load re-trained model error

[leviome]

at the beginning, 'demo.py' ran successfully with pretrained model.
then, after preparing and converting data, I retrained the model by

python train.py --dataset ./DIV2K_BIN --outdir ./output --profile wdsr-a-8 --scale 2

it ran normally for 300 epochs, and I got many new h5 files. when I loaded one of new h5 files, error occurred:

TypeError: Expected float32, got {'type': 'ndarray', 'value': [114.44399999999999, 111.4605, 103.02000000000001]} of type 'dict' instead.
could you offer me some suggestions?

[krasserm]

Please provide details how you loaded the h5 file.

[leviome]

python demo.py -i ./demo -o ./output --model ./output/20181204-230641/models/epoch-300-psnr-37.7157.h5
'epoch-300-psnr-37.7157.h5' is the new h5 file. And,
python bench.py -i ./output/20181204-230641/models -o bench.json had the same error information.

[krasserm]

I'm afraid I cannot reproduce your problem i.e. the commands you provided work as expected in my environment. What versions of Keras and Tensorflow do you use?

[krasserm]

Probably you ran into keras-team/keras#11281

[leviome]

tensorflow==1.9.0
tensorflow-gpu==1.9.0
keras==2.2.0
I upgraded keras version to 2.2.4, the error still existed.
may I know your version of tf and keras

[krasserm]

You will need to re-train the model as the h5 files have been written with the older Keras version. For a quick test, can you please train the model again for only 1 epoch and then use the generated h5 file with the demo? This should work now. I'm using Keras 2.2.4 and tested with Tensorflow 1.10 and 1.12.

[leviome]

thanks a lot!
I retrained it with keras 2.2.4, and the new model is available for demo.py.
Your suggestions helped much!

[krasserm]

You're welcome, glad it worked!

[ink1]

this helped me too. consider adding keras=2.2.4 as a requirement.
and many thanks for sharing. much appreciated!

[ngonthier]

Sorry but it doesn't work for me to update keras to 2.2.4.