Hello. I have images dataset of ECG Signal which has 6 classes but the classes are imbalanced. Now I wanna use data augmentation on my dataset to balance the classes. You know ECG Signal needs to be augmented to have a benefit so I do not see it benefiting by croping, rotating etc so Im doing scaling, translation. My goal is these two techniques. I have read about this in pytorch and came to know that transforms.affine has attributes that can perform scaling and translation.
Now my query is when I apply it like:
from torchvision import transforms as T, datasets
import os
import random
Define the data augmentation transformations
train_transform = T.Compose([
T.RandomApply([T.RandomAffine(degrees=0, translate=(0.1, 0), scale=(0.9, 1.1))], p=0.5),
T.Resize(size=(CFG.img_size,CFG.img_size)),
T.ToTensor(), # convert the dimension from (height,weight,channel) to (channel,height,weight) convention of PyTorch
T.Grayscale(1)
])
Define the validation and test transformations
validate_transform = T.Compose([
T.RandomApply([T.RandomAffine(degrees=0, translate=(0.1, 0), scale=(0.9, 1.1))], p=0.5),
T.Resize(size=(CFG.img_size,CFG.img_size)),
T.ToTensor(),
T.Grayscale(1)
])
test_transform = T.Compose([
T.RandomApply([T.RandomAffine(degrees=0, translate=(0.1, 0), scale=(0.9, 1.1))], p=0.5),
T.Resize(size=(CFG.img_size,CFG.img_size)),
T.ToTensor(),
T.Grayscale(1)
])
But my images looked like this :
by running the code: dataiter = iter(trainloader)
images,labels = next(dataiter)
out = make_grid(images,nrow=4)
show_grid(out,title = [class_name[x] for x in labels])
Now my problem is how can I fix it? why it is happening ? besides as I said i need class balance but how will I ensure that classes got balanced. ? I have say 9000 images in train 1000 in validation and 4000 in test. i need to see my dataset getting large by balancing classes by introducing translation or scaling.
I tried this code chunk but as PIL needs rgb image I get error I have to work with grayscale Images through out for my application.
from PIL import Image
Define the number of images you want per class
num_images_per_class = 2200
going to be used for loading dataset
train_path=‘/content/train’
validate_path= ‘/content/val’
test_path=‘/content/ECG_Image_data/test’
Loop through the train, val, and test folders
for folder_path in [train_path,validate_path,test_path]:
# Loop through each sub-folder (class) in the current folder
for class_folder in os.listdir(folder_path):
class_path = os.path.join(folder_path, class_folder)
# Count the number of images in the current class
num_images = len(os.listdir(class_path))
# Calculate the number of images to generate
num_images_to_generate = num_images_per_class - num_images
# Generate new images by applying the transform to existing images
for i in range(num_images_to_generate):
# Choose a random image from the class folder
image_path = os.path.join(class_path, random.choice(os.listdir(class_path)))
# Load the image
image = Image.open(image_path)
# Apply the transform
transformed_image = train_transform(image)
# Save the transformed image with a new name
new_image_name = f"{class_folder}transformed{i}.png"
transformed_image.save(os.path.join(class_path, new_image_name))
Plz help me solve this. I can’t fix it.
