我的 Pytorch 卷积神经网络中的每个时期是否应用了随机变换？ （数据增强）

Question

I'm new to Pytorch and I try to make a convolutional neural net to classify a set of images (personal iris recognition problem). My issue is that I have a small number of images (10 classes and 20 images per class). I tried to make data augmentation (random transforms for every epoch) but I'm not sure that these are applied at each epoch as I entended. Here's my code. If anyone can confirm that I'm doing it right or if it's not ok, is there a way to make the transforms inside the loop?

from torch import utils, nn, optim, no_grad
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from ConvNet import ConvNet
from ImagesDataset import ImagesDataset, AddGaussianNoise

DATABASE_PATH = "C://Users//Maria//Downloads//ees//CASIA-IrisV2"
MODEL_PATH = "entire_model.pt"
dataArray = []

# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# data augmentation by applying some transforms randomly for every batch
transform = transforms.Compose([transforms.RandomCrop(5), transforms.RandomHorizontalFlip(p=0.1),
                                transforms.ColorJitter(brightness=0.1, contrast=0.2, saturation=0, hue=0),
                                AddGaussianNoise(0.1, 0.05), transforms.ToTensor()])
dataset = ImagesDataset(csv_file="generate_csv//generate_csv_correctly_detected.csv", root_dir=DATABASE_PATH, transform=transforms.ToTensor())
num_epochs = 300
num_classes = 10
batch_size = 100
learning_rate = 0.01

# the dataset is partitioned in 5 subsets to perform cross validation
sum_percents = 0
data_set = utils.data.random_split(dataset, [40, 40, 40, 40, 40])

for i in range(5):
    test_set = data_set[i]
    train_set = []
    for j in range(5):
        if j != i:
            train_set += data_set[j]

    train_loader = DataLoader(dataset=train_set, batch_size=batch_size, shuffle=True)
    test_loader = DataLoader(dataset=test_set, batch_size=batch_size, shuffle=True)

    model = ConvNet(0).to(device)

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=learning_rate)
    
    # Train the model
    total_step = len(train_loader)
    loss_list = []
    acc_list = []

    # delete contents of loss1 file
    file = open("loss1.txt", "r+")
    file.truncate(0)
    file.close()

    for epoch in range(num_epochs):
        print("Epoch: " + str(epoch))
        for i, (images, labels) in enumerate(train_loader):
            # Run the forward pass
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images)
            loss = criterion(outputs, labels)  
            # set the gradients to zero
            optimizer.zero_grad()
            # compute gradients
            loss.backward()
            # update the parameters
            optimizer.step()

            # Track the accuracy
            total = labels.size(0)
            _, predicted = torch.max(outputs.data, 1)
            correct = (predicted == labels).sum().item()
            acc_list.append(correct / total)

    # Save
    torch.save(model, MODEL_PATH)

    # Test the model
    model.eval()
    with no_grad():
        correct = 0
        total = 0
        for images, labels in test_loader:
            outputs = model(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()

        print('Test Accuracy of the model on the 132 test images: {} %'.format((correct / total) * 100))
    sum_percents += (correct / total) * 100
print('Average accuracy is {}%'.format((sum_percents/5)))

Answer 1

Hi what i meant wasn't like that but the following and i cannot completely reproduce since i dont have your function AddGaussianNoise

import torchvision.transforms as T
import numpy as np

transforms = T.Compose([
    T.ToPILImage(), # You need to add this to pil image
    T.RandomCrop(5), T.RandomHorizontalFlip(p=0.1),
    T.ColorJitter(brightness=0.1, contrast=0.2, saturation=0, hue=0),
    T.ToTensor()
])

transforms(np.random.randn(224, 224, 3).astype(np.uint8))
>>>tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0039],
         [0.0000, 0.0000, 0.9882, 0.0000, 0.9882],
         [0.0039, 0.9882, 0.9882, 0.0000, 0.9882],
         [0.0000, 0.0039, 0.0000, 0.0000, 0.0000]],

        [[0.0039, 0.0000, 0.0000, 0.0039, 0.9882],
         [0.9882, 0.0000, 0.0000, 0.0000, 0.9882],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.9882, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.9882, 0.0000, 0.0000, 0.0039]],

        [[0.0000, 0.9882, 0.0000, 0.9882, 0.0000],
         [0.0000, 0.0039, 0.0000, 0.0000, 0.0000],
         [0.0039, 0.0000, 0.0000, 0.0000, 0.0039],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0039, 0.0039, 0.0000, 0.0000, 0.0000]]])

So this is another assumption but a transform should work like this no? Since i dont have any of your code so here it is

import torchvision.transforms as T

transforms = T.Compose([
    T.ToPILImage(), # You need to add this to pil image
    T.RandomCrop(5), T.RandomHorizontalFlip(p=0.1),
    T.ColorJitter(brightness=0.1, contrast=0.2, saturation=0, hue=0),
    T.ToTensor(), # Maybe you can add you gaussian noise augment here
])
dataset = ImagesDataset(csv_file="generate_csv//generate_csv_correctly_detected.csv", root_dir=DATABASE_PATH, transform=transforms)

Answer 2

I tried adding the ToPILImage() transform (without the gaussian noise transform) and the above errors disappear but I get a size mismatch later when the images are fed in the neural net. Is there any possibility that one of the transforms changes the image size?

RuntimeError: size mismatch, m1: [100 x 32], m2: [55296 x 1000] at ..\aten\src\TH/generic/THTensorMath.cpp:41

the code for the Gaussian Noise transform is (currently not used)

class AddGaussianNoise(object):
    def __init__(self, mean=0., std=1.):
        self.std = std
        self.mean = mean

    def __call__(self, tensor):
        return tensor + torch.randn(tensor.size()) * self.std + self.mean

    def __repr__(self):
        return self.__class__.__name__ + '(mean={0}, std={1})'.format(self.mean, self.std)