about the pre-train weight

[Fahim-F]

Hi,
Did you test the pre-train model on tensorflow backend?
I got an error when use it while the theano backend works.
I need a tensorflow pre-train model.

Thanks

[flyyufelix]

Oh yes, I forgot to erase a line in custom_layers.py which resulted in an error. It should work now!

[Fahim-F]

Hi
Thanks for your attention!
But again it does not work!!
I think it is not related to custom_layers.py, the *.h5 weights file of tensorflow has a problem and I think you save it with the old version of tensorflow.
Thanks

[flyyufelix]

@fahim7 , I think the *.h5 weights file is backward compatible. Which model did you try to run? What was the corresponding error message?

[Fahim-F]

I tested it with all three models in a tensorflow backend and I got a dimension error.
for example if I only load the weights of the pre-trained model (DenseNet 121 (k=32)) as follow:

model.load_weights('/weights/densenet121_weights_tf.h5',by_name = True)

I get the following error:
ValueError: Dimension 0 in both shapes must be equal, but are 192 and 256 for 'Assign_41' (op: 'Assign') with input shapes: [192], [256].

while if I change the backend as theano and load a model in the theano backend, every things done well.
But I want to use a tensoflow backend.

[flyyufelix]

I am using Keras 2.0.5 and Tensorflow 1.2.1 and all models work fine with Tensorflow backend. What version of Keras and Tensorflow are you using?

[Fahim-F]

My keras 2.0.5 and tensorflow 1.2.0

[datlife]

Same error with DenseNet169 TF weights.

Here is my system info:

>>> print(keras.__version__)
2.0.5
>>> print(tensorflow.__version__)
1.2.0

...
ValueError: Dimension 0 in both shapes must be equal, but are 1 and 128 for 'Assign_97' (op: 'Assign') with input shapes: [1,1,256,256], [128,256,1,1].

[Tangzy7]

same problems with tensorflow 1.2.1 and keras 2.0.6

[taoddiao]

set reduction=0.5 to fix the ValueError issue when loading weights

[Rabia-Metis]

It's also not working with reduction=0.5. Before it was giving error in 'Assign_97' and now by changing reduction=0.5 it is giving error in 'Assign_1126'

[taoddiao]

@Rabia-Metis which densenet did you run. I am doing fine with densenet121 on tensorflow 1.3 and keras 2.0.9

[Rabia-Metis]

@taoddiao Thanks for quick reply
I am using densenet161, keras 2.1.2 and tf 1.4.0
And my input shape is 224,224,3
Following is the exact wording of the error
"ValueError: Dimension 1 in both shapes must be equal, but are 2 and 1000 for 'Assign_1126' (op: 'Assign') with input shapes: [2208,2], [2208,1000]."

[taoddiao]

@Rabia-Metis It is seems that you have changed the output shape classes=1000 to classes=2. In this case, you cannot load weights directly because the weights have the output shape classes=1000. One feasible solution is to assign weights layer by layer.

[Rabia-Metis]

@taoddiao Oh I forgot it. Thanks for pointing it out

[rahul-321]

@Rabia-Metis your observation is correct, but how do we do that?