Train with smaller classes about rotnet HOT 7 OPEN

Hi @KazuhideMimura , thanks for the reply.
error is fixed, I just gave the value for parameter nb_classes in RotNetDataGenerator class as 4.
But after training the model, while testing I'm not getting the proper output i.e., the images are as it is, they are not rotating.
I'm using custom data and I'm giving the raw images, do I need to manipulate the images before training?

from rotnet.

HI @KazuhideMimura , I tried to train the model for 4 classes, its giving
ValueError: Error when checking target: expected fc360 to have shape (4,) but got array with shape (360,).
do we need to change any files other than utils.py

from rotnet.

Hi @HAsarvesh , thank you for the repo.
Would you provide me with a detailed error log? I'd like to know where the error occurred.

from rotnet.

You don't need add special treatment to the original image. I think the possibilities are:

1. Training did not go well
2. nb_classes is not designated

For example, if you run a function display example...

display_examples(
    model, 
    test_filenames,
    num_images=num_images,
    size=(224, 224),
    crop_center=True,
    crop_largest_rect=True,
    preprocess_func=preprocess_input,
    nb_classes=4, # designate nb_class
)

Please note that nb_class will be set at 360 unless designated. This is because I didn't want to disturb the original concept.

from rotnet.

Thanks for the clarification, I'll train it again and check.

from rotnet.

Here?

    predictions = model.predict_generator(
        RotNetDataGenerator(
            image_paths,
            input_shape=(224, 224, 3),
            batch_size=64,
            one_hot=True,
            preprocess_func=preprocess_input,
            rotate=False,
            crop_largest_rect=True,
            crop_center=True,
            nb_classes = model.layers[-1].units, # added. Not sure it works, but statement is required to obtain model's class number.
        ),
        val_samples=len(image_paths)
    )

from rotnet.

Thanks @KazuhideMimura, I'll try this.

from rotnet.