aitech-eks-pub/cw/08_regresja_logistyczna.ipynb

Regresja logistyczna

import bibliotek

import numpy as np
import gensim
import torch
import pandas as pd
from sklearn.model_selection import train_test_split

from sklearn.datasets import fetch_20newsgroups
# https://scikit-learn.org/0.19/datasets/twenty_newsgroups.html

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics import accuracy_score

/media/kuba/ssdsam/anaconda3/lib/python3.8/site-packages/gensim/similarities/__init__.py:15: UserWarning: The gensim.similarities.levenshtein submodule is disabled, because the optional Levenshtein package <https://pypi.org/project/python-Levenshtein/> is unavailable. Install Levenhstein (e.g. `pip install python-Levenshtein`) to suppress this warning.
  warnings.warn(msg)

CATEGORIES = ['soc.religion.christian', 'alt.atheism']

newsgroups_train_dev = fetch_20newsgroups(subset = 'train', categories=CATEGORIES)
newsgroups_test = fetch_20newsgroups(subset = 'test', categories=CATEGORIES)

newsgroups_train_dev_text = newsgroups_train_dev['data']
newsgroups_test_text = newsgroups_test['data']

Y_train_dev = newsgroups_train_dev['target']
Y_test = newsgroups_test['target']

newsgroups_train_text, newsgroups_dev_text, Y_train, Y_dev = train_test_split(newsgroups_train_dev_text, Y_train_dev, random_state=42)

Y_names = newsgroups_train_dev['target_names']

Y_names

['alt.atheism', 'soc.religion.christian']

baseline

zadanie (5 minut)

• stworzyć baseline

PYTANIE: co jest nie tak z regresją liniową?

Regresja logistyczna

wektoryzacja

zadanie (5 minut)

• na podstawie newsgroups_train_text stworzyć tfidf wektoryzer ze słownikiem max 10_000
• wygenerować wektory: X_train, X_dev, X_test

model - inicjalizacja

class LogisticRegressionModel(torch.nn.Module):

    def __init__(self):
        super(LogisticRegressionModel, self).__init__()
        self.fc = torch.nn.Linear(FEAUTERES,1)

    def forward(self, x):
        x = self.fc(x)
        x = torch.sigmoid(x)
        return x

lr_model = LogisticRegressionModel()

lr_model(torch.Tensor(X_train[0:5].astype(np.float32).todense()))

tensor([[0.4978],
        [0.5009],
        [0.4998],
        [0.4990],
        [0.5018]], grad_fn=<SigmoidBackward>)

lr_model

LogisticRegressionModel(
  (fc): Linear(in_features=10000, out_features=1, bias=True)
)

list(lr_model.parameters())

[Parameter containing:
 tensor([[-0.0059,  0.0035,  0.0021,  ..., -0.0042, -0.0057, -0.0049]],
        requires_grad=True),
 Parameter containing:
 tensor([-0.0023], requires_grad=True)]

model - trenowanie

BATCH_SIZE = 5

criterion = torch.nn.BCELoss()

optimizer = torch.optim.SGD(lr_model.parameters(), lr = 0.1)

Y_train.shape[0]

809

loss_score = 0
acc_score = 0
items_total = 0
lr_model.train()
for i in range(0, Y_train.shape[0], BATCH_SIZE):
    X = X_train[i:i+BATCH_SIZE]
    X = torch.tensor(X.astype(np.float32).todense())
    Y = Y_train[i:i+BATCH_SIZE]
    Y = torch.tensor(Y.astype(np.float32)).reshape(-1,1)
    Y_predictions = lr_model(X)
    acc_score += torch.sum((Y_predictions > 0.5) == Y).item()
    items_total += Y.shape[0] 
    
    optimizer.zero_grad()
    loss = criterion(Y_predictions, Y)
    loss.backward()
    optimizer.step()
    

    loss_score += loss.item() * Y.shape[0] 

Y_predictions

tensor([[0.5667],
        [0.5802],
        [0.5757],
        [0.5670]], grad_fn=<SigmoidBackward>)

Y

tensor([[0.],
        [1.],
        [1.],
        [0.]])

acc_score

452

items_total

809

print(f'accuracy: {acc_score / items_total}')

accuracy: 0.5587144622991347

print(f'BCE loss: {loss_score / items_total}')

BCE loss: 0.6745463597170355

model - ewaluacja

def get_loss_acc(model, X_dataset, Y_dataset):
    loss_score = 0
    acc_score = 0
    items_total = 0
    model.eval()
    for i in range(0, Y_dataset.shape[0], BATCH_SIZE):
        X = X_dataset[i:i+BATCH_SIZE]
        X = torch.tensor(X.astype(np.float32).todense())
        Y = Y_dataset[i:i+BATCH_SIZE]
        Y = torch.tensor(Y.astype(np.float32)).reshape(-1,1)
        Y_predictions = model(X)
        acc_score += torch.sum((Y_predictions > 0.5) == Y).item()
        items_total += Y.shape[0] 

        optimizer.zero_grad()
        loss = criterion(Y_predictions, Y)
        loss.backward()
        optimizer.step()


        loss_score += loss.item() * Y.shape[0] 
    return (loss_score / items_total), (acc_score / items_total)

get_loss_acc(lr_model, X_train, Y_train)

(0.6443227143826974, 0.622991347342398)

get_loss_acc(lr_model, X_dev, Y_dev)

(0.6369243131743537, 0.6037037037037037)

get_loss_acc(lr_model, X_test, Y_test)

(0.6323775731785694, 0.6499302649930265)

wagi modelu

list(lr_model.parameters())

[Parameter containing:
 tensor([[ 0.0314, -0.0375,  0.0131,  ..., -0.0057, -0.0008, -0.0089]],
        requires_grad=True),
 Parameter containing:
 tensor([0.0563], requires_grad=True)]

list(lr_model.parameters())[0][0]

tensor([ 0.0314, -0.0375,  0.0131,  ..., -0.0057, -0.0008, -0.0089],
       grad_fn=<SelectBackward>)

torch.topk(list(lr_model.parameters())[0][0], 20)

torch.return_types.topk(
values=tensor([0.3753, 0.2305, 0.2007, 0.2006, 0.1993, 0.1952, 0.1930, 0.1898, 0.1831,
        0.1731, 0.1649, 0.1647, 0.1543, 0.1320, 0.1314, 0.1303, 0.1296, 0.1261,
        0.1245, 0.1243], grad_fn=<TopkBackward>),
indices=tensor([8942, 6336, 1852, 9056, 1865, 4039, 7820, 5002, 8208, 1857, 9709,  803,
        1046,  130, 4306, 6481, 4370, 4259, 4285, 1855]))

for i in torch.topk(list(lr_model.parameters())[0][0], 20)[1]:
    print(vectorizer.get_feature_names()[i])

the
of
christ
to
church
god
rutgers
jesus
sin
christians
we
and
athos
1993
hell
our
his
he
heaven
christian

torch.topk(list(lr_model.parameters())[0][0], 20, largest = False)

torch.return_types.topk(
values=tensor([-0.3478, -0.2578, -0.2455, -0.2347, -0.2330, -0.2265, -0.2205, -0.2050,
        -0.2044, -0.1979, -0.1876, -0.1790, -0.1747, -0.1745, -0.1734, -0.1647,
        -0.1639, -0.1617, -0.1601, -0.1592], grad_fn=<TopkBackward>),
indices=tensor([5119, 8096, 5420, 4436, 6194, 1627, 6901, 5946, 9970, 3116, 1036, 9906,
        5654, 8329, 7869, 1039, 1991, 4926, 5035, 4925]))

for i in torch.topk(list(lr_model.parameters())[0][0], 20, largest = False)[1]:
    print(vectorizer.get_feature_names()[i])

keith
sgi
livesey
host
nntp
caltech
posting
morality
you
edu
atheism
wpd
mathew
solntze
sandvik
atheists
com
islamic
jon
islam

sieć neuronowa

class NeuralNetworkModel(torch.nn.Module):

    def __init__(self):
        super(NeuralNetworkModel, self).__init__()
        self.fc1 = torch.nn.Linear(FEAUTERES,500)
        self.fc2 = torch.nn.Linear(500,1)

    def forward(self, x):
        x = self.fc1(x)
        x = self.fc2(x)
        x = torch.sigmoid(x)
        return x

nn_model = NeuralNetworkModel()

BATCH_SIZE = 5

criterion = torch.nn.BCELoss()

optimizer = torch.optim.SGD(nn_model.parameters(), lr = 0.1)

for epoch in range(5):
    loss_score = 0
    acc_score = 0
    items_total = 0
    nn_model.train()
    for i in range(0, Y_train.shape[0], BATCH_SIZE):
        X = X_train[i:i+BATCH_SIZE]
        X = torch.tensor(X.astype(np.float32).todense())
        Y = Y_train[i:i+BATCH_SIZE]
        Y = torch.tensor(Y.astype(np.float32)).reshape(-1,1)
        Y_predictions = nn_model(X)
        acc_score += torch.sum((Y_predictions > 0.5) == Y).item()
        items_total += Y.shape[0] 

        optimizer.zero_grad()
        loss = criterion(Y_predictions, Y)
        loss.backward()
        optimizer.step()


        loss_score += loss.item() * Y.shape[0] 

    display(epoch)
    display(get_loss_acc(nn_model, X_train, Y_train))
    display(get_loss_acc(nn_model, X_dev, Y_dev))

0

(0.6605833534551934, 0.5908529048207664)

(0.6379233609747004, 0.6481481481481481)

1

(0.4341224195120214, 0.896168108776267)

(0.3649017943276299, 0.9074074074074074)

2

(0.18619558424660096, 0.9765142150803461)

(0.16293201995668588, 0.9888888888888889)

3

(0.09108264647580784, 0.9962917181705809)

(0.08985773311858927, 0.9962962962962963)

4

(0.053487053708540566, 0.9987639060568603)

(0.05794332528279887, 1.0)

get_loss_acc(nn_model, X_test, Y_test)

(0.16834938257537793, 0.9428172942817294)

Zadanie domowe

• wybrać jedno z poniższych repozytoriów i je sforkować:
  • https://git.wmi.amu.edu.pl/kubapok/paranormal-or-skeptic-ISI-public
  • https://git.wmi.amu.edu.pl/kubapok/sport-text-classification-ball-ISI-public
• stworzyć klasyfikator bazujący na prostej sieci neuronowej feed forward w pytorchu (można bazować na tym jupyterze). Zamiast tfidf proszę skorzystać z jakieś reprezentacji gęstej (np. word2vec).
• stworzyć predykcje w plikach dev-0/out.tsv oraz test-A/out.tsv
• wynik accuracy sprawdzony za pomocą narzędzia geval (patrz poprzednie zadanie) powinien wynosić conajmniej 0.67
• proszę umieścić predykcję oraz skrypty generujące (w postaci tekstowej a nie jupyter) w repo, a w MS TEAMS umieścić link do swojego repo termin 25.05, 70 punktów