The loaded model behaves inconsistently

First of all, thank you for this awesome library!

I exported, and imported the model with this code:

# save weights
model.save_weights("model.h5")

from keras_segmentation.models.unet import vgg_unet

# load model
model_weight_path = 'model.h5'
model = vgg_unet(n_classes=6,  input_height=640, input_width=640)
model.load_weights(model_weight_path, by_name=True)

In some cases when I load the model, I got good prediction results that is middle in the table, and then if I use the same code, the same model, I got this strange result which is bottom in the table.

I use Google Colab, Keras 2.1.0. keras-segmentation-0.2.0

I installed keras-segmentation with pip. I tried with Keras 2.2.5 as well, and I got the same strange behavior.

What can it cause this phenomenon?

Thanks, for your help!

Original image
Good result
Bad result

[divamgupta]

@bessszilard , replace model.load_weights(model_weight_path, by_name=True) with model.load_weights(model_weight_path) and it should work fine!

[divamgupta]

That is strange. Could you share some reproducible snippet or the maybe the collab notebook? I can take a look.

One reason causing the problem could be that weights of all the layers are not being loaded properly. That could be checked by comparing the weights.