Loading .nii images while saving memory

[BRN]

Hello everybody,
for my Deep Learning exam, I have to develop a project that provides for the generation of 3D images in format .nii using the DCGAN algorithm trained with some real images of MRI scans of brains of patients with Alzahimer.

I have a serious problem. I should load three different datasets that weigh 3GB, 5GB and 8GB respectively. Unfortunately I am forced to use an old PC with only 4GB of RAM and 2GB of Swap memory, so I am impossible to upload the files at one time using this simple code:

train_data = []
data_path = './data/ADNI_test/'
for i in range(len(filenames)):
    mri_file = data_path + filenames[i]
    train_data.append(nib.load(mri_file).get_fdata())

Would any of you know how to give me some alternative solution? Is there a way to upload the files a little at a time without overloading memory? In DCGAN algorithm the batch size is set to 64 files, but I will certainly have to decrease to 30.

Thank you!

 

[pbuk]

I wouldn't have thought it would be necessary to have all the image files in memory; can you provide the ML engine with an iterator that loads each file on demand?

 

[BRN]

I hope it can be done, but at the moment I don't know how.
Should the iterator be implemented directly in the loss function of the GAN discriminator?

def discriminator_model(strides, kernel_size, input_img, weight_initializer, downsample_layers):
    rate = 0.2
    filters = input_img.shape[1]
    
    model = Sequential()

    model.add(tf.keras.layers.Conv3D(strides, kernel_size, filters, input_img.shape, padding = 'same', 
                                     kernel_initializer = weight_initializer))
    model.add(tf.keras.layers.LeakyReLU())
    model.add(tf.keras.layers.Dropout(rate = rate))

    for l in range(downsample_layers - 1):
        filters = int(filters * 2)
        model.add(tf.keras.layers.Conv3D(strides, kernel_size, filters, padding = 'same'))
        model.add(tf.keras.layers.LeakyReLU())
        model.add(tf.keras.layers.Dropout(rate = rate))
        
    model.add(tf.keras.layers.Flatten())
    model.add(tf.keras.layers.Dense(1))
    return model

def discriminator_loss(real_output, fake_output):
    real_loss = cross_entropy(tf.ones_like(real_output), real_output)
    fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)
    total_loss = real_loss + fake_loss
    return total_loss

where 'real_output' would be the real image that is compared with that artificially generated.

Some idea?

 

[pbuk]

You are using keras? I don't have a development environment handy but this might help https://www.tensorflow.org/api_docs/python/tf/keras/utils/image_dataset_from_directory