Preprocessing images about kerasgen

[ma7555]

Hello @jaimerodric ,

Sorry for the late response. As the output is a normal tf.data.Dataset you can use the apply function normally after creating the dataset from the generator.

[jaimerodric]

@ma7555
Thanks for your response! That's a smart solution. If I am not wrong, the images are loaded as RGB, right? I want to confirm this because when testing, if I load images with opencv, they are in BGR form and should I convert it ro RGB?

I am trying to train a network from scratch and I am struggling, I do not know if it is because working with only 200k images, the generator, the network...

[ma7555]

The images are loaded in RGB as tf.Tensor, to use a preprocessing function you need to use the apply method I mentioned but make sure all the transformations are done using tf ops and not numpy.
If you need to use numpy or other pythonic functions please refer to tf.numpy_function and tf.function

[jaimerodric]

@ma7555
Thanks again for the response. What margin have you tested that is the better for training? Because the TripletSemiHardLoss, it is by default established on 1, but from my point of view it does not make sense when there is a normalization L2 at the end of the network (normalized output of 1)

[ma7555]

I have not experimented with changing the margin. But keeping it default worked well for me.

[jaimerodric]

Hi again @ma7555 !
I am struggling applying the preprocessing function. I have implemented one that is like this:

def pre_process_facenet(dataset):
x_images = []
for ret in dataset:
# standardize pixel values across channels (global)
mean, std = ret.mean(), ret.std()
ret = (ret - mean) / std
x_images.append(ret)
return np.array(x_images)

train_ds = balanced_image_dataset_from_directory(
directory, num_classes_per_batch=5,
num_images_per_class=5, image_size=(160, 160),
seed=6, validation_split=0.2, subset='training',
safe_triplet=True)

But when using apply on the dataset generated with kerasgen I got some errors
train_ds = train_ds.apply(tf.numpy_function(pre_process_facenet, train_ds, tf.float32))
ValueError: could not broadcast input array from shape (25,160,160,3) into shape (25,)

What I am doing wrong? I want to apply pre_process_facenet in train_ds and then into the fit method of TF

[simkarwin]

mimo-Keras package (https://pypi.org/project/mimo-keras/) can help you to create your pipeline without additional coding