cnn.py

import matplotlib.pyplot as plt
plt.style.use('ggplot')

import os
import time

from typing import Iterable
from dataclasses import dataclass

import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

from torchvision import datasets, transforms

from torch.optim import lr_scheduler
from torch.optim.lr_scheduler import StepLR

training_data_path = "./data/training"
validation_data_path = "./data/validation"

classifier_name = "torsk_sei_classifier.pt"

def image_preprocess_transforms():

    preprocess = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()
        ])

    return preprocess

def image_common_transforms(mean=(0.4611, 0.4359, 0.3905), std=(0.2193, 0.2150, 0.2109)):
    preprocess = image_preprocess_transforms()

    common_transforms = transforms.Compose([
        preprocess,
        transforms.Normalize(mean, std)
    ])

    return common_transforms

def get_mean_std(data_root, num_workers=4):

    transform = image_preprocess_transforms()

    loader = data_loader(data_root, transform)

    mean = 0.
    std = 0.

    for images, _ in loader:
        batch_samples = images.size(0) # batch size (the last batch can have smaller size!)
        images = images.view(batch_samples, images.size(1), -1)
        mean += images.mean(2).sum(0)
        std += images.std(2).sum(0)

    mean /= len(loader.dataset)
    std /= len(loader.dataset)

    print('mean: {}, std: {}'.format(mean, std))

    return mean, std

def data_loader(data_root, transform, batch_size=16, shuffle=False, num_workers=2):
    dataset = datasets.ImageFolder(root=data_root, transform=transform)

    loader = torch.utils.data.DataLoader(dataset,
                                         batch_size=batch_size,
                                         num_workers=num_workers,
                                         shuffle=shuffle)

    return loader

def get_data(batch_size, data_root, num_workers=4, data_augmentation=False):

    train_data_path = os.path.join(data_root, 'training')

    mean, std = get_mean_std(data_root=train_data_path, num_workers=num_workers)

    common_transforms = image_common_transforms(mean, std)


    # if data_augmentation is true
    # data augmentation implementation
    if data_augmentation:
        train_transforms = data_augmentation_preprocess(mean, std)
    # else do common transforms
    else:
        train_transforms = common_transforms


    # train dataloader

    train_loader = data_loader(train_data_path,
                               train_transforms,
                               batch_size=batch_size,
                               shuffle=True,
                               num_workers=num_workers)

    # test dataloader

    test_data_path = os.path.join(data_root, 'validation')

    test_loader = data_loader(test_data_path,
                              train_transforms,
                              batch_size=batch_size,
                              shuffle=True,
                              num_workers=num_workers)

    return train_loader, test_loader

@dataclass
class SystemConfiguration:
    '''
    Describes the common system setting needed for reproducible training
    '''
    seed: int = 21  # seed number to set the state of all random number generators
    cudnn_benchmark_enabled: bool = True  # enable CuDNN benchmark for the sake of performance
    cudnn_deterministic: bool = True  # make cudnn deterministic (reproducible training)

@dataclass
class TrainingConfiguration:
    '''
    Describes configuration of the training process
    '''
    batch_size: int = 10
    epochs_count: int = 100
    init_learning_rate: float = 0.1
    log_interval: int = 5
    test_interval: int = 1
    data_root: str = "./data"
    num_workers: int = 2
    device: str = 'cuda'

def setup_system(system_config: SystemConfiguration) -> None:
    torch.manual_seed(system_config.seed)
    if torch.cuda.is_available():
        torch.backends.cudnn_benchmark_enabled = system_config.cudnn_benchmark_enabled
        torch.backends.cudnn.deterministic = system_config.cudnn_deterministic

def train(
    train_config: TrainingConfiguration, model: nn.Module, optimizer: torch.optim.Optimizer,
    train_loader: torch.utils.data.DataLoader, epoch_idx: int
) -> None:

    # change model in training mood
    model.train()

    # to get batch loss
    batch_loss = np.array([])

    # to get batch accuracy
    batch_acc = np.array([])

    for batch_idx, (data, target) in enumerate(train_loader):

        # clone target
        indx_target = target.clone()
        # send data to device (its is medatory if GPU has to be used)
        data = data.to(train_config.device)
        # send target to device
        target = target.to(train_config.device)

        # reset parameters gradient to zero
        optimizer.zero_grad()

        # forward pass to the model
        output = model(data)

        # cross entropy loss
        loss = F.cross_entropy(output, target)

        # find gradients w.r.t training parameters
        loss.backward()
        # Update parameters using gardients
        optimizer.step()

        batch_loss = np.append(batch_loss, [loss.item()])

        # Score to probability using softmax
        prob = F.softmax(output, dim=1)

        # get the index of the max probability
        pred = prob.data.max(dim=1)[1]

        # correct prediction
        correct = pred.cpu().eq(indx_target).sum()

        # accuracy
        acc = float(correct) / float(len(data))

        batch_acc = np.append(batch_acc, [acc])

    # Decay learning rate
    scheduler.step()
    print('Stepping scheduler this epoch. ', 'LR:', scheduler.get_lr())

    epoch_loss = batch_loss.mean()
    epoch_acc = batch_acc.mean()
    print('Epoch: {} \nTrain Loss: {:.6f} Acc: {:.4f}'.format(epoch_idx, epoch_loss, epoch_acc))
    return epoch_loss, epoch_acc

def validate(
    train_config: TrainingConfiguration,
    model: nn.Module,
    test_loader: torch.utils.data.DataLoader,
) -> float:
    #
    model.eval()
    test_loss = 0
    count_corect_predictions = 0
    for data, target in test_loader:
        indx_target = target.clone()
        data = data.to(train_config.device)

        target = target.to(train_config.device)

        output = model(data)
        # add loss for each mini batch
        test_loss += F.cross_entropy(output, target).item()

        # Score to probability using softmax
        prob = F.softmax(output, dim=1)

        # get the index of the max probability
        pred = prob.data.max(dim=1)[1]

        # add correct prediction count
        count_corect_predictions += pred.cpu().eq(indx_target).sum()

    # average over number of mini-batches
    test_loss = test_loss / len(test_loader)

    # average over number of dataset
    accuracy = 100. * count_corect_predictions / len(test_loader.dataset)

    print(
        '\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
            test_loss, count_corect_predictions, len(test_loader.dataset), accuracy
        )
    )

    return test_loss, accuracy/100.0

def save_model(model, device, model_dir='models', model_file_name=classifier_name):


    if not os.path.exists(model_dir):
        os.makedirs(model_dir)

    model_path = os.path.join(model_dir, model_file_name)

    # make sure you transfer the model to cpu.
    if device == 'cuda':
        model.to('cpu')

    # save the state_dict
    torch.save(model.state_dict(), model_path)

    if device == 'cuda':
        model.to('cuda')

    return

def load_model(model, model_dir='models', model_file_name=classifier_name):
    model_path = os.path.join(model_dir, model_file_name)

    # loading the model and getting model parameters by using load_state_dict
    model.load_state_dict(torch.load(model_path))

    return model

def main(model, optimizer, scheduler=None, system_configuration=SystemConfiguration(),
         training_configuration=TrainingConfiguration(), data_augmentation=True):

    # system configuration
    setup_system(system_configuration)

    # batch size
    batch_size_to_set = training_configuration.batch_size
    # num_workers
    num_workers_to_set = training_configuration.num_workers
    # epochs
    epoch_num_to_set = training_configuration.epochs_count

    # if GPU is available use training config,
    # else lowers batch_size, num_workers and epochs count
    if torch.cuda.is_available():
        device = "cuda"
    else:
        device = "cpu"
        batch_size_to_set = 16
        num_workers_to_set = 4

    # data loader
    train_loader, test_loader = get_data(
        batch_size=batch_size_to_set,
        data_root=training_configuration.data_root,
        num_workers=num_workers_to_set,
        data_augmentation=data_augmentation
    )

    # Update training configuration
    training_configuration = TrainingConfiguration(
        device=device,
        batch_size=batch_size_to_set,
        num_workers=num_workers_to_set
    )

    # send model to device (GPU/CPU)
    model.to(training_configuration.device)

    best_loss = torch.tensor(np.inf)

    # epoch train/test loss
    epoch_train_loss = np.array([])
    epoch_test_loss = np.array([])

    # epch train/test accuracy
    epoch_train_acc = np.array([])
    epoch_test_acc = np.array([])

    # trainig time measurement
    t_begin = time.time()
    for epoch in range(training_configuration.epochs_count):

#         Calculate Initial Test Loss
        init_val_loss, init_val_accuracy = validate(training_configuration, model, test_loader)
        print("Initial Test Loss : {:.6f}, \nInitial Test Accuracy : {:.3f}%\n".format(init_val_loss, init_val_accuracy*100))

        # Train
        train_loss, train_acc = train(training_configuration, model, optimizer, train_loader, epoch)

        epoch_train_loss = np.append(epoch_train_loss, [train_loss])

        epoch_train_acc = np.append(epoch_train_acc, [train_acc])

        elapsed_time = time.time() - t_begin
        speed_epoch = elapsed_time / (epoch + 1)
        speed_batch = speed_epoch / len(train_loader)
        eta = speed_epoch * training_configuration.epochs_count - elapsed_time

        print(
            "Elapsed {:.2f}s, {:.2f} s/epoch, {:.2f} s/batch, ets {:.2f}s".format(
                elapsed_time, speed_epoch, speed_batch, eta
            )
        )

        # Validate
        if epoch % training_configuration.test_interval == 0:
            current_loss, current_accuracy = validate(training_configuration, model, test_loader)

            epoch_test_loss = np.append(epoch_test_loss, [current_loss])

            epoch_test_acc = np.append(epoch_test_acc, [current_accuracy])

            if current_loss < best_loss:
                best_loss = current_loss
                print('Model Improved. Saving the Model...\n')
                save_model(model, device=training_configuration.device)


    print("Total time: {:.2f}, Best Loss: {:.3f}".format(time.time() - t_begin, best_loss))

    return model, epoch_train_loss, epoch_train_acc, epoch_test_loss, epoch_test_acc

def plot_loss_accuracy(train_loss, val_loss, train_acc, val_acc, colors,
                       loss_legend_loc='upper center', acc_legend_loc='upper left',
                       fig_size=(20, 10), sub_plot1=(1, 2, 1), sub_plot2=(1, 2, 2)):

    plt.rcParams["figure.figsize"] = fig_size
    fig = plt.figure()

    plt.subplot(sub_plot1[0], sub_plot1[1], sub_plot1[2])

    for i in range(len(train_loss)):
        x_train = range(len(train_loss[i]))
        x_val = range(len(val_loss[i]))

        min_train_loss = train_loss[i].min()

        min_val_loss = val_loss[i].min()

        plt.plot(x_train, train_loss[i], linestyle='-', color='tab:{}'.format(colors[i]),
                 label="TRAIN LOSS ({0:.4})".format(min_train_loss))
        plt.plot(x_val, val_loss[i], linestyle='--' , color='tab:{}'.format(colors[i]),
                 label="VALID LOSS ({0:.4})".format(min_val_loss))

    plt.xlabel('epoch no.')
    plt.ylabel('loss')
    plt.legend(loc=loss_legend_loc)
    plt.title('Training and Validation Loss')

    plt.subplot(sub_plot2[0], sub_plot2[1], sub_plot2[2])

    for i in range(len(train_acc)):
        x_train = range(len(train_acc[i]))
        x_val = range(len(val_acc[i]))

        max_train_acc = train_acc[i].max()

        max_val_acc = val_acc[i].max()

        plt.plot(x_train, train_acc[i], linestyle='-', color='tab:{}'.format(colors[i]),
                 label="TRAIN ACC ({0:.4})".format(max_train_acc))
        plt.plot(x_val, val_acc[i], linestyle='--' , color='tab:{}'.format(colors[i]),
                 label="VALID ACC ({0:.4})".format(max_val_acc))

    plt.xlabel('epoch no.')
    plt.ylabel('accuracy')
    plt.legend(loc=acc_legend_loc)
    plt.title('Training and Validation Accuracy')

    fig.savefig('sample_loss_acc_plot.png')
    plt.show()

    return

classes = 3
nodes = 128

class MyModel(nn.Module):
    def __init__(self):
        super().__init__()

        # convolution layers
        self._body = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=64, kernel_size=7),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2),
            
            nn.Conv2d(in_channels=64, out_channels=nodes, kernel_size=5),
            nn.BatchNorm2d(nodes),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=2)
        )

        
        # Fully connected layers
        self._head = nn.Sequential(
            nn.Linear(in_features=nodes*52*52, out_features=1024), 
            nn.ReLU(inplace=True),
            
            nn.Linear(in_features=1024, out_features=classes)
            
        )
    
    def forward(self, x):
        self.drop_out = nn.Dropout()
        # apply feature extractor
        x = self._body(x)
        # flatten the output of conv layers
        # dimension should be batch_size * number_of weight_in_last conv_layer
        x = x.view(x.size()[0], -1)
        # apply classification head
        x = self._head(x)
        
        return x

def prediction(model, device, batch_input):

    # send model to cpu/cuda according to your system configuration
    model.to(device)

    # it is important to do model.eval() before prediction
    model.eval()

    data = batch_input.to(device)

    output = model(data)

    # Score to probability using softmax
    prob = F.softmax(output, dim=1)

    # get the max probability
    pred_prob = prob.data.max(dim=1)[0]

    # get the index of the max probability
    pred_index = prob.data.max(dim=1)[1]

    return pred_index.cpu().numpy(), pred_prob.cpu().numpy()

def get_sample_prediction(model, data_root, mean, std):
    batch_size = 15

    if torch.cuda.is_available():
        device = "cuda"
        num_workers = 8
    else:
        device = "cpu"
        num_workers = 2



    # transformed data
    test_dataset_trans = datasets.ImageFolder(root=data_root, transform=image_common_transforms(mean, std))

    # original image dataset
    test_dataset = datasets.ImageFolder(root=data_root, transform=image_preprocess_transforms())

    data_len = test_dataset.__len__()

    interval = int(data_len/batch_size)

    imgs = []
    inputs = []
    targets = []
    for i in range(batch_size):
        index = i * interval
        trans_input, target = test_dataset_trans.__getitem__(index)
        img, _ = test_dataset.__getitem__(index)

        imgs.append(img)
        inputs.append(trans_input)
        targets.append(target)

    inputs = torch.stack(inputs)

    cls, prob = prediction(model, device, batch_input=inputs)

    plt.style.use('default')
    plt.rcParams["figure.figsize"] = (15, 9)
    fig = plt.figure()


    for i, target in enumerate(targets):
        plt.subplot(3, 5, i+1)
        img = transforms.functional.to_pil_image(imgs[i])
        plt.imshow(img)
        plt.gca().set_title('P:{0}({1:.2}), T:{2}'.format(test_dataset.classes[cls[i]],
                                                     prob[i],
                                                     test_dataset.classes[targets[i]]))
    fig.savefig('sample_prediction.png')
    plt.show()

    return


if __name__ == '__main__':
    model = MyModel()
    #model = torchvision.models.resnet50()

    print(model)

    # get optimizer
    train_config = TrainingConfiguration()

    optimizer = torch.optim.Adam(params=model.parameters(), lr=train_config.init_learning_rate, betas=(0.9, 0.999), eps=1e-08, weight_decay=0, amsgrad=False)

    # optimizer
    #optimizer = optim.Adam(
    #    model.parameters(),
    #    lr = train_config.init_learning_rate
    #)

    #scheduler = StepLR(optimizer, step_size=1, gamma=0.1)

    decayRate = 0.96
    scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=decayRate)

    #optimizer = optim.SGD(
    #    model.parameters(),
    #    lr=train_config.init_learning_rate
    #)

    plot_loss_accuracy(train_loss=[train_loss],
                       val_loss=[val_loss],
                       train_acc=[train_acc],
                       val_acc=[val_acc],
                       colors=['blue'],
                       loss_legend_loc='upper center',
                       acc_legend_loc='upper left')


    m = MyModel()

    m = load_model(m)

    train_config = TrainingConfiguration()

    test_data_path = os.path.join(train_config.data_root, 'validation')

    train_data_path = os.path.join(train_config.data_root, 'training')

    mean, std = get_mean_std(train_data_path)


    get_sample_prediction(m, test_data_path, mean, std)