Uncertain behavior when used transform.py

Hey. Using the https://github.com/fizyr/keras-retinanet/blob/master/keras_retinanet/utils/transform.py#L236 at the training stage, I encountered strange behavior, the error increases with each epoch:

Epoch 1/50 car 0.1321 mAP: 0.1321 Epoch 2/50 car 0.1134 mAP: 0.1134 Epoch 3/50 car 0.1032 mAP: 0.1032 Epoch 4/50 car 0.0870 mAP: 0.0870

I use the following conversion options:

def random_transform(
    min_rotation=0,
    max_rotation=0.6,
    min_translation=(0, 0),
    max_translation=(0.2, 0.2),
    min_shear=0,
    max_shear=1,
    min_scaling=(0.8, 0.8),
    max_scaling=(1.5, 1.5),
    flip_x_chance=0.5,
    flip_y_chance=0.5,
    prng=DEFAULT_PRNG
) 

return np.linalg.multi_dot([
        random_rotation(min_rotation, max_rotation, prng),
        random_scaling(min_scaling, max_scaling, prng),
        random_flip(flip_x_chance, flip_y_chance, prng)
    ])

At the same time, if I use only the flip_x / flip_y, everything is fine:

Epoch 00001 car 0.7827 mAP: 0.7827 Epoch 00002 car 0.8040 mAP: 0.8040 Epoch 00003 car 0.8143 mAP: 0.8143

What could be the reason for this behavior, and how do I use data augmentation?

return random_flip(flip_x_chance, flip_y_chance, prng)

[de-vri-es]

@Maswor: I don’t think that is the same issue. If you have a recurring problem, could you open a new issue?

[hgaiser]

Closing due to inactivity.