB[0m [31m12.1 MB/s[0m eta [36m0:00:00[0m
[?25hRequirement already satisfied: fsspec[http]>=2021.11.1 in /usr/local/lib/python3.8/dist-packages (from datasets) (2023.1.0)
Requirement already satisfied: dill<0.3.7 in /usr/local/lib/python3.8/dist-packages (from datasets) (0.3.6)
Collecting responses<0.19
  Downloading responses-0.18.0-py3-none-any.whl (38 kB)
Requirement already satisfied: pyarrow>=6.0.0 in /usr/local/lib/python3.8/dist-packages (from datasets) (9.0.0)
Requirement already satisfied: aiohttp in /usr/local/lib/python3.8/dist-packages (from datasets) (3.8.4)
Requirement already satisfied: pandas in /usr/local/lib/python3.8/dist-packages (from datasets) (1.3.5)
Requirement already satisfied: typing-extensions in /usr/local/lib/python3.8/dist-packages (from torch) (4.5.0)
Requirement already satisfied: aiosignal>=1.1.2 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (1.3.1)
Requirement already satisfied: multidict<7.0,>=4.5 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (6.0.4)
Requirement already satisfied: charset-normalizer<4.0,>=2.0 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (3.0.1)
Requirement already satisfied: frozenlist>=1.1.1 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (1.3.3)
Requirement already satisfied: yarl<2.0,>=1.0 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (1.8.2)
Requirement already satisfied: attrs>=17.3.0 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (22.2.0)
Requirement already satisfied: async-timeout<5.0,>=4.0.0a3 in /usr/local/lib/python3.8/dist-packages (from aiohttp->datasets) (4.0.2)
Requirement already satisfied: urllib3<1.27,>=1.21.1 in /usr/local/lib/python3.8/dist-packages (from requests->transformers) (1.24.3)
Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.8/dist-packages (from requests->transformers) (2022.12.7)
Requirement already satisfied: chardet<5,>=3.0.2 in /usr/local/lib/python3.8/dist-packages (from requests->transformers) (4.0.0)
Requirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.8/dist-packages (from requests->transformers) (2.10)
Collecting urllib3<1.27,>=1.21.1
  Downloading urllib3-1.26.14-py2.py3-none-any.whl (140 kB)
[2K     [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m140.6/140.6 KB[0m [31m9.2 MB/s[0m eta [36m0:00:00[0m
[?25hRequirement already satisfied: pytz>=2017.3 in /usr/local/lib/python3.8/dist-packages (from pandas->datasets) (2022.7.1)
Requirement already satisfied: python-dateutil>=2.7.3 in /usr/local/lib/python3.8/dist-packages (from pandas->datasets) (2.8.2)
Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.8/dist-packages (from python-dateutil>=2.7.3->pandas->datasets) (1.15.0)
Installing collected packages: tokenizers, xxhash, urllib3, multiprocess, responses, huggingface-hub, transformers, datasets
  Attempting uninstall: urllib3
    Found existing installation: urllib3 1.24.3
    Uninstalling urllib3-1.24.3:
      Successfully uninstalled urllib3-1.24.3
Successfully installed datasets-2.9.0 huggingface-hub-0.12.1 multiprocess-0.70.14 responses-0.18.0 tokenizers-0.13.2 transformers-4.26.1 urllib3-1.26.14 xxhash-3.2.0

Załadowanie pakietów

from datasets import load_dataset
from transformers import GPT2Tokenizer
import torch
from torch.utils.data import TensorDataset, random_split
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from transformers import GPT2ForSequenceClassification, GPT2Config
from transformers import get_linear_schedule_with_warmup
import numpy as np
import time
import datetime
import random

Załadowanie datasetu

sms_spam

dataset = load_dataset("sms_spam")

Downloading builder script:   0%|          | 0.00/3.21k [00:00<?, ?B/s]

Downloading metadata:   0%|          | 0.00/1.69k [00:00<?, ?B/s]

Downloading readme:   0%|          | 0.00/4.87k [00:00<?, ?B/s]

Downloading and preparing dataset sms_spam/plain_text to /root/.cache/huggingface/datasets/sms_spam/plain_text/1.0.0/53f051d3b5f62d99d61792c91acefe4f1577ad3e4c216fb0ad39e30b9f20019c...

Downloading data:   0%|          | 0.00/203k [00:00<?, ?B/s]

Generating train split:   0%|          | 0/5574 [00:00<?, ? examples/s]

Dataset sms_spam downloaded and prepared to /root/.cache/huggingface/datasets/sms_spam/plain_text/1.0.0/53f051d3b5f62d99d61792c91acefe4f1577ad3e4c216fb0ad39e30b9f20019c. Subsequent calls will reuse this data.

  0%|          | 0/1 [00:00<?, ?it/s]

dataset['train'][0]

{'sms': 'Go until jurong point, crazy.. Available only in bugis n great world la e buffet... Cine there got amore wat...\n',
 'label': 0}

Tokenizer GPT2

tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
tokenizer.pad_token = tokenizer.eos_token

Downloading (…)olve/main/vocab.json:   0%|          | 0.00/1.04M [00:00<?, ?B/s]

Downloading (…)olve/main/merges.txt:   0%|          | 0.00/456k [00:00<?, ?B/s]

Downloading (…)lve/main/config.json:   0%|          | 0.00/665 [00:00<?, ?B/s]

sms = dataset['train'][0]['sms']
print('Original: ', sms)
print('Tokenized: ', tokenizer.tokenize(sms))
print('Token IDs: ', tokenizer.convert_tokens_to_ids(tokenizer.tokenize(sms)))

Original:  Go until jurong point, crazy.. Available only in bugis n great world la e buffet... Cine there got amore wat...

Tokenized:  ['Go', 'Ġuntil', 'Ġjur', 'ong', 'Ġpoint', ',', 'Ġcrazy', '..', 'ĠAvailable', 'Ġonly', 'Ġin', 'Ġbug', 'is', 'Ġn', 'Ġgreat', 'Ġworld', 'Ġla', 'Ġe', 'Ġbuffet', '...', 'ĠC', 'ine', 'Ġthere', 'Ġgot', 'Ġam', 'ore', 'Ġwat', '...', 'Ċ']
Token IDs:  [5247, 1566, 8174, 506, 966, 11, 7165, 492, 14898, 691, 287, 5434, 271, 299, 1049, 995, 8591, 304, 44703, 986, 327, 500, 612, 1392, 716, 382, 4383, 986, 198]

Check maximum length of a sentence

max_len = 0

for sentence in dataset['train']:
    input_ids = tokenizer.encode(sentence['sms'], add_special_tokens=True)
    max_len = max(max_len, len(input_ids))

print('Max sentence length: ', max_len)

Max sentence length:  258

Special tokenization

input_ids = []
attention_masks = []

for sentence in dataset['train']:
    encoded_dict = tokenizer.encode_plus(
                        sentence['sms'],
                        add_special_tokens = True,
                        max_length = 260,
                        padding = 'max_length',
                        truncation=True,
                        return_attention_mask = True,
                        return_tensors = 'pt',
                   )
    
    input_ids.append(encoded_dict['input_ids'])
    attention_masks.append(encoded_dict['attention_mask'])

input_ids = torch.cat(input_ids, dim=0)
attention_masks = torch.cat(attention_masks, dim=0)
labels = torch.tensor([sentence['label'] for sentence in dataset['train']])

print('Original: ', dataset['train'][0])
print('Token IDs:', input_ids[0])

Original:  {'sms': 'Go until jurong point, crazy.. Available only in bugis n great world la e buffet... Cine there got amore wat...\n', 'label': 0}
Token IDs: tensor([ 5247,  1566,  8174,   506,   966,    11,  7165,   492, 14898,   691,
          287,  5434,   271,   299,  1049,   995,  8591,   304, 44703,   986,
          327,   500,   612,  1392,   716,   382,  4383,   986,   198, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256,
        50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256, 50256])

Split dataset

Class balance ratio should be similar to base dataset ratio.

def check_class_balance(dataset):
  spam_count = 0.0
  not_spam_count = 0.0
  for row in dataset:
    if row[2].item() == 1:
      spam_count += 1.0
    else:
      not_spam_count += 1.0
  return spam_count / not_spam_count

dataset = TensorDataset(input_ids, attention_masks, labels)
print("Spam to not spam messages ratio: {}\n".format(check_class_balance(dataset)))

test_size = 1000
dataset_len = len(dataset)
train_size = int(0.9 * (dataset_len-test_size))
val_size = (dataset_len-test_size) - train_size

test_dataset, train_dataset, val_dataset = random_split(dataset, [test_size, train_size, val_size])

print('{:>5,} test samples'.format(test_size))
print("Ratio: {}\n".format(check_class_balance(test_dataset)))
print('{:>5,} training samples'.format(train_size))
print("Ratio: {}\n".format(check_class_balance(train_dataset)))
print('{:>5,} validation samples'.format(val_size))
print("Ratio: {}\n".format(check_class_balance(val_dataset)))

Spam to not spam messages ratio: 0.15475450590428838

1,000 test samples
Ratio: 0.15606936416184972

4,116 training samples
Ratio: 0.15294117647058825

  458 validation samples
Ratio: 0.1683673469387755

Create train and validation loaders

batch_size = 8

train_dataloader = DataLoader(
            train_dataset,
            sampler = RandomSampler(train_dataset),
            batch_size = batch_size
        )

validation_dataloader = DataLoader(
            val_dataset,
            sampler = SequentialSampler(val_dataset),
            batch_size = batch_size
        )

Device check

if torch.cuda.is_available():    
    device = torch.device("cuda")

    print('There are %d GPU(s) available.' % torch.cuda.device_count())
    print('We will use the GPU:', torch.cuda.get_device_name(0))

else:
    print('No GPU available, using the CPU instead.')
    device = torch.device("cpu")

There are 1 GPU(s) available.
We will use the GPU: Tesla T4

Load GPT2 model

model = GPT2ForSequenceClassification.from_pretrained(
    'gpt2',
    num_labels = 2,
)

Downloading (…)"pytorch_model.bin";:   0%|          | 0.00/548M [00:00<?, ?B/s]

Some weights of GPT2ForSequenceClassification were not initialized from the model checkpoint at gpt2 and are newly initialized: ['score.weight']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

model.resize_token_embeddings(len(tokenizer))
model.config.pad_token_id = model.config.eos_token_id

model.cuda()

GPT2ForSequenceClassification(
  (transformer): GPT2Model(
    (wte): Embedding(50257, 768)
    (wpe): Embedding(1024, 768)
    (drop): Dropout(p=0.1, inplace=False)
    (h): ModuleList(
      (0): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (1): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (2): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (3): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (4): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (5): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (6): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (7): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (8): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (9): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (10): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
      (11): GPT2Block(
        (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (attn): GPT2Attention(
          (c_attn): Conv1D()
          (c_proj): Conv1D()
          (attn_dropout): Dropout(p=0.1, inplace=False)
          (resid_dropout): Dropout(p=0.1, inplace=False)
        )
        (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        (mlp): GPT2MLP(
          (c_fc): Conv1D()
          (c_proj): Conv1D()
          (act): NewGELUActivation()
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
    )
    (ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
  )
  (score): Linear(in_features=768, out_features=2, bias=False)
)

Init training parameters

optimizer = torch.optim.AdamW(model.parameters(),
                  lr = 2e-5,
                  eps = 1e-8
                )

epochs = 4
total_steps = len(train_dataloader) * epochs

scheduler = get_linear_schedule_with_warmup(optimizer, 
                                            num_warmup_steps = 0,
                                            num_training_steps = total_steps)

Helper functions

def flat_accuracy(preds, labels):
    pred_flat = np.argmax(preds, axis=1).flatten()
    labels_flat = labels.flatten()
    return np.sum(pred_flat == labels_flat) / len(labels_flat)

def format_time(elapsed):
    '''
    Takes a time in seconds and returns a string hh:mm:ss
    '''
    elapsed_rounded = int(round((elapsed)))
    return str(datetime.timedelta(seconds=elapsed_rounded))

Training

# This training code is based on the `run_glue.py` script here:
# https://github.com/huggingface/transformers/blob/5bfcd0485ece086ebcbed2d008813037968a9e58/examples/run_glue.py#L128

seed_val = 42

random.seed(seed_val)
np.random.seed(seed_val)
torch.manual_seed(seed_val)
torch.cuda.manual_seed_all(seed_val)

training_stats = []
total_t0 = time.time()

for epoch_i in range(0, epochs):
    
    # ========================================
    #               Training
    # ========================================

    print("")
    print('======== Epoch {:} / {:} ========'.format(epoch_i + 1, epochs))
    print('Training...')

    t0 = time.time()
    total_train_loss = 0

    model.train()

    for step, batch in enumerate(train_dataloader):
        if step % 40 == 0 and not step == 0:
            elapsed = format_time(time.time() - t0)
            print('  Batch {:>5,}  of  {:>5,}.    Elapsed: {:}.'.format(step, len(train_dataloader), elapsed))

        b_input_ids = batch[0].to(device)
        b_input_mask = batch[1].to(device)
        b_labels = batch[2].to(device)

        model.zero_grad()        

        outputs = model(b_input_ids, 
                             token_type_ids=None, 
                             attention_mask=b_input_mask, 
                             labels=b_labels)

        loss = outputs['loss']
        total_train_loss += loss.item()

        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)

        optimizer.step()
        scheduler.step()

    avg_train_loss = total_train_loss / len(train_dataloader)            
    training_time = format_time(time.time() - t0)

    print("")
    print("  Average training loss: {0:.2f}".format(avg_train_loss))
    print("  Training epcoh took: {:}".format(training_time))
        
    # ========================================
    #               Validation
    # ========================================

    print("")
    print("Running Validation...")

    t0 = time.time()
    model.eval()

    total_eval_accuracy = 0
    total_eval_loss = 0
    nb_eval_steps = 0

    for batch in validation_dataloader:
        b_input_ids = batch[0].to(device)
        b_input_mask = batch[1].to(device)
        b_labels = batch[2].to(device)
        
        with torch.no_grad():        
            outputs = model(b_input_ids, 
                                   token_type_ids=None, 
                                   attention_mask=b_input_mask,
                                   labels=b_labels)
            loss = outputs['loss']
            logits = outputs['logits']
            
        total_eval_loss += loss.item()

        logits = logits.detach().cpu().numpy()
        label_ids = b_labels.to('cpu').numpy()

        total_eval_accuracy += flat_accuracy(logits, label_ids)
        
    avg_val_accuracy = total_eval_accuracy / len(validation_dataloader)
    print("  Accuracy: {0:.2f}".format(avg_val_accuracy))

    avg_val_loss = total_eval_loss / len(validation_dataloader)
    validation_time = format_time(time.time() - t0)
    
    print("  Validation Loss: {0:.2f}".format(avg_val_loss))
    print("  Validation took: {:}".format(validation_time))

    training_stats.append(
        {
            'epoch': epoch_i + 1,
            'Training Loss': avg_train_loss,
            'Valid. Loss': avg_val_loss,
            'Valid. Accur.': avg_val_accuracy,
            'Training Time': training_time,
            'Validation Time': validation_time
        }
    )

print("")
print("Training complete!")

print("Total training took {:} (h:mm:ss)".format(format_time(time.time()-total_t0)))

======== Epoch 1 / 4 ========
Training...
  Batch    40  of    515.    Elapsed: 0:00:19.
  Batch    80  of    515.    Elapsed: 0:00:35.
  Batch   120  of    515.    Elapsed: 0:00:52.
  Batch   160  of    515.    Elapsed: 0:01:08.
  Batch   200  of    515.    Elapsed: 0:01:25.
  Batch   240  of    515.    Elapsed: 0:01:42.
  Batch   280  of    515.    Elapsed: 0:01:58.
  Batch   320  of    515.    Elapsed: 0:02:15.
  Batch   360  of    515.    Elapsed: 0:02:32.
  Batch   400  of    515.    Elapsed: 0:02:49.
  Batch   440  of    515.    Elapsed: 0:03:06.
  Batch   480  of    515.    Elapsed: 0:03:24.

  Average training loss: 0.14
  Training epcoh took: 0:03:38

Running Validation...
  Accuracy: 0.97
  Validation Loss: 0.21
  Validation took: 0:00:08

======== Epoch 2 / 4 ========
Training...
  Batch    40  of    515.    Elapsed: 0:00:17.
  Batch    80  of    515.    Elapsed: 0:00:35.
  Batch   120  of    515.    Elapsed: 0:00:52.
  Batch   160  of    515.    Elapsed: 0:01:09.
  Batch   200  of    515.    Elapsed: 0:01:27.
  Batch   240  of    515.    Elapsed: 0:01:44.
  Batch   280  of    515.    Elapsed: 0:02:01.
  Batch   320  of    515.    Elapsed: 0:02:19.
  Batch   360  of    515.    Elapsed: 0:02:36.
  Batch   400  of    515.    Elapsed: 0:02:54.
  Batch   440  of    515.    Elapsed: 0:03:11.
  Batch   480  of    515.    Elapsed: 0:03:29.

  Average training loss: 0.04
  Training epcoh took: 0:03:44

Running Validation...
  Accuracy: 0.97
  Validation Loss: 0.19
  Validation took: 0:00:08

======== Epoch 3 / 4 ========
Training...
  Batch    40  of    515.    Elapsed: 0:00:17.
  Batch    80  of    515.    Elapsed: 0:00:35.
  Batch   120  of    515.    Elapsed: 0:00:52.
  Batch   160  of    515.    Elapsed: 0:01:10.
  Batch   200  of    515.    Elapsed: 0:01:27.
  Batch   240  of    515.    Elapsed: 0:01:45.
  Batch   280  of    515.    Elapsed: 0:02:02.
  Batch   320  of    515.    Elapsed: 0:02:20.
  Batch   360  of    515.    Elapsed: 0:02:37.
  Batch   400  of    515.    Elapsed: 0:02:55.
  Batch   440  of    515.    Elapsed: 0:03:12.
  Batch   480  of    515.    Elapsed: 0:03:30.

  Average training loss: 0.03
  Training epcoh took: 0:03:45

Running Validation...
  Accuracy: 0.97
  Validation Loss: 0.16
  Validation took: 0:00:08

======== Epoch 4 / 4 ========
Training...
  Batch    40  of    515.    Elapsed: 0:00:17.
  Batch    80  of    515.    Elapsed: 0:00:35.
  Batch   120  of    515.    Elapsed: 0:00:52.
  Batch   160  of    515.    Elapsed: 0:01:10.
  Batch   200  of    515.    Elapsed: 0:01:27.
  Batch   240  of    515.    Elapsed: 0:01:45.
  Batch   280  of    515.    Elapsed: 0:02:02.
  Batch   320  of    515.    Elapsed: 0:02:20.
  Batch   360  of    515.    Elapsed: 0:02:37.
  Batch   400  of    515.    Elapsed: 0:02:55.
  Batch   440  of    515.    Elapsed: 0:03:12.
  Batch   480  of    515.    Elapsed: 0:03:30.

  Average training loss: 0.01
  Training epcoh took: 0:03:45

Running Validation...
  Accuracy: 0.98
  Validation Loss: 0.11
  Validation took: 0:00:08

Training complete!
Total training took 0:15:24 (h:mm:ss)

Train summary

import pandas as pd

pd.set_option('precision', 2)
df_stats = pd.DataFrame(data=training_stats)

df_stats = df_stats.set_index('epoch')
df_stats

	Training Loss	Valid. Loss	Valid. Accur.	Training Time	Validation Time
epoch
1	0.14	0.21	0.97	0:03:38	0:00:08
2	0.04	0.19	0.97	0:03:44	0:00:08
3	0.03	0.16	0.97	0:03:45	0:00:08
4	0.01	0.11	0.98	0:03:45	0:00:08

import matplotlib.pyplot as plt
%matplotlib inline

import seaborn as sns

sns.set(style='darkgrid')

sns.set(font_scale=1.5)
plt.rcParams["figure.figsize"] = (12,6)

plt.plot(df_stats['Training Loss'], 'b-o', label="Training")
plt.plot(df_stats['Valid. Loss'], 'g-o', label="Validation")

plt.title("Training & Validation Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.legend()
plt.xticks([1, 2, 3, 4])

plt.show()

Create test loader

prediction_dataloader = DataLoader(
            test_dataset,
            sampler = SequentialSampler(test_dataset),
            batch_size = batch_size
        )

Evaluate on test dataset

print('Predicting labels for {:,} test sentences...'.format(len(test_dataset)))

model.eval()
predictions , true_labels = [], []

for batch in prediction_dataloader:
  batch = tuple(t.to(device) for t in batch)
  
  b_input_ids, b_input_mask, b_labels = batch
  
  with torch.no_grad():
      outputs = model(b_input_ids, token_type_ids=None, 
                      attention_mask=b_input_mask)

  logits = outputs['logits']

  logits = logits.detach().cpu().numpy()
  label_ids = b_labels.to('cpu').numpy()

  predictions.append(logits)
  true_labels.append(label_ids)

print('    DONE.')

Predicting labels for 1,000 test sentences...
    DONE.

results_ok = 0
results_false = 0
for idx, true_labels_batch in enumerate(true_labels):
  predictions_i = np.argmax(predictions[idx], axis=1).flatten()
  for bidx, true_label in enumerate(true_labels_batch):
    if true_label == predictions_i[bidx]:
      results_ok += 1
    else:
      results_false += 1

print("Correct predictions: {}, incorrect results: {}, accuracy: {}".format(results_ok, results_false, float(results_ok) / (results_ok + results_false)))

Correct predictions: 990, incorrect results: 10, accuracy: 0.99

MCC Score

from sklearn.metrics import matthews_corrcoef

matthews_set = []
print('Calculating Matthews Corr. Coef. for each batch...')

for i in range(len(true_labels)):
  pred_labels_i = np.argmax(predictions[i], axis=1).flatten()
  
  matthews = matthews_corrcoef(true_labels[i], pred_labels_i)                
  matthews_set.append(matthews)

Calculating Matthews Corr. Coef. for each batch...

ax = sns.barplot(x=list(range(len(matthews_set))), y=matthews_set, ci=None)

plt.title('MCC Score per Batch')
plt.ylabel('MCC Score (-1 to +1)')
plt.xlabel('Batch #')

plt.show()

flat_predictions = np.concatenate(predictions, axis=0)
flat_predictions = np.argmax(flat_predictions, axis=1).flatten()

flat_true_labels = np.concatenate(true_labels, axis=0)

mcc = matthews_corrcoef(flat_true_labels, flat_predictions)

print('Total MCC: %.3f' % mcc)

Total MCC: 0.960

Save model

from google.colab import drive

drive.mount('/content/gdrive/', force_remount=True)

output_dir = '/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model'
print("Saving model to %s" % output_dir)

model_to_save = model.module if hasattr(model, 'module') else model
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)

Mounted at /content/gdrive/
Saving model to /content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model

('/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model/tokenizer_config.json',
 '/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model/special_tokens_map.json',
 '/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model/vocab.json',
 '/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model/merges.txt',
 '/content/gdrive/My Drive/UAM/Przetwarzanie-tekstu/GPT2_Model/added_tokens.json')

289 KiB Raw Permalink Blame History Unescape Escape

Instalacja pakietów

Załadowanie pakietów

Załadowanie datasetu

Tokenizer GPT2

Check maximum length of a sentence

Special tokenization

Split dataset

Create train and validation loaders

Device check

Load GPT2 model

Init training parameters

Helper functions

Training

Train summary

Create test loader

Evaluate on test dataset

MCC Score

Save model

Bibliografia

289 KiB

Raw Permalink Blame History