DEV Community: Quoc Bao

Convert my Pytorch model to Pytorch Lightning

Quoc Bao — Wed, 29 Jun 2022 16:52:43 +0000

Hello, everybody! Today I am going to show how you how to convert my model from Pytorch to Pytorch Lightning. Pytorch Lightning is a light-weight deep learning framework built upon Pytorch. It removes a lot of boilerplate code (standard code that can be found in almost any deep learning pipeline) and adds in many functions that helps to interfere training at a specific position.

Firstly, I import the libraries.

pip install pytorch-lightning
import pytorch_lightning as pl

Pytorch LightningModule resembles nn.Module. Forward function can be defined in a pl class.

# an nn class can be converted to a pl class by replacing nn with pl

class NeuralNet(nn.Module): 
# --> class NeuralNet(pl.LightningModule):
    def __init__(self, input_size, num_classes):
        super(NeuralNet, self).__init__()
        self.fc1 = nn.Linear(input_size, 50)
        self.fc2 = nn.Linear(50, num_classes)

# --> specific functions belong to nn class should not be changed!
    def forward(self, x):
        out = self.fc1(x)
        out = torch.sigmoid(out)
        out = self.fc2(out)
        return out

FruitsClassifier on Fruits 360

Quoc Bao — Mon, 27 Jun 2022 17:20:18 +0000

On this project, I build a classifier on the Fruits 360 dataset using Pytorch. I use a pretrained model and transfer learning, as well as do hyper-parameter search to help increase the accuracy.

✍️ Documentation:

- Full Training on Kaggle | GoogleNet | 94% Acc

- Efficient Training with Pytorch Lightning

- GitHub Page

📰 Dataset structure

train/test
    |___apple_braeburn_1
        |___r0_0.jpg
        |___r0_10.jpg
            ...
    |___carrot_1
        |___r0_0.jpg
        |___r0_10.jpg
        ...

Large batch visualization

Separate visualization

Single fruit visualization

Single fruit visualization with transforms

Separate visualization with transforms

Batch visualization with transforms

Pixel Intensity

Visualize the number of classes

Visualize loss and accuracy

Visualize model predictions: green(correct) and red(incorrect)

FruitsClassifier with Pytorch Lightning 🍉

Quoc Bao — Mon, 27 Jun 2022 17:19:04 +0000

On this project, I convert my FruitsClassifier from Pytorch to Pytorch Lightning. I employ a callback module, a checkpoint mode to store and save the model state_dict. I conduct resume training, plot loss and accuracy from previous trains. I check the accuracy for each class and make predictions on images from outside the dataset.

✍️ Documentation:

- Convert my Pytorch model to Pytorch Lightning

- FruitsClassifier on Fruits 360

- Full Training on Kaggle | GoogleNet | 94% Acc

- GitHub Page

💥 Training Epochs

🍇 Visualize Loss and Accuracy

🍌 Visualize model performance on test dataset

🍎 Accuracy of each class

🍍 Image Predictions from outside the Dataset

🍑 Accuracy of each class: green(correct predictions), red(incorrect predictions)

My Pytorch Training Pipeline in Image Classification

Quoc Bao — Sat, 11 Jun 2022 10:21:12 +0000

On this project, I create a custom dataset of 5 male models and conduct a full Pytorch training pipeline. I use a pretrained model and transfer learning, as well as do hyper-parameter search to help increase the accuracy.

✍️ Analysis and Evaluation:

Full Pytorch Training Pipeline on Image Classification part1
Full Pytorch Training Pipeline on Image Classification part2

📺 My implementations are based off Aladdin Persson and Python Engineer.

My repository contains:

A training script (using pretrained vgg16 and transfer learning)
A script for Hyper-parameter Search
A script for loading the model for either resumed training or inference
A trained model 😬
Some helper functions
A dataset

📰 Dataset Structure

train/val
|___chau_minh_chi
       |___chau_minh_chi_01.jpg
          |___chau_minh_chi_02.jpg
          ...
|___keita_machida
       |___keita_machida_01.jpg
          |___keita_machida_02.jpg
          ...

Small batch visualization

Large batch visualization

Visualize the number of classes

😑 Hyper-parameter Search

💥 Training Epochs

==> Saving new best
Epoch 1/25
Step 34/34, train Loss = 1.84,  train Acc = 0.29
Step 20/20, val loss = 1.58,  val acc = 0.25
Time spent for this epoch -----> 0m 32s

==> Saving new best
Epoch 2/25
Step 34/34, train Loss = 0.91,  train Acc = 0.63
Step 20/20, val loss = 1.53,  val acc = 0.43
Time spent for this epoch -----> 0m 13s

==> Validation accuracy did not improve.
Epoch 3/25
Step 34/34, train Loss = 0.61,  train Acc = 0.82
Step 20/20, val loss = 1.67,  val acc = 0.27
Time spent for this epoch -----> 0m 8s

😌 Visualize Loss and Accuracy

😅 Accuracy of Each Class

Test Acc
Got 13/30 correct samples over 43.33%
Accuracy of timmy_xu: 33.33%
Accuracy of corbyn_besson: 62.50%
Accuracy of keita_machida: 16.67%
Accuracy of wang_kai: 30.00%
Accuracy of chau_minh_chi: 100.00%

😞 Classification Report and Confusion Matrix Heatmap

😍 Predict a single image

😐 Prediction In Batch

Binary Search Tree Class Implementation

Quoc Bao — Mon, 02 May 2022 14:01:36 +0000

Hello everybody! Today, I will show you an implementation of the binary search tree (BST) class which borrows the structure of a linked list but is a little bit different. The BST has two reference fields (left subtree and right subtree) instead of just only one. Though I have implemented a linked list class before, it has not quite come as an official post. Therefore, if you want to compare between the two, see it for yourself.

Without further ado, let’s get on!

struct node
{
    int info;
    node *left;
    node *right;
};

class BST
{
private:
    node *root;
    void deepcopy(const BST &other);

public:
    // node *search(int);
    // node *search(node *, int);
    BST();
    ~BST();
    BST(int);
    BST(const BST &other);
    BST &operator=(const BST &other);
    void insert(int);
    void insert(node *, int);
};

Read more here!

String class implementation

Quoc Bao — Wed, 27 Apr 2022 09:24:08 +0000

Hello everybody! Today, I would like to show you a quick implementation of a string class.

A string is an array of characters. So, we initialize it with an array of char type and a length.

class STRING
{
private:
    char *content;
    int len;

    void deepcopy(const STRING &other);

public:
    STRING();
    // ~STRING();
    STRING(std::string);

    STRING(const STRING &other);
    STRING &operator=(const STRING &other);

    bool operator==(STRING &other);
    bool operator==(std::string);
    bool operator!=(STRING other);

    char getContent(int);
    char *getWholeContent();
    void deleteContent();
    void allocateContent(int);
    int getLen();
    void setLen(int);
    void setContent(int, char);

    void input();
    void output();

    void removeBlankHead();
    void removeBlankTail();
    void RemoveExtraSpaces();

    friend std::istream &operator>>(std::istream &in, STRING &a);
    friend std::ostream &operator<<(std::ostream &out, STRING a);
};

Read more here!

Matrix class implementation

Quoc Bao — Sun, 17 Apr 2022 17:02:13 +0000

The concept of matrix and multidimensional array had always been an intriguing topic for me when I was in programming classes. Since computer memory isn’t organized in tabular form, there should be a way to represent it. Today, I am going to show you how to implement a matrix class with a component of an array. It will simply has some basic operations of matrix such as addition, subtraction, and multiplication.

class Matrix
{
private:
    Array *A;
    int rows = 0, cols = 0;

    void deepcopy(const Matrix &other);

public:
    Matrix();
    Matrix(int x, int y);
    void input();
    void output();

    void insert(int, int, int);
    int findCelebrity();

    // ~Matrix();

    Matrix(const Matrix &other);
    Matrix &operator=(const Matrix &other);
    void create_identity(int);
    void create_celeb_test(int);

    // Matrix &operator+(const Matrix &other);
    Matrix &operator-(const Matrix &other);

    int view_rows();
    int view_cols();
    int getElement(int, int);
    void setElement(int, int, int);
    bool knows(int, int);

    friend istream &operator>>(istream &in, Matrix &a);
    friend ostream &operator<<(ostream &out, Matrix a);
};

Read more here!

Introduction to OOP and array class implementation [part 2]

Quoc Bao — Sun, 20 Mar 2022 02:53:17 +0000

Last time, we left off with the default constructor. Remember that it has no arguments, which is why it is also called non-parameterized constructor. Therefore, this time we’ll kick off with parameterized constructor, which takes in user’s initial values.

Array::Array(int num_elements)
{
    allocate(this->A, num_elements);
    this->na = 0;
    this->capacity_a = num_elements;
}

Read more here!

Introduction to OOP and array class implementation [part 1]

Quoc Bao — Sun, 20 Mar 2022 02:50:33 +0000

Object-oriented programming (OOP) is an important knowledge taught in schools and programming classes. It is a concept that relies on the creation of classes and objects which helps coding more efficient with reusable code blocks of well-defined functions. Today, I will show you how to implement a simple 1D array class. Throughout the post, important OOP terminologies and concepts will be discussed for the sake of better understanding.

class Array
{
private:
    int *A;
    int na = 0;
    int capacity_a = 1;

    void deepcopy(const Array &a);

public:
    Array();
    Array(int);
    ~Array();

    void insert(int);
    void print();

    Array(const Array &other);
    Array &operator=(const Array &other);
    Array operator+(Array other);

    int view_na();
    void set_na(int);
    int getElement(int index);
    void setElement(int index, int value);

    void quicksort(int left, int right);

    friend istream &operator>>(istream &in, Array &a);
    friend ostream &operator<<(ostream &out, Array a);
};

Read more here!

Full Pytorch Training Pipeline on Image Classification [ part2 ]

Quoc Bao — Fri, 25 Feb 2022 14:25:27 +0000

Hello everybody! Today, we will continue with what we left off in the previous blog.

I first talk about the Tensorboard. Tensorboard is a visualization toolkit of TensorFlow, which is also a machine learning and deep learning framework. Tensorboard can be used for displaying the images in the dataset, drawing loss and accuracy graphs, analyzing the components of a neural network.

Continue to read this here!

Full Pytorch Training Pipeline on Image Classification [ part1 ]

Quoc Bao — Fri, 25 Feb 2022 14:08:34 +0000

Hello everybody! Today, I will show you how to train a full deep learning pipeline on image classification using Pytorch. For somebody who didn’t know Pytorch, it is a machine learning and deep learning framework similar to Keras or TensorFlow. It is a very fun and cool framework to work with, especially on understanding the structure of a neural network.

Continue to read it here!