voxelmorph: A wrong pre-trained weights?
Task (what are you trying to do/register?)
I am trying to create deformable templates by using the pre-trained weights you provided (conditional atlas). And I found the outputs of layer atlas_gen always give me noisy maps, it is a simple test, so I guess maybe you put the wrong one?
What have you tried
Using this model: Creation of Deformable Templates.
Details of experiments key code:
atlas_gen_model = tf.keras.models.Model(model.inputs[:1], model.get_layer('atlas_gen').output)
input_samples = [tf.keras.utils.to_categorical(np.arange(pheno_shape), pheno_shape)]
pred = atlas_gen_model.predict(input_samples)
About this issue
- Original URL
- State: open
- Created 2 years ago
- Comments: 17
hmm, did either/both of you use the code in the colab? It might be helpful to try to debug/figure this out with something simple like MNIST.
@kvttt the reason to we add the initial rough template is that it acts like an “initialization”, in a sense having the atlas part of the network learn just the residual.